Add Batch dceca8c9-fb0f-4e4d-8635-84ff1b773143

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/1de0f91e-0150-4eab-a1fc-f5394801fd7c_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9a99a512cee0199080bbb525b38eea7d2270a4bab268ff10617d47d02576a73
+size 74674

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/1de0f91e-0150-4eab-a1fc-f5394801fd7c_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edf4980f250e3e615dccf2e1ea22d000e61de1b3adc5e4383b8e5b1167948759
+size 91319

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/1de0f91e-0150-4eab-a1fc-f5394801fd7c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e76ea4f112a14023f56ff16c0ca40dbb5ffcbe2c8395547b529f551b8648126
+size 1995587

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/full.md

@@ -0,0 +1,277 @@
+# POMATO: Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction
+
+Songyan Zhang $^{1*}$  Yongtao Ge $^{2,3*}$  Jinyuan Tian $^{2*}$  Guangkai Xu $^{2}$  Hao Chen $^{2\boxtimes}$  Chen Lv $^{1}$  Chunhua Shen $^{2}$
+
+<sup>1</sup>Nanyang Technological University, Singapore <sup>2</sup>Zhejiang University, China <sup>3</sup>The University of Adelaide, Australia
+
+![](images/7d7d39de5cf1840271816b75909cf2854e95e233438c2a390e2761672d9cfe1d.jpg)  
+Figure 1. 3D reconstruction from an arbitrary dynamic video with POMATO. Without relying on external modules, POMATO can directly perform 3D reconstruction along with temporal 3D point tracking and dynamic mask estimation.
+
+![](images/9d756e08d3e2d94f27849af6b4e03f65b0e8cf5fee25fbbabbc5bbe174eee64c.jpg)
+
+![](images/bfcaaab9a5ec21305cfa58e3d67b19c49ade436db127ec3a7bcdd3ad0e48ee84.jpg)
+
+# Abstract
+
+Recent approaches to 3D reconstruction in dynamic scenes primarily rely on the integration of separate geometry estimation and matching modules, where the latter plays a critical role in distinguishing dynamic regions and mitigating the interference caused by moving objects. Furthermore, the matching module explicitly models object motion, enabling the tracking of specific targets and advancing motion understanding in complex scenarios. Recently, the proposed representation of pointmap in DUSt3R suggests a potential solution to unify both geometry estimation and matching in 3D space, effectively reducing computational overhead by eliminating the need for redundant auxiliary modules. However, it still struggles with ambiguous correspondences in dynamic regions, which limits reconstruction
+
+performance in such scenarios. In this work, we present POMATO, a unified framework for dynamic 3D reconstruction by marrying POintmap MAatching with Temporal mOtion. Specifically, our method first learns an explicit matching relationship by mapping RGB pixels across different views to 3D pointmaps within a unified coordinate system. Furthermore, we introduce a temporal motion module for dynamic motions that ensures scale consistency across different frames and enhances performance in 3D reconstruction tasks requiring both precise geometry and reliable matching, most notably 3D point tracking. We show the effectiveness of our proposed POMATO by demonstrating the remarkable performance across multiple downstream tasks, including video depth estimation, 3D point tracking, and pose estimation. Code and models are publicly available at https://github.com/wyddmw/POMATO.
+
+![](images/d4790d7013afc4e3913de26bff86a1058157e1f8f2647138f96914e1b1e2a19f.jpg)  
+Image1
+
+![](images/da7cfadfc356bf2c1cf00c8b8279df0074817de3b6f5c9b690134e137e9a30f0.jpg)  
+Figure 2. Ambiguity in 3D point matching in dynamic scenes with DUSt3R. Given representative corresponding pixels of background (orange) and moving foreground (red) in two different views, DUSt3R outputs a pair of 3D points within the same coordinate system. In static regions, identical pixels share the same 3D coordinates which provide an accurate matching relationship in 3D space, but in moving regions, the 3D coordinates are inconsistent for corresponding pixels across views, leading to ambiguous 3D matching relationships.
+
+![](images/80ba9d359d10fae8be286af68910bd22527b49e2e0e9982c65725f9e7b2f3600.jpg)
+
+# 1. Introduction
+
+Image-based 3D reconstruction is a fundamental task in computer vision with a wide range of applications including SLAM [39], robotics [19, 48], autonomous driving [52], and novel view synthesis [5]. While substantial progress has been achieved in static 3D reconstruction [16, 23, 26, 44, 50], dynamic scenes remain a major hurdle due to complexities like non-rigid motion and deformation, which may hamper the learning of local structure and camera motion, thereby complicating accurate 3D reconstruction for dynamic scenes. These scenarios require explicit modeling of both scene geometry and object motion. Moreover, downstream reconstruction tasks, such as 3D point tracking, demand precise geometry estimation and robust matching across views. To effectively distinguish dynamic regions, it is essential to establish reliable correspondences between different frames. Some pioneering works have attempted to address dynamic motion by incorporating additional auxiliary matching modules, such as optical flow [42, 51] or 2D tracking [46]. However, these approaches may suffer from domain gaps and accumulated errors between modules, limiting their effectiveness. A unified framework that seamlessly integrates geometry estimation and matching for dynamic 3D reconstruction remains a critical and underexplored challenge.
+
+Recently, DUSt3R [44] proposes a promising solution to address this challenge. It introduces the concept of a pointmap that assigns each pixel in an image to a corresponding 3D coordinate. The network utilizes a standard transformer-based encoder-decoder architecture and receives a pair of images as input. The system incorporates two parallel decoders to predict pointmaps for each view within the same coordinate system. However, this representation is limited to static matching and struggles in dynamic scenes, as illustrated in Fig. 2.
+
+To address this problem, we present POMATO, a unified network for dynamic 3D reconstruction by marrying
+
+POintmap MMatching with Temporal mOtion. We argue that with iterative cross-attention modules across different views, matching features are well preserved in the decoder tokens. We thus introduce an auxiliary pointmap matching head to learn explicit correspondences. Specifically, for each pixel in the second view, the pointmap matching head predicts the corresponding 3D coordinates of its counterpart in the first view, under the shared coordinate system. Our proposed pointmap-based matching representation enables the establishment of explicit correspondences in 3D space, which can be directly leveraged for motion analysis, especially the estimation of dynamic regions. Moreover, we further extend our POMATO to handle 4D video sequences by introducing a temporal motion module that enhances the learning of temporal motions. This motion module promotes scale consistency across different frames and improves performance in tasks where both accurate geometry and reliable matching are paramount, most notably 3D point tracking. Compared with recent temporal 3D reconstruction methods [41, 43] based on an autoregression manner where the previous frames are blocked from the recently added frames, our temporal motion module is based on the self-attention mechanism along the temporal dimension, facilitating a comprehensive interaction across all frames. Our POMATO is trained in a two-stage manner. In the first stage, we used pairwise input images to learn fundamental geometry and matching capabilities. In the second stage, we extend the input to sequential video frames and incorporate the temporal motion module, enabling the model to effectively capture motions over time.
+
+Our contributions can be summarized in threefold: First, we propose a novel approach that unifies the fundamental geometry estimation and motion understanding for dynamic 3D reconstruction into a single network by incorporating the representation of pointmap matching. Second, we introduce a temporal motion module to facilitate the interactions of motion features along the temporal dimension, which significantly improves the performance in tasks
+
+where both accurate geometry and precise matching are required for video sequential input, most notably 3D point tracking. Third, we demonstrate promising performance on 3D vision tasks, including video depth estimation, 3D point tracking, and camera pose estimation.
+
+# 2. Related Work
+
+Geometry estimation refers to the process of determining the spatial properties and structures from different forms of visual data. Direct recovery of 3D geometry from a single RGB image is by nature an ill-posed problem. Many recent works [3, 16, 23, 50] have tried to leverage strong pre-trained models to learn generalizable depthmaps from large-scale real and synthetic datasets to solve ambiguities. For example, Marigold [23], Geowizard [11], and GenPercept [47] aim at leveraging the generative priors from pre-trained diffusion models by finetuning them on synthetic datasets. Depthanything V2 [50] proposes to estimate scale-and-shift invariant disparity map by finetuning DINOV2 [29] model on synthetic datasets and largescale pseudo labels. Depth Pro [3] further propose a FOV head to estimate the metric depthmap from a single image without relying on camera intrinsics as input. Due to the scale ambiguity in the monocular depth estimation models, ChronoDepth [36], DepthCrafter [17], and Depth-any-video [49] proposes to learn temporal consistent depthmaps by leveraging the priors from a video generative model, i.e. SVD [2]. In another line of the research, multi-view stereo reconstruction (MVS) methods seek to reconstruct visible surfaces from multiple viewpoints. Traditional MVS [12] and SfM pipelines break the reconstruction pipeline into several sub-problems, e.g., feature extraction [8], image matching [1, 27], triangulation, and bundle adjustment [7]. The chain is complicated and accumulates noise for every single step, thus often resulting in unsatisfactory performance in complex real-world scenes. Recognizing the limitations of previous MVS methods, seminal work DUSt3R [44] proposes 3D pointmaps representation, and trains a network from large-scale data to regress the dense and accurate pointmaps from a pair of images. The camera intrinsics and relative camera poses can be implicitly inferred from the two-view pointmaps. However, it still can not handle reconstruction for dynamic scenes. MonST3R [51] directly finetuned the original DUSt3R model upon synthetic datasets that contain dynamic scenes. Motion representation. Optical flow is a commonly used representation for 2D motion. RAFT [38] is a representative work for pairwise optical flow estimation, which employs a 4D cost volume and recurrently estimates the optical flow. Some follow-up methods further extend it to multi-frame (3-5 frames) settings, which is still insufficient for long-range tracking. To resolve the problem, Particle Video [35] represent video motion by using a set of particles. Each
+
+particle is an image point sample with a long-duration trajectory and other properties. Particle videos have two key advantages over optical flow: (1) persistence through occlusions, and (2) multi-frame temporal context. Some recent works, PIPs [15], TAPIR [9] and Cotracker [22] have renewed interest in this representation and show promising long-term 2D point tracking results. Recognizing the advantage of point representation, SpatialTracker [46] lifts the 2D points into 3D and performs tracking in the 3D space. Though it can handle occlusions and enhance 3D tracking accuracy, it still relies on a separate monocular depth estimator, which prevents it performing 3D point tracking in an end-to-end fashion.
+
+Multi-view dynamic reconstruction. Our work is closely connected to multi-view dynamic 3D reconstruction techniques. Early works [32, 34] take the straightforward idea that first pre-segment the scene into different regions, each corresponding to a single rigid part of an object, then apply the rigid-SfM technique to each of the regions. Some of the recent Neural Radiance Fields (NeRF) [28] and Gaussian Splatting [24] based methods have achieved state-of-the-art results. However, most of these methods require simultaneous multi-view video inputs or require predefined templates [18]. Shape of motion [42], proposes a new dynamic scene representation to represent the dynamic scene as a set of persistent 3D Gaussians, and optimize the representation from a monocular video by leveraging monocular depth estimation priors and 2D track estimates across frames.
+
+# 3. Method
+
+# 3.1. Preliminary
+
+The overview of our POMATO is demonstrated in Fig.3. We adopt the definition of pointmap  $\mathbf{X} \in \mathbb{R}^{H \times W \times 3}$  in DUSt3R [44] as a dense 2D field of 3D points where each point corresponds to its respective RGB pixel. Given a pair of input images  $\mathbf{I}^1, \mathbf{I}^2 \in \mathbb{R}^{H \times W \times 3}$  from two different views, a weight-sharing ViT first extracts the corresponding features  $\mathbf{F}^1, \mathbf{F}^2$  for each view. Two parallel branches are employed to decode the geometric structures and enhance the feature alignment via cross-attention in decoder modules, following a regression head to estimate pointmaps  $\mathbf{X}^{1,1}, \mathbf{X}^{2,1} \in \mathbb{R}^{H \times W \times 3}$  along with a confidence map  $\mathbf{C}^{1,1}, \mathbf{C}^{2,1} \in \mathbb{R}^{H \times W}$  for each image view. Generally,  $\mathbf{X}^{n,m}$  indicates the pointmap  $\mathbf{X}^n$  from camera  $n$  expressed in camera  $m$ 's coordinate frame, which is obtained by a rigid transformation:
+
+$$
+\mathbf {X} ^ {n, m} = \mathbf {P} _ {m} \mathbf {P} _ {n} ^ {- 1} h \left(\mathbf {X} ^ {n}\right), \tag {1}
+$$
+
+where  $\mathbf{P}_m, \mathbf{P}_n \in \mathbb{R}^{3 \times 4}$  are world-to-camera poses for camera  $m$  and camera  $n$ , respectively, and  $h(\mathbf{X}^n)$  is a homogeneous mapping for the 3D coordinate in camera coordinate
+
+![](images/6c2a4e30e8ce899d819fb5ecc880bda38163cdd7993ced19b28b4dd5846716c8.jpg)  
+Figure 3. Overview of our training pipeline. (1) Stage I: build upon DUSt3R [44] architecture, we introduce a third regression point-matching head:  $\mathrm{Head}_3$ , which is in parallel to  $\mathrm{Head}_2$  for explicit pointmap matching in 3D space. For each pixel in the second view, the output pointmap coordinate is the 3D point map of the corresponding pixel in the first view. (2) Stage II: we introduce a temporal fusion module in three heads that enables multi-style sequential input for learning temporal motions.
+
+![](images/605719015a3f01a385bb6b41f66a977cb3406db4d9da2a81979a20f7ae60c171.jpg)
+
+of camera  $n$
+
+The task for Decoder 1 and its regression head estimate the 3D points for  $\mathbf{I}^1$  in its own coordinate system while Decoder 2 and its regression head are responsible for estimating pixel-wise 3D coordinates for  $\mathbf{I}^2$  in  $\mathbf{I}^1$ 's coordinate system after a rigid transformation of global rotation and translation. In the following contents, we will first introduce our POMATO with pairwise input images and then extend it to the video sequence input with our temporal motion module.
+
+# 3.2. Pointmap Matching with Pairwise Input
+
+As discussed before, the definition of  $\mathbf{X}^{2,1}$  depicts a rigid camera transformation that is ambiguous to reflect explicit matching relationships for dynamic regions. To tackle this, we propose to formulate an explicit pointmap matching  $\mathbf{X}_m^{2,1} \in \mathbb{R}^{H \times W \times 3}$  that maps dense RGB pixels of  $\mathbf{I}^2$  to 3D coordinates of corresponding pixels in  $\mathbf{I}^1$  under the first image's coordinate system. Given a 2D query pixel at  $(x_2, y_2)$  in  $\mathbf{I}^2$  and its corresponding pixel at  $(x_1, y_1)$  in  $\mathbf{I}^1$ , the matched pointmap at  $(x_2, y_2)$  in  $\mathbf{I}^2$  is:
+
+$$
+\mathbf {X} _ {m} ^ {2, 1} \left(x _ {2}, y _ {2}\right) = \mathbf {X} ^ {1, 1} \left(x _ {1}, y _ {1}\right), \tag {2}
+$$
+
+where  $(x,y)$  indicates the coordinates of 2D grid. For the representative dynamic point (red) in Fig. 2, the pointmap matching result is the 3D coordinate of point A in the coordinate system of the first image. As shown in Fig. 3,  $\mathbf{X}_m^{2,1}$  and  $\mathbf{X}^{1,1}$  are supposed to match perfectly in 3D space on the premise of neglecting occluded regions. We argue that the set of decoder tokens from the second branch preserves abundant matching information with iterative cross-attentions, so we introduce a matching head with the same architecture of  $\mathrm{Head}_1$  and  $\mathrm{Head}_2$ . The supervision for pointmap matching  $\mathbf{X}_m^{2,1}$  still follows the 3D regression loss which is defined as the Euclidean distance:
+
+$$
+\mathcal {L} _ {\mathrm {m}} = \left\| \frac {1}{z _ {m}} \mathbf {X} _ {m} ^ {2, 1} - \frac {1}{\bar {z} _ {m}} \bar {\mathbf {X}} _ {m} ^ {2, 1} \right\|, \tag {3}
+$$
+
+where  $\bar{\mathbf{X}}_m^{2,1}$  is the ground truth pointmap matching, which can be obtained following Eq. 2 on the 2D tracking dataset
+
+with the depth and camera information.  $z_{m},\bar{z}_{m}$  are the same norm factor defined in DUSt3R. The matching confidence  $\mathbf{C}_m^{2,1}$  is also learned following the confidence loss for Head1 and Head2 within valid regions:
+
+$$
+\mathcal {L} _ {\mathrm {m c o n f}} = \mathbf {C} _ {m} ^ {2, 1} \mathcal {L} _ {\mathrm {m}} - \alpha \log \mathbf {C} _ {m} ^ {2, 1} \tag {4}
+$$
+
+The final loss  $\mathcal{L}$  of our POMATO for pairwise input is a combination of predefined DUSt3R loss  $\mathcal{L}_{\mathrm{DUSt3R}}$ , matching loss  $\mathcal{L}_{\mathrm{m}}$ , and matching confidence loss  $\mathcal{L}_{\mathrm{mconf}}$ . When training our POMATO for pairwise input images at the first stage, the parameters in the encoder are frozen.
+
+# 3.3. Dynamic Mask Estimation
+
+Taking advantage of the explicit pointmap matching head, our POMATO can directly perform dynamic mask estimation without introducing an assistant module such as the optical flow model, getting rid of the additional computation cost and the potential domain gap. For an image pair  $\{\mathbf{I}^i,\mathbf{I}^j\}$  along with the estimation of  $\mathbf{X}^{j,i}$  from  $\mathrm{Head}_2$  and  $\mathbf{X}_{m}^{j,i}$  from  $\mathrm{Head}_3$ , the dynamic mask  $\mathbf{D}^{j,i}$  can be obtained by comparing the difference between  $\mathbf{X}^{j,i}$  and  $\mathbf{X}_{m}^{j,i}$ :
+
+$$
+\mathbf {D} ^ {j, i} = \left| \left| \mathbf {X} _ {m} ^ {j, i} - \mathbf {X} ^ {j, i} \right| \right| > \alpha , \tag {5}
+$$
+
+where  $\alpha$  is a dynamic threshold defined as  $3 \times$  median  $(||\mathbf{X}_m^{j,i} - \mathbf{X}^{j,i}||)$ . The explicit dynamic mask can be incorporated into the global alignment process to minimize the interference of moving objects for pose estimation and 3D reconstruction. Details on the incorporation of dynamic masks for global alignment are provided in the supplementary materials.
+
+# 3.4. Temporal Motion Module
+
+With the fundamental capability of geometric estimation and pointmap matching for pairwise images, we follow [6] and extend our POMATO to 4D video sequences by inserting a transformer-based motion module into the vanilla DPT head to construct the "temporal DPT head", which is illustrated in Fig.4. For a set of decoder tokens  $\mathbf{G} \in \mathbb{R}^{B,T,N,C}$
+
+![](images/e73694188c1b8131f76e305cf3ef1fcdd588d7576f6399241e71b55f09d6872f.jpg)  
+Figure 4. Architecture of our temporal motion module. We insert a transformer-based motion module (in shallow yellow) into the vanilla DPT [33] head to enhance the temporal consistency.
+
+where  $B, T, N, C$  represent the batch size, window length of a video sequence, token number, and token dimension, respectively, we merge the token number dimension into the batch axis and apply the motion module which consists of two blocks of standard multi-head self-attention modules and feed-forward networks along the temporal dimension  $T$ . To reduce the computation cost, the temporal motion modules are applied to features of low resolution.
+
+# 3.5. Downstream Temporal Tasks
+
+Given a video sequence of  $T$  frames  $\mathbf{I}^{t_1},\mathbf{I}^{t_2},\dots ,\mathbf{I}^{t_T}$  we construct a unique set of stereo image pairs for each task. As illustrated in Fig. 5, the flexible construction of input pairs—combined with the proposed temporal motion module and pointmap matching head—enables POMATO to seamlessly address downstream temporal tasks, including 3D point tracking, video depth estimation, and 3D reconstruction. The keyframe selection strategy and input formulation for each task are detailed in the following section.
+
+Besides the default regression losses for  $\mathrm{Head}_1$  and  $\mathrm{Head}_2$ , and predefined losses Eq. 3 and Eq. 4 for  $\mathrm{Head}_3$ , we further employ a temporal consistency loss,  $\mathcal{L}_{\mathrm{t}}$ , which will be described in detail below.
+
+In addition to the default regression losses for  $\mathrm{Head}_1$  and  $\mathrm{Head}_2$ , and the predefined losses in Eq. 3 and Eq. 4 for  $\mathrm{Head}_3$ , we further introduce a temporal consistency loss,  $\mathcal{L}_{\mathrm{t}}$ , which will also be described in detail below.
+
+3D Point Tracking. As illustrated at the top of Fig.5, the keyframe is set to the first image of the global video sequence and fed to the proposed  $\mathrm{Head}_3$  to obtain the pointmap matching result of each query point (initialized at the first image) under the coordinate system of each reference frame  $\{\mathbf{X}_{m}^{t_{1},t_{1}},\mathbf{X}_{m}^{t_{1},t_{2}},\mathbf{X}_{m}^{t_{1},t_{3}},\dots \mathbf{X}_{m}^{t_{1},t_{T}}\}$ , while the set of reference frames  $\{\mathbf{I}^{t_1},\mathbf{I}^{t_2},\mathbf{I}^{t_3},\dots \mathbf{I}^{t_T}\}$  is fed to the  $\mathrm{Head}_1$  to obtain the pointmap under each ego coordinate system. The dense tracking results can be further sparsified by indexing the 2D coordinates. When inference on a video
+
+![](images/03e46edafee22826cca167e638e68f533c2191534b14727c1a972191680c8ac9.jpg)  
+Figure 5. Inference pipelines for point tracking, video depth, and multi-view reconstruction.  $t_k$  indicates the keyframe. With the help of the motion module and flexible input construction, PO-MATO can be easily applied to downstream temporal tasks.
+
+longer than  $T$  frames, a simple sliding-window approach with an overlap of four frames is adopted to enhance the consistency between adjacent video windows. The temporal consistency loss  $\mathcal{L}_{\mathrm{t}}$  for tracking is:
+
+$$
+\mathcal {L} _ {\mathrm {t}} = \frac {1}{T} \sum_ {i = 1} ^ {T} \left\| \frac {\mathbf {X} _ {m} ^ {t _ {1} , t _ {i}}}{z _ {m} ^ {T}} - \frac {\bar {\mathbf {X}} _ {m} ^ {t _ {1} , t _ {i}}}{\bar {z} _ {m} ^ {T}} \right\| + \left\| \frac {\mathbf {X} ^ {t _ {i} , t _ {i}}}{z ^ {T}} - \frac {\bar {\mathbf {X}} ^ {t _ {i} , t _ {i}}}{\bar {z} ^ {T}} \right\|, \tag {6}
+$$
+
+where  $z_{m}^{T} = \mathrm{norm}\left(\mathbf{X}_{m}^{t_{1},t_{1}},\mathbf{X}_{m}^{t_{1},t_{2}},\dots,\mathbf{X}_{m}^{t_{1},t_{T}}\right)$  and  $\bar{z}_T =$  norm  $(\bar{\mathbf{X}}_m^{t_1,t_1},\bar{\mathbf{X}}_m^{t_1,t_2},\dots,\bar{\mathbf{X}}_m^{t_1,t_T})$  .  $z_{m}^{T}$  and  $\bar{z}_T$  are defined analogously.
+
+Video Depth Estimation. As shown in the middle part of the Fig. 5, the input video sequence is formulated to a set of identical image pairs  $\{(\mathbf{I}^{t_1},\mathbf{I}^{t_1}),(\mathbf{I}^{t_2},\mathbf{I}^{t_2}),\dots,(\mathbf{I}^{t_T},\mathbf{I}^{t_T})\}$  and fed to  $\mathrm{Head}_1$  and  $\mathrm{Head}_2$ , where the predictions from each head are identical:  $\{\mathbf{X}^{t_1,t_1},\mathbf{X}^{t_2,t_2},\dots,\mathbf{X}^{t_N,t_N}\}$ . We use the output of  $\mathrm{Head}_1$  as our final video depth estimation. The temporal consistency loss  $\mathcal{L}_{\mathrm{t}}$  is defined as:
+
+$$
+\mathcal {L} _ {\mathrm {t}} = \frac {1}{T} \sum_ {i = 1} ^ {T} \left\| \frac {\mathbf {X} _ {1} ^ {t _ {i} , t _ {i}}}{z _ {1} ^ {T}} - \frac {\bar {\mathbf {X}} ^ {t _ {i} , t _ {i}}}{\bar {z} ^ {T}} \right\| + \left\| \frac {\mathbf {X} _ {2} ^ {t _ {i} , t _ {i}}}{z _ {2} ^ {T}} - \frac {\bar {\mathbf {X}} ^ {t _ {i} , t _ {i}}}{\bar {z} ^ {T}} \right\|, \tag {7}
+$$
+
+where  $\mathbf{X}_1^{t_i,t_i}$  and  $\mathbf{X}_2^{t_i,t_i}$  indicate the output from Head_1 and Head_2, respectively.  $\bar{\mathbf{X}}^{t_i,t_i}$  is the pointmap groundtruth.
+
+3D Reconstruction. Assisted by the temporal motion module, redundant post-process operations such as global alignment can be omitted, allowing the reconstructed 3D point cloud to be obtained in a feed-forward manner. As shown in the bottom part of Fig.5, the keyframe is set to the last frame
+
+<table><tr><td rowspan="2">Alignment</td><td rowspan="2">Method</td><td rowspan="2" colspan="2">Optim. Onl.</td><td colspan="2">Sintel [4]</td><td colspan="2">BONN [30]</td><td colspan="2">KITTI [13]</td></tr><tr><td>Abs Rel ↓</td><td>δ&lt;1.25 ↑</td><td>Abs Rel ↓</td><td>δ&lt;1.25 ↑</td><td>Abs Rel ↓</td><td>δ&lt;1.25 ↑</td></tr><tr><td rowspan="6">Per-sequence scale</td><td>DUSt3R-GA [44]</td><td>✓</td><td></td><td>0.656</td><td>45.2</td><td>0.155</td><td>83.3</td><td>0.144</td><td>81.3</td></tr><tr><td>MASt3R-GA [26]</td><td>✓</td><td></td><td>0.641</td><td>43.9</td><td>0.252</td><td>70.1</td><td>0.183</td><td>74.5</td></tr><tr><td>MonST3R-GA [51]</td><td>✓</td><td></td><td>0.378</td><td>55.8</td><td>0.067</td><td>96.3</td><td>0.168</td><td>74.4</td></tr><tr><td>Spann3R [41]</td><td></td><td>✓</td><td>0.622</td><td>42.6</td><td>0.144</td><td>81.3</td><td>0.198</td><td>73.7</td></tr><tr><td>CUT3R [43]</td><td></td><td>✓</td><td>0.421</td><td>47.9</td><td>0.078</td><td>93.7</td><td>0.118</td><td>88.1</td></tr><tr><td>POMATO</td><td></td><td>✓</td><td>0.416</td><td>53.6</td><td>0.074</td><td>96.1</td><td>0.085</td><td>93.3</td></tr><tr><td rowspan="3">Per-sequence scale &amp; shift</td><td>MonST3R-GA [51]</td><td>✓</td><td></td><td>0.335</td><td>58.5</td><td>0.063</td><td>96.4</td><td>0.104</td><td>89.5</td></tr><tr><td>CUT3R [43]</td><td></td><td>✓</td><td>0.466</td><td>56.2</td><td>0.111</td><td>88.3</td><td>0.075</td><td>94.3</td></tr><tr><td>POMATO</td><td></td><td>✓</td><td>0.345</td><td>57.9</td><td>0.072</td><td>96.5</td><td>0.084</td><td>93.4</td></tr></table>
+
+Table 1. Video depth evaluation. We report scale-invariant depth and scale & shift invariant depth accuracy on Sintel [4], Bonn [30], and KITTI [13] datasets. Methods requiring global alignment are marked “GA”, while “Optim.” and “Onl.” indicate optimization-based and online methods, respectively. The best and second best results in each category are bold and underlined, respectively.
+
+$\mathbf{I}^{t_T}$  within the temporal window of length  $T$  and is fed to  $\mathrm{Head}_1$  with a set output of  $\{\mathbf{X}^{t_T,t_T},\mathbf{X}^{t_T,t_T},\dots,\mathbf{X}^{t_T,t_T}\}$ . All the reference frames are input to the  $\mathrm{Head}_2$  so the target pointmaps  $\{\mathbf{X}^{t_1,t_T},\mathbf{X}^{t_2,t_T},\dots,\mathbf{X}^{t_T,t_T}\}$  are aligned under the coordinate system of the keyframe. The temporal consistency loss  $\mathcal{L}_{\mathrm{t}}$  is:
+
+$$
+\mathcal {L} _ {\mathrm {t}} = \frac {1}{T} \sum_ {i = 1} ^ {T} \left\| \frac {\mathbf {X} ^ {t _ {T} , t _ {T}}}{z _ {1} ^ {T}} - \frac {\bar {\mathbf {X}} ^ {t _ {T} , t _ {T}}}{\bar {z} _ {1} ^ {T}} \right\| + \left\| \frac {\mathbf {X} ^ {t _ {i} , t _ {T}}}{z _ {2} ^ {T}} - \frac {\bar {\mathbf {X}} ^ {t _ {i} , t _ {T}}}{\bar {z} _ {2} ^ {T}} \right\| \tag {8}
+$$
+
+We further freeze the parameters in Decoder1 and Decoder2 when training the temporal downstream tasks at the second stage. In our work, the temporal window length  $T$  is set to 12. Additional explorations on the temporal length can be found in Sec.4.
+
+# 4. Experiments
+
+# 4.1. Experimental Details
+
+Training data. We train our network with a mixture of five datasets: PointOdyssey [53], Tartanair [45], ParallelDomain4D [40], DynamicReplica [21] and Carla (0.9.15) [10]. The specific number and the usage ratio of each dataset can be found in the supplementary materials. All datasets include pixel-accurate ground truth depth, as well as camera intrinsics and extrinsics, and encompass a wide variety of dynamic scenes across both indoor and outdoor environments. Among them, PointOdyssey and DynamicReplica have additional 2D trajectory annotations for dynamic objects which can be used to construct pointmap matching ground truth following Eq. 2. All datasets are used to supervise geometry learning on  $\mathrm{Head}_1$  and  $\mathrm{Head}_2$ , while only PointOdyssey, DynamicReplica, and TartanAir are used to train the proposed pointmap matching head.
+
+Training and inference details. Our model architecture is based on the publicly available DUSt3R [51] model, utilizing the same backbone consisting of a ViT-Large encoder
+
+and a ViT-Base decoder. To fully leverage MonST3R's geometry estimation capabilities in dynamic scenes, we initialize our model using the publicly available MonST3R checkpoint. For the newly introduced pointmap matching head, we initialize its weights from the pretrained  $\mathrm{Head}_2$  weights of MonST3R. The temporal motion module is initialized following [14]. We train our network for 10 epochs with a cosine learning rate schedule, with an initial learning rate of 1e-4. In the first stage, which involves pairwise training, we use a batch size of 16 on 4 A100 GPUs (40G). In the second stage, where the temporal motion module is introduced, the batch size is set to 4 with a fixed temporal window length of 12. During each training iteration, we randomly sample a downstream task—3D point tracking, video depth estimation, or 3D reconstruction—to construct the input pairs and apply the corresponding loss function.
+
+# 4.2. Video Depth Estimation
+
+Following MonST3R [51] and CUT3R [43], we rescale all predictions from the same video to align them together by conducting two forms of alignment: per-sequence scale and shift alignment and per-sequence scale alignment. Thus, we can measure the per-frame depth quality and inter-frame depth consistency. We employ our proposed motion module for video depth estimation in a feed-forward manner as described in Sec.3.5 and compare our method against several variants of DUSt3R, including DUSt3R [44], MAST3R [26], MonST3R [51], Spann3R [41], and CUT3R [43]. Given 6 frames of  $288\times 512$  on an NVIDIA 4070 GPU, POMATO reconstructs the 3D point cloud in 0.7 seconds, whereas global alignment-based methods such as MonST3R require 5.8 seconds. As shown in Tab. 1, our method demonstrates comparable performance to the global alignment (GA)-based MonST3R [51] on the Sintel [4] and BONN [30] datasets, while surpassing it on KITTI dataset. Besides, we
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">PointOdyssey [53]</td><td colspan="2">ADT [31]</td><td colspan="2">PStudio [20]</td><td colspan="2">Average</td></tr><tr><td>L-12</td><td>L-24</td><td>L-12</td><td>L-24</td><td>L-12</td><td>L-24</td><td>L-12</td><td>L-24</td></tr><tr><td>SpatialTracker* [46]</td><td>20.46</td><td>20.71</td><td>21.64</td><td>20.67</td><td>30.41</td><td>25.87</td><td>24.17</td><td>22.42</td></tr><tr><td>DUSt3R [44]</td><td>19.03</td><td>19.03</td><td>29.02</td><td>25.55</td><td>9.72</td><td>6.50</td><td>19.26</td><td>17.03</td></tr><tr><td>MASt3R [26]</td><td>16.58</td><td>17.35</td><td>27.36</td><td>26.46</td><td>11.78</td><td>8.09</td><td>18.57</td><td>17.30</td></tr><tr><td>MonST3R [51]</td><td>27.31</td><td>27.92</td><td>28.30</td><td>26.13</td><td>16.50</td><td>11.06</td><td>24.03</td><td>21.70</td></tr><tr><td>POMATO</td><td>33.20</td><td>33.58</td><td>31.57</td><td>28.22</td><td>24.59</td><td>19.79</td><td>29.79</td><td>27.20</td></tr></table>
+
+Table 2. 3D tracking evaluation. We report the APD metric to evaluate 3D point tracking on the PointOdyssey [53], ADT [31], and PStudio [20] datasets. L-12 and L-24 indicate tracking within the temporal length of 12 frames and 24 frames, respectively.
+
+![](images/03a55f67a47387cf7a0f1955fb4795a7aa96e25d9492413951874f029abfce2e.jpg)  
+Figure 6. Qualitative comparison of dynamic scenes. Compared to MonST3R, our POMATO can provide more reliable motion masks, 3D point tracking, and reconstruction performance.
+
+consistently outperform the state-of-the-art online method, CUT3R [43], across various settings. These results underscore the effectiveness of our approach, specifically (1) the joint learning of geometry and pointmap matching, and (2) the temporal motion module.
+
+# 4.3. 3D Point Tracking
+
+For 3D point tracking task, we use the Aria Digital Twin (ADT) [31], and Panoptic Studio (PStudio) [20] benchmarks from the TAPVid-3D [25] dataset along with the validation set on the PointOdyssey [53] dataset. We report the Average Percent Deviation (APD) metric, which quantifies the average percentage of points within a threshold relative to the ground truth depth. The APD metric serves as a direct measure of the accuracy of the predicted tracking. We reformulate the datasets and project all the query points within a temporal window to the first frame. We report tracking results on the length of 12 and 24 frames. As shown in Tab.2, our POMATO achieves the best performance on both PointOdyssey and ADT datasets. It's worth mentioning that SpatialTracker [46] is a state-of-the-art network tailored for 3D point tracking with ground truth camera intrinsic as ad
+
+ditional input data. POMATO surpasses it on two datasets and improves the average APD metric by  $23.3\%$  and  $21.4\%$  for 12 frames and 24 frames, respectively. For DUSt3R-based methods, we use the output of  $\mathrm{Head}_2$  as tracking results. Obviously, the ambiguous matching representation limits its capability to handle this fine-grained 3D reconstruction task in dynamic scenes. 1
+
+# 4.4. Camera Pose Estimation
+
+Following DUSt3R-based methods, we perform global alignment with the model trained in the first stage on the Bonn [30] and TUM [37] datasets. The sampling stride is set to 5 for the Bonn dataset and 3 for the TUM dataset. Compared with optical-flow assisted global alignment in MonST3R, the dynamic mask is computed according to Eq. 5 while the 2D pseudo label is replaced by projecting the pointmap matching results to 2D coordinates with estimated camera intrinsic. Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot) are reported. The evaluation results over 40 frames are reported in Tab. 4. Notably, POMATO obtains an overall state-of-the-art performance and signifi
+
+<table><tr><td rowspan="3">Temporal Length</td><td colspan="6">Video Depth</td><td colspan="3">Tracking (12 Frames)</td></tr><tr><td colspan="2">Sintel [4]</td><td colspan="2">Bonn [30]</td><td colspan="2">KITTI [13]</td><td colspan="3">PointOdyssey [53] ADT [31] PStudio [20]</td></tr><tr><td>Abs Rel ↓ δ&lt;1.25 ↑</td><td>Abs Rel ↓ δ&lt;1.25 ↑</td><td>Abs Rel ↓ δ&lt;1.25 ↑</td><td>Abs Rel ↓ δ&lt;1.25 ↑</td><td>APD↑</td><td>APD↑</td><td>APD↑</td><td></td><td></td></tr><tr><td>Pair-wise</td><td>0.548</td><td>46.2</td><td>0.087</td><td>94.0</td><td>0.113</td><td>89.5</td><td>32.06</td><td>29.87</td><td>23.10</td></tr><tr><td>6 frames</td><td>0.436</td><td>51.3</td><td>0.076</td><td>95.9</td><td>0.085</td><td>93.5</td><td>32.69</td><td>30.93</td><td>24.52</td></tr><tr><td>12 frames</td><td>0.416</td><td>53.6</td><td>0.075</td><td>96.1</td><td>0.086</td><td>93.3</td><td>33.20</td><td>31.57</td><td>24.59</td></tr></table>
+
+Table 3. Ablation study on the temporal motion module. The introduction of the temporal motion module brings a significant improvement. As the temporal window length enlarges from 6 frames to 12 frames, we obtain an overall consistent improvement.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">TUM [37]</td><td colspan="3">Bonn [30]</td></tr><tr><td>ATE ↓</td><td>RPE trans ↓</td><td>RPE rot ↓</td><td>ATE ↓</td><td>RPE trans ↓</td><td>RPE rot ↓</td></tr><tr><td>DUSt3R [44]</td><td>0.025</td><td>0.013</td><td>2.361</td><td>0.030</td><td>0.025</td><td>2.522</td></tr><tr><td>MASt3R [26]</td><td>0.027</td><td>0.015</td><td>1.910</td><td>0.031</td><td>0.025</td><td>2.478</td></tr><tr><td>MonST3R [51]</td><td>0.021</td><td>0.006</td><td>1.142</td><td>0.025</td><td>0.021</td><td>2.120</td></tr><tr><td>CUT3R [43]</td><td>0.023</td><td>0.016</td><td>0.510</td><td>0.028</td><td>0.033</td><td>2.569</td></tr><tr><td>POMATO</td><td>0.020</td><td>0.010</td><td>0.509</td><td>0.037</td><td>0.016</td><td>1.782</td></tr></table>
+
+![](images/1c6c752e80c60fcd5532fb788c94cf5b5034109a6a0a070aefad9cd940462a1d.jpg)  
+Figure 7. Effectiveness of our pointmap matching head. Without explicitly filtering out the motion area, both pose and geometry estimation will be degraded.
+
+Table 4. Pose estimation. Our method achieves an overall best performance and improves the RPE rot metric significantly.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Bonn [30]</td><td>PointOdyssey [53]</td><td>ADT [31]</td><td>PStudio [20]</td></tr><tr><td>ATE ↓</td><td>RPE trans ↓</td><td>RPE rot ↓</td><td>APD ↑</td><td>APD ↑</td><td>APD ↑</td></tr><tr><td>W/O Head3</td><td>0.040</td><td>0.015</td><td>1.721</td><td>29.10</td><td>29.62</td><td>16.94</td></tr><tr><td>W/ Head3</td><td>0.037</td><td>0.016</td><td>1.782</td><td>32.06</td><td>29.87</td><td>23.10</td></tr></table>
+
+Table 5. Ablation study on the effectiveness of the pointmap matching head. The comparisons are reported on the pose estimation and 3D point tracking tasks.
+
+cantly improves the RPE-rot metric, surpassing MonST3R by  $55.4\%$  and  $13.3\%$  on the TUM and Bonn datasets.
+
+# 4.5. Ablation Study
+
+We conduct extensive ablation studies to evaluate the effectiveness of the temporal motion module and the proposed pointmap matching head. As shown in Table 3, we report results for three models: one trained with only pairwise
+
+images (first-stage training), one using a shorter temporal window of 6 frames, and another using the default temporal window length of 12 frames. Incorporating temporal consistency yields substantial improvements across all datasets for video depth estimation and 3D point tracking. Further improvement is achieved when the temporal window length increases from 6 frames to 12 frames. In Table 5, we evaluate the effectiveness of the pointmap matching head. While it introduces only a modest improvement in the ATE metric, we attribute this to the limited motion and minimal viewpoint variation in the indoor evaluation dataset. As illustrated in Fig. 7, under challenging in-the-wild conditions with significant motion and rapid viewpoint changes, removing the pointmap matching head introduces ambiguity in explicit rigid transformation estimation, resulting in a clear degradation in performance. To further demonstrate the impact of the pointmap matching head on 3D point tracking, we conduct tracking experiments over 12 frames using the pairwise input setup. Clearly, removing the pointmap matching head (using only  $\mathrm{Head}_2$ ) leads to an inevitable performance drop, emphasizing explicit correspondence modeling for reliable long-term tracking.
+
+# 5. Discussion and Conclusion
+
+We introduce POMATO, a unified framework for geometry estimation and motion understanding in dynamic scenes. By leveraging the proposed pointmap matching head, our method effectively distinguishes moving regions, thereby mitigating the interference introduced by dynamic objects. The temporal motion module further facilitates the learning of temporal dynamics across frames, enhancing scale consistency and improving performance in tasks where both geometry and matching are critical—most notably, 3D point tracking. The downstream temporal tasks including 3D point tracking, video depth estimation, and 3D reconstruction can be easily applied in a feed-forward manner. In future work, we plan to scale up training with more dynamic reconstruction and matching datasets to further enhance 3D reconstruction and tracking performance.
+
+Acknowledgement. This work was supported by the National Natural Science Foundation of China (No. 62206244)
+
+# References
+
+[1] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. Affineglue: Joint matching and robust estimation. arXiv preprint arXiv:2307.15381, 2023. 3  
+[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. abs/2311.15127, 2023. 3  
+[3] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv, 2024. 3  
+[4] Daniel J. Butler, Jonas Wulff, Garrett B. Stanley, and Michael J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625, 2012. 6, 8  
+[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. CVPR, 2023. 2  
+[6] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. arXiv preprint arXiv:2501.12375, 2025. 4  
+[7] Yu Chen, Yisong Chen, and Guoping Wang. Bundle adjustment revisited. arXiv preprint arXiv: 1912.03858, 2019. 3  
+[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3  
+[9] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 3  
+[10] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 6  
+[11] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv: 2403.12013, 2024. 3  
+[12] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 3  
+[13] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361, 2012. 6, 8  
+[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning, 2023. 6
+
+[15] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, pages 59-75. Springer, 2022. 3  
+[16] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3  
+[17] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. In CVPR, 2025. 3  
+[18] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 3  
+[19] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv: 2410.20220, 2024. 2  
+[20] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 7, 8  
+[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 6  
+[22] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In Proc. ECCV, 2024. 3  
+[23] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3  
+[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 3  
+[25] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. arXiv preprint arXiv: 2407.05921, 2024. 7  
+[26] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. European Conference on Computer Vision, 2024. 2, 6, 7, 8  
+[27] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 3  
+[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:
+
+Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3  
+[29] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, and Marc Szafraniec et al. DINOv2: Learning robust visual features without supervision. Trans. Mach. Learn. Research, 2024. 3  
+[30] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 6, 7, 8  
+[31] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. arXiv preprint arXiv: 2306.06362, 2023. 7, 8  
+[32] Rene Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4058-4066, 2016. 3  
+[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 5  
+[34] Chris Russell, Rui Yu, and Lourdes Agapito. Video popu- up: Monocular 3d reconstruction of dynamic scenes. In European conference on computer vision, pages 583-598. Springer, 2014. 3  
+[35] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 3  
+[36] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. abs/2406.01493, 2024. 3  
+[37] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. pages 573-580, 2012. 7, 8  
+[38] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 3  
+[39] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Neural Information Processing Systems, 2021. 2  
+[40] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. arXiv preprint arXiv:2405.14868, 2024. 6  
+[41] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 6
+
+[42] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2, 3  
+[43] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state, 2025. 2, 6, 7, 8  
+[44] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In CVPR, pages 20697-20709, 2024. 2, 3, 4, 6, 7, 8  
+[45] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual SLAM. pages 4909-4916, 2020. 6  
+[46] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 7  
+[47] Guangkai Xu, Yongtao Ge, Mingyu Liu, Chengxiang Fan, Kangyang Xie, Zhiyue Zhao, Hao Chen, and Chunhua Shen. Diffusion models trained with large data are transferable visual models. arXiv preprint arXiv: 2403.06090, 2024. 3  
+[48] Yueming Xu, Haochen Jiang, Zhongyang Xiao, Jianfeng Feng, and Li Zhang. Dg-slam: Robust dynamic gaussian splatting slam with hybrid pose optimization. arXiv preprint arXiv: 2411.08373, 2024. 2  
+[49] Honghui Yang, Di Huang, Wei Yin, Chunhua Shen, Haifeng Liu, Xiaofei He, Binbin Lin, Wanli Ouyang, and Tong He. Depth any video with scalable synthetic data. arXiv preprint arXiv:2410.10815, 2024. 3  
+[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 2, 3  
+[51] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arxiv:2410.03825, 2024. 2, 3, 6, 7, 8  
+[52] Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, Wenjun Mei, and Xingang Wang. Drivedreamer4d: World models are effective data machines for 4d driving scene representation. 2024. 2  
+[53] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. PointOdyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 6, 7, 8

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb54e7a6a5818da12bf328f41937e70e9cfc5f15bf14cdc6ce72fe14cbac4734
+size 700141

ICCV/2025/POMATO_ Marrying Pointmap Matching with Temporal Motions for Dynamic 3D Reconstruction/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39e581dcbfd335a5ad0b24b0eedd28b749215bbd7ec57fe7e0f7808b56f86c8d
+size 377092

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/521e29cf-6292-4b2a-8333-77ba2b903010_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:169f5adb90a74e8902dbb5df387c9cae9f04b2cc577f53bde191aa040b9167df
+size 89010

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/521e29cf-6292-4b2a-8333-77ba2b903010_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abb3e249211ec69c5d1835f10542142440b8fdf553692e4f9dce3e481982fc0f
+size 115445

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/521e29cf-6292-4b2a-8333-77ba2b903010_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53814b710b4f9967d579ec4995bb1b14a64a0a8ac556edfba1d7639b1628ef13
+size 25834177

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/full.md

@@ -0,0 +1,329 @@
+![](images/53fc065fe924d277a3c87ed3bce9fcbd4e5e70514d50127ac3e8581566f8ad50.jpg)
+
+# PRE-Mamba: A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining
+
+Ciyu Ruan $^{1,\ast}$ , Ruishan Guo $^{1,\ast}$ , Zihang Gong $^{2}$ , Jingao Xu $^{3}$ , Wenhan Yang $^{4}$ , Xinlei Chen $^{1,\dagger}$ $^{1}$ Shenzhen International Graduate School, Tsinghua University,  $^{2}$ Harbin Institute of Technology,  $^{3}$ Carnegie Mellon University,  $^{4}$ Pengcheng Laboratory,
+
+{rcy23, grs24}@mails.tsinghua.edu.cn, gongzihang0201@gmail.com,
+
+jingaox@andrew.cmu.edu, yangwh@pcl.ac.cn, chen.xinlei@sz.tsinghua.edu.cn
+
+![](images/d10678f1157603b25cce2a297987f9ffbbacff341b09ab667bb14a33332a6c63.jpg)  
+(a) Comparison in rainy scenario
+
+![](images/bb73f351aa4fbc02674ee348315a6fda8cfaf01711fde45e7cee74b3b1e671af.jpg)  
+(b) Direct Sampling of Raw Data
+
+![](images/bf6ce5de68327edbfdeab985b3d647efcaccc378a50653a841853683685f6ba9.jpg)  
+(c) Multi-scale State Space Model  
+Figure 1. (a) Rain severely degrades event camera data compared to conventional cameras. (b) Unlike frame-based aggregation methods, our approach directly samples raw data, preserving temporal resolution. (c) The multi-scale state space model integrates dual-temporal scales for intra- and inter-window information and multi-spatial scales for rain dynamics. (d) Deraining results of our method.
+
+![](images/b16275764078c2a8382e2d9085cbfac504c8eb212e79b9dbf756b8f508cb9744.jpg)  
+(d) Deraining Performance
+
+# Abstract
+
+Event cameras excel in high temporal resolution and dynamic range but suffer from dense noise in rainy conditions. Existing event deraining methods face trade-offs between temporal precision, deraining effectiveness, and computational efficiency. In this paper, we propose PRE-Mamba, a novel point-based event camera deraining framework that fully exploits the spatiotemporal characteristics of raw event and rain. Our framework introduces a 4D event cloud representation that integrates dual temporal scales to preserve high temporal precision, a Spatio-Temporal Decoupling and Fusion module (STDF) that enhances deraining capability by enabling shallow decoupling and interaction of temporal and spatial information, and a Multi-Scale State Space Model (MS3M) that captures deeper rain dynamics across dual-temporal and multi-spatial scales with linear computational complexity. Enhanced by frequency-domain regularization, PRE-Mamba achieves superior performance (0.95 SR, 0.91 NR, and 0.4s/M events) with only 0.26M parameters on EventRain-27K, a comprehensive
+
+dataset with labeled synthetic and real-world sequences. Moreover, our method generalizes well across varying rain intensities, viewpoints, and even snowy conditions. Code and dataset: https://github.com/softword-ii/PRE-Mamba.
+
+# 1. Introduction
+
+Event cameras have revolutionized mobile applications with their high temporal resolution (microsecond level), high dynamic range ( $>120$  dB), and low power consumption ( $<10$  mW) [1, 2], driving breakthroughs in high-speed tracking [3, 4], SLAM [5-7], localization [8] and obstacle avoidance [9, 10]. Unlike frame cameras, they asynchronously capture pixel-level intensity changes without global exposure, eliminating motion blur even at extreme speeds. However, despite excelling in dynamic scenes, event cameras suffer significant performance degradation under adverse weather, particularly in rain [11, 12]. The rapid motion of rain streaks triggers excessive intensity changes, generating dense noise that overwhelms valid scene information and exacerbates sampling rates. This severely impacts accuracy and latency in downstream tasks on drones [13-15] and autonomous vehicles [16-18], limiting real-world reliability.
+
+Despite significant advances in deep learning for frame
+
+camera deraining [19-22], event camera deraining remains largely unexplored. Existing solutions predominantly follow a frame-based paradigm, converting event streams into grayscale frames [12] and applying frame-based deraining techniques [11], which sacrifice the fine-grained temporal resolution and sparsity of event data. Point-based methods align naturally with event streams' asynchronous, sparse nature, but remain unapplied to deraining due to notable limitations. Point Networks [23, 24], GNNs [25, 26], and SNNs [27-29] still underperform compared to frame-based methods, while Point Transformers face computational bottlenecks under high event rates ( $>10\mathrm{M}$  events/s) due to quadratic attention complexity [30, 31]. As a result, balancing temporal precision, deraining effectiveness, and computational efficiency remains an open challenge.
+
+Recently, the State Space Model (SSM) [32], particularly Mamba, provides a promising solution with linear complexity and long-range context modeling. However, introducing Mamba to event camera deraining is non-trivial. The original design of the Mamba is aimed at solving causal sequential language tasks, which differ from the asynchronous, sparse, and high-temporal-resolution nature of event data. Rain streaks further complicate this by introducing diverse spatiotemporal patterns, requiring the model to capture complex dependencies across both time and space.
+
+To address this, we propose PRE-Mamba, the first Point-based deRaining framework for Event camera. As a point-based method, it processes raw events directly instead of quantized frames, fully preserving the sensor's native microsecond  $(\mu s)$  temporal resolution and facilitating downstream high-speed vision tasks. Our framework first introduces a 4D event cloud representation (§3.1), integrating inter- and intra-temporal windows to preserve high temporal precision. To handle the spatial irregularity and motion characteristics of rain, we propose a Spatio-Temporal Decoupling and Fusion module (STDF, §3.2), enabling shallow decoupling and interaction of temporal and spatial information. For deeper modeling, we propose a Multi-Scale State Space Model (MS3M, §3.3) that efficiently captures complex rain dynamics across dual temporal and multispatial scales with linear computational complexity. To improve feature discrimination, we introduce a frequency regularization term (§3.4) in the loss function, guiding the model to learn rain distribution patterns in the frequency domain. To evaluate our approach, we construct EventRain27K (§4), the first comprehensive point-based event deraining dataset, comprising 18K labeled synthetic and 9K unlabeled real-world rain sequences. Our model achieves state-of-the-art performance (0.95 SR, 0.91 NR, and 0.4s/M events) with only 0.26M parameters, validated through extensive experiments (§5). Moreover, it generalizes well across diverse rain intensities and snowy conditions. Our contributions are summarized as follows:
+
+- To the best of our knowledge, we propose the first point-based event deraining framework.  
+- To effectively extend Mamba for event deraining, we address the asynchronous, sparse nature of event data and the diverse spatiotemporal patterns of rain streaks by introducing a 4D event cloud, an STDF module, and an MS3M enhanced with frequency regularization loss.  
+- We introduce EventRain-27K, the first point-based event deraining with labeled synthetic and real-world sequences, providing a benchmark for future research.  
+- We achieve superior performance on synthetic and real datasets with high efficiency and a lightweight design.
+
+# 2. Related Work
+
+Existing deraining approaches for frame camera. Deraining methods for frame cameras are categorized into traditional and deep learning-based approaches. Traditional methods, such as dictionary learning and low-rank matrix decomposition [33-36], rely on handcrafted assumptions about the directional nature or sparsity of rain streaks, often failing to generalize to complex real-world scenarios. In contrast, deep learning-based methods, leveraging CNNs and Transformer [37, 38] architectures, have shown significant advances. CNNs effectively model local spatial patterns, capturing rain stripe textures and background structures [19, 39], while Transformer-based models [20-22] enhance performance by capturing long-range dependencies and global context, offering superior deraining capabilities.
+
+Existing deraining approaches involving event camera. Research on event-based deraining remains limited despite its importance for outdoor applications. Existing methods, such as EGVD [40], RainVID&SS [41], and NESD [42], leverage the high temporal resolution and dynamic range of event cameras to assist frame camera by providing prior knowledge of rain location and density. Other methods designed for event-based deraining typically convert asynchronous event streams into synchronous frames like representations, aligning with image deraining but sacrificing temporal resolution and sparsity. For instance, DistillNet [11] partially preserves temporal resolution through voxel representations but still loses their asynchronous nature. Moreover, the unique motion and spatiotemporal distribution of rain create a domain barrier, limiting the applicability of traditional event denoising methods like filtering [43-45] and clustering [46]. In addition, existing deraining datasets primarily focus on frame camera [47-50] or represent event data in image-like formats [51], further hinder the event camera deraining research.
+
+Event Data Processing Architectures. Event data processing approaches are broadly categorized into frame-based and point-based frameworks. Frame-based methods [45, 52-54] not only face the aforementioned limitations but also incur redundant computational overhead by
+
+![](images/1457058ce192e44dcc23e44dd6d41481b45575731f2842eabaf3a9f9cbd274f0.jpg)  
+Figure 2. Overview of Our Proposed PRE-Mamba. The model takes  $n$  temporal windows as input to generate a 4D event cloud, serialized via four scanning modes. The STDF module decouples and enhances spatiotemporal features across dual temporal scales, which flow through a U-Net style encoder-decoder with multiple blocks. Each block employs MS3M as its core, integrating intra- and interwindow branches to capture appearance and motion features, alongside a multi-spatial-scale pathway for handling diverse rain streak scales while preserving local dynamics. Finally, the decoder outputs are fed into a linear layer followed by softmax for per-event predictions.
+
+processing event-free regions [55], making them inefficient for resource-constrained environments [53, 56]. In contrast, point-based methods process events in their native form, preserving temporal resolution and sparsity. These include Point Networks [57, 58], Graph Neural Networks (GNNs) [59], and Spiking Neural Networks (SNNs) [29, 60], excel in capturing rapid motion and fine-grained temporal dynamics. However, GNNs and SNNs often require specialized hardware or face performance issues, while Point Cloud networks, based on assumptions of permutation and transformation invariance, are poorly suited for temporal event data and require time-aware feature extraction.
+
+To address these limitations, this paper proposes a novel approach based on selective State Space Models (SSMs) [61-63]. Renowned for its linear computational complexity and parameter efficiency [32, 64], Mamba is well-suited for processing long event sequences exceeding 1M length [30]. It provides a promising direction to bridge performance gap between point-based and frame-based methods [30, 65] while preserving the inherent advantages of event data, facilitating the development of lightweight, real-time deraining systems.
+
+# 3. Method
+
+To perform event deraining effectively and efficiently while preserving temporal precision, we propose a novel framework, PRE-Mamba. We formulate event deraining as an event-by-event classification task, akin to event denoising, reducing computational complexity and avoiding artifacts
+
+compared to reconstruction-based methods. Figure 2 illustrates the architecture of our system. At the input level, a 4D Event Cloud Representation (§3.1) integrates inter- and intra-temporal windows to preserve high temporal precision. For feature extraction, a Spatio-Temporal Decoupling and Fusion module (§3.2) enables effective decoupling and interaction of temporal and spatial information to capture complex rain dynamics. For deep modeling, a Multi-Scale State Space Model (§3.3) captures rain dynamics across dual temporal and multi-spatial scales with linear computational complexity. Finally, a Frequency-Regularized Optimization (§3.4) enhances feature discrimination by learning rain distribution patterns in the frequency domain, ensuring robust deraining accuracy. Together, these components form a unified framework that achieves both computational efficiency and deraining performance.
+
+# 3.1. 4D Event Cloud Representation
+
+To preserve fine-grained spatiotemporal information while enabling efficient processing, we transform raw event streams into a structured 4D event cloud representation. The raw event data is represented as a sequence  $\mathcal{E} = \{e_i\}_{i=1}^N$ , where each event  $e_i = (x_i, y_i, t_i, p_i)$  consists of spatial coordinates  $(x_i, y_i)$ , timestamp  $t_i$ , and polarity  $p_i$ .
+
+3D Pseudo-Point Cloud Construction: The event sequence is divided into fixed-duration temporal windows  $\{\mathcal{W}_k\}_{k=1}^L$ , each spanning a time interval  $T$ . Within each window  $\mathcal{W}_k$ , events are normalized into a 3D pseudo-point cloud representation:
+
+$$
+p _ {i} = \left(x _ {i}, y _ {i}, z _ {i}\right) = \left(x _ {i}, y _ {i}, \frac {t _ {i} - t _ {0}}{t _ {e} - t _ {0}}\right), \tag {1}
+$$
+
+where  $t_0$  and  $t_e$  denote the window's start and end timestamps, respectively. The normalized coordinate  $z_i$  encodes the relative temporal position of the event within the window, mapping event timestamps to a unified temporal scale. 4D Cross-Window Temporal Modeling: To model long-term dependencies across windows, we introduce a fourth dimension,  $T_n$ , which represents the global temporal index of the window. This extends the 3D representation to a 4D event cloud  $E_n = \{(x_i, y_i, z_i, T_n)\}_{i=1}^{M_n}$ , where  $M_n$  is the number of events in window  $n$ . The complete 4D event cloud is then  $E = \bigcup_{n=1}^{L} E_n$ .
+
+This hierarchical dual-scale framework captures local intra-window dynamics through normalized timestamp  $z$  and models global inter-window evolution via window index  $T_{\mathrm{n}}$ . Unlike fixed event-count methods that fragment or regroup events [31, 66], compromising temporal integrity, our approach dynamically adapts to event sparsity and irregularity by processing variable-length sequences within fixed time windows. We also serialize 4D event clouds into SSM-compatible sequences using  $z$ -order and Hilbert curves, aligning with point cloud processing paradigms [67, 68].
+
+# 3.2. Spatio-Temporal Decoupling and Fusion
+
+To address the complex spatiotemporal noise in rain events while maximizing the temporal discrimination power of event camera, we propose a Spatio-Temporal Decoupling and Fusion Module (STDF). This module explicitly decouples spatial and temporal features while modeling their interactions across dual scales in the 4D event cloud.
+
+We extract spatial features  $f_{\mathrm{s}}$  via depthwise 1D convolutions on  $(x, y, z, p)$  coordinates and temporal features  $f_{\mathrm{t}}$  through embeddings of window indices  $T_{\mathrm{n}}$ . To exploit rain events' temporal characteristics and event cameras' high resolution, we prioritize temporal information over spatial information through cross-domain feature modulation. Specifically, spatial features  $f_{\mathrm{s}}$  are refined through Hadamard products with complementary temporal representations: intra-window dynamics  $f_{t}^{\mathrm{intra}}$  from 1D convolutions along  $z$ , encoding microsecond correlations within windows, and inter-window trends  $f_{t}^{\mathrm{inter}}$  from 1D convolutions on  $T_{\mathrm{n}}$ , capturing long-term temporal trends across consecutive windows. This modulation is formulated as:
+
+$$
+f _ {\mathrm {s}} ^ {*} = f _ {\mathrm {s}} \otimes \left(f _ {\mathrm {t}} ^ {\text {i n t r a}} + f _ {t} ^ {\text {i n t e r}}\right). \tag {2}
+$$
+
+Finally, the temporally-modulated features are combined with the original spatial and temporal representations via residual addition  $(\mathcal{R})$ . The integrated features are sequentially processed by 1D convolution  $(\Phi_{1d})$ , batch normalization (BN), and activation  $(\sigma)$  to ensure training stability:
+
+$$
+\begin{array}{l} \mathcal {R} \left(f _ {\mathrm {s}}, f _ {\mathrm {t}}\right) = f _ {\mathrm {s}} + f _ {\mathrm {t}} + f _ {\mathrm {s}} ^ {*}, \tag {3} \\ f _ {\mathrm {S T D F}} = \sigma \left(\mathrm {B N} \left(\Phi_ {1 d} \left(\mathcal {R} (f _ {\mathrm {s}}, f _ {\mathrm {t}})\right)\right)\right). \\ \end{array}
+$$
+
+# 3.3. Multi-Scale State Space Model
+
+Building upon the hierarchical input representation and spatiotemporal feature embedding, we propose the Multi-Scale State Space Model (MS3M) to further model deeper rain dynamics from multi-temporal and spatial scales. This framework incorporates parallel intra- and inter-window branches to simultaneously capture appearance features and motion patterns, coupled with a multi-scale spatial pathway that preserves structural details and local rain dynamics.
+
+Dual-Temporal-Scale Global Modeling: While Mamba excels in sequence modeling, it struggles to leverage temporal features [69] for event deraining. Inspired by the Motion-aware State Space Model (MSSM), which excels in LiDAR-based moving object segmentation [70], we integrate MSSM into our dual-temporal-scale architecture. This enables explicit modeling of rain events' appearance and motion, improving rain-background separation.
+
+The architecture mainly consists of two complementary branches: an intra-window branch processes spatial appearance features  $f_{\mathrm{s}}^{\mathrm{intra}}$  through Reversed Aggregation (RA) and 1D convolution, capturing rain patterns and structural details for rain-background separation; an inter-window branch extracts motion-aware features  $f_{\mathrm{m}}^{\mathrm{inter}}$  by analyzing temporal dependencies, enabling separation of dynamic rain streaks from moving objects or scene changes. An adaptive gating mechanism [71] generates weight  $f_{\mathrm{G}}$  to emphasize rain-relevant features. Outputs from both branches are fused via cross-product attention with inter-branch interaction, then passed to SSM for global feature learning via scanning [72], followed by gating signal modulation.
+
+$$
+\begin{array}{l} f _ {\mathrm {s}} ^ {\text {i n t r a}} = \sigma \left(\Phi_ {1 d} \left(\mathrm {R A} \left(f _ {1}\right)\right)\right) \\ f _ {\text {f u s e}} = \sigma \left(f _ {\mathrm {m}} ^ {\text {i n t e r}}\right) \otimes f _ {\mathrm {s}} ^ {\text {i n t r a}} + f _ {\mathrm {m}} ^ {\text {i n t e r}} \tag {4} \\ f _ {\text {d u a l}} = \operatorname {S S M} \left(\sigma \left(f _ {\text {f u s e}}\right)\right) \otimes f _ {\mathrm {G}}. \\ \end{array}
+$$
+
+Multi-Spatial-Scale Local Pathway: To address rain scale and appearance diversity, we introduce a Multi-Scale Local Spatial Pathway. This pathway extracts multi-scale local features from the intra-window branch using convolutional layers with varying kernel sizes. Small kernel sizes ( $k_{s}$ ) capture fine details of thin rain streaks, while larger kernels model broader and more diffuse rain patterns. This design enhances the model's ability to handle diverse rain scenarios while preserving local structural details.
+
+$$
+\begin{array}{l} f _ {i} = \sigma \left(\Phi_ {1 d} ^ {k _ {i}} \left(f _ {i - 1}\right)\right), \quad i = 1, 2, \dots , n \\ f _ {0} = \Phi_ {1 d} ^ {k _ {0}} (\mathbb {L} (\mathrm {R A} (f _ {\mathrm {I}}))) \quad f _ {\mathrm {m s}} = \sigma \left(\sum_ {i = 1} ^ {n} f _ {i}\right) \tag {5} \\ \end{array}
+$$
+
+The multi-scale features  $f_{\mathrm{ms}}$  are aggregated and fused with the global outputs  $f_{\mathrm{dual}}$  via residual addition, seamlessly integrating local details and global context. Finally, a linear layer  $(\mathbb{L})$  refines the fused features, formulated as  $f_{\mathrm{out}} = \mathbb{L}(f_{\mathrm{ms}} + f_{\mathrm{dual}})$ , enhancing spatial precision while maintaining structural integrity.
+
+# 3.4. Training Loss
+
+![](images/40e0f1483264f087ea2224a6702c67b41a232614b322150a7af6359f9f3d9317.jpg)
+
+![](images/e4366231ffefaff397d82315ed5ebb9094ad3654b0d42fdd0ddd43be4cc850b4.jpg)
+
+![](images/ef1afa446e9b7a89434e17d8d7aaf8929bb14e374a2393e659143e9cf6b0cecf.jpg)
+
+![](images/16bc3bfd3439bfc46fbc0938199f13cd63d51b5a5dd673884f466de4a7a44b62.jpg)  
+Figure 3. Power Spectrum Heatmap and Segment Length Distribution for Rainfall Intensities (5, 20, 60 mm). Top: Heatmaps show higher frequencies at peak power for lower densities, reflecting sparse rain. Bottom: Distributions of continuous signals reveal longer segments at higher densities, indicating sustained rain.
+
+![](images/9da8306bcd63da747d601fc2e4d69c90037fb1d43968f949075e751d6222c5c6.jpg)
+
+![](images/86218fad09f5f9eb62b99210f2095a405eb2933dc1aeea99fe2d48ec74a53166.jpg)
+
+Our training objective combines a standard binary cross-entropy loss  $(\mathcal{L}_{\mathrm{ce}})$  with a frequency-domain regularization term  $(\mathcal{L}_{\mathrm{fit}})$  to guide the model in learning both point-wise accuracy and global frequency patterns, forming the joint loss as  $\mathcal{L} = \mathcal{L}_{\mathrm{ce}} + \lambda \mathcal{L}_{\mathrm{fit}}$ .
+
+Frequency-Domain Regularization: Existing image deraining methods integrate frequency layers to capture frequency-domain features [61, 73, 74]. However, for event camera with millions of event data, embedding Fast Fourier Transform (FFT) layer directly introduces significant computational overhead. Alternatively, we propose incorporating  $\mathcal{L}_{\mathrm{fft}}$  as a regularization term during model optimization, maintaining efficiency while mitigating overfitting.
+
+While  $\mathcal{L}_{\mathrm{ce}}$  focuses on per-event accuracy, it may overlook global distribution patterns, resulting in unrealistic predictions like isolated rain events. In contrast,  $\mathcal{L}_{\mathrm{ftt}}$  enforces statistical consistency by aligning frequency-domain distributions via amplitude and phase constraints:
+
+$$
+\mathcal {L} _ {\mathrm {f f t}} = \frac {1}{L} \sum_ {i = 1} ^ {L} \left(\frac {| \mathcal {F} (P) _ {i} - \mathcal {F} (Y) _ {i} |}{\operatorname* {m a x} (| \mathcal {F} (P) - \mathcal {F} (Y) | , \epsilon)} + \epsilon^ {\prime}\right) ^ {2}, \tag {6}
+$$
+
+where  $\mathcal{F}(\cdot)$  denotes the 1D FFT,  $L$  is the sequence length, and  $\epsilon, \epsilon'$  ensure numerical stability, and  $Y$  and  $P$  denote the ground truth labels and predicted outputs, respectively. Unlike traditional  $\mathcal{L}_{\mathrm{fft}}$  used in image reconstruction [61, 75], our formulation operates directly on event labels, leveraging event cameras' microsecond-level resolution to capture rain dynamics. Rainfall intensity directly correlated to event density and label 1's density (rain events labeled as 1). As illustrated in Figure 3, high-density rainfall manifests as low-frequency continuous patterns in the label sequence, while low-density rain exhibits high-frequency sparse patterns. This spectral distinction enables  $\mathcal{L}_{\mathrm{fft}}$  to align predictions with physically consistent global patterns [76].
+
+![](images/77bf151f5740c1e9b2c617eb867f43577a00b746df66a317940cb87d51ca331a.jpg)  
+Figure 4. EventRain-27K: The First Point-based Event Deraining Dataset. Additional dataset visualizations with varying rainfall intensities and scenarios are included in the Appendix.
+
+# 4. Dataset
+
+To support model training and testing, we published EventRain-27K, the first point-based benchmark dataset for event-based deraining. The dataset comprises over 7K self-generated synthetic rain samples, over 7K self-recorded artificial rain samples, and over 9K real rain samples.
+
+Synthetic Dataset (labeled). Supervised event deraining relies on labeled data; however, annotating discrete, sparse, and highly dynamic event points is non-trivial. To address this limitation, we use a two-step synthetic data generation approach. We first utilize the KITTI dataset [77] and the SPAC dataset [78] as clean video sources, then synthesize rain effects using state-of-the-art rendering simulators [79, 80]. The KITTI rainy dataset provides videos with varying camera motion speeds and rain intensities (4(a)), covering rainfall rates ranging from light rain (5 mm/hour) to severe storms (200 mm/hour). The SPAC rainy dataset includes 16 diverse scenes, such as cityscapes and natural environments, each augmented with 3 to 4 distinct rain patterns (4(b)). Finally, we generate event sequences from these rainy videos using the Vid2E simulator [81].
+
+Self-recorded Artificial Dataset (labeled). To bridge the domain gap between synthetic and real-world event data for model training, we introduce a self-recorded artificial dataset, capturing real event camera data under controlled artificial rainfall. A showerhead was employed to simulate raindrop free-fall motion, captured by a Prophesee EVK4 event camera  $(1280\times 720$  resolution,  $10\mathrm{k}$  fps,  $120\mathrm{dB}$  dy
+
+<table><tr><td rowspan="2">Metrics</td><td colspan="3">5mm</td><td colspan="3">20mm</td><td colspan="3">50mm</td><td colspan="3">80mm</td><td colspan="3">125mm</td><td colspan="3">150mm</td></tr><tr><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td><td>SR ↑</td><td>NR ↑</td><td>DA ↑</td></tr><tr><td>TS[82]</td><td>0.888</td><td>0.265</td><td>0.576</td><td>0.887</td><td>0.305</td><td>0.596</td><td>0.883</td><td>0.231</td><td>0.557</td><td>0.881</td><td>0.271</td><td>0.576</td><td>0.877</td><td>0.237</td><td>0.556</td><td>0.872</td><td>0.243</td><td>0.557</td></tr><tr><td>DWF[83]</td><td>0.703</td><td>0.394</td><td>0.549</td><td>0.734</td><td>0.464</td><td>0.599</td><td>0.755</td><td>0.352</td><td>0.553</td><td>0.777</td><td>0.432</td><td>0.604</td><td>0.786</td><td>0.371</td><td>0.578</td><td>0.782</td><td>0.375</td><td>0.578</td></tr><tr><td>Knoise[84]</td><td>0.860</td><td>0.375</td><td>0.618</td><td>0.870</td><td>0.377</td><td>0.624</td><td>0.884</td><td>0.241</td><td>0.563</td><td>0.890</td><td>0.257</td><td>0.574</td><td>0.901</td><td>0.211</td><td>0.556</td><td>0.896</td><td>0.214</td><td>0.555</td></tr><tr><td>Ynoise[85]</td><td>0.679</td><td>0.533</td><td>0.606</td><td>0.677</td><td>0.568</td><td>0.623</td><td>0.663</td><td>0.481</td><td>0.572</td><td>0.658</td><td>0.525</td><td>0.591</td><td>0.647</td><td>0.482</td><td>0.564</td><td>0.634</td><td>0.487</td><td>0.561</td></tr><tr><td>RED[86]</td><td>0.862</td><td>0.172</td><td>0.517</td><td>0.853</td><td>0.251</td><td>0.552</td><td>0.833</td><td>0.208</td><td>0.520</td><td>0.822</td><td>0.185</td><td>0.503</td><td>0.808</td><td>0.207</td><td>0.507</td><td>0.789</td><td>0.208</td><td>0.499</td></tr><tr><td>EDnCNN[86]</td><td>0.968</td><td>0.905</td><td>0.937</td><td>0.954</td><td>0.904</td><td>0.929</td><td>0.948</td><td>0.888</td><td>0.918</td><td>0.935</td><td>0.870</td><td>0.903</td><td>0.933</td><td>0.867</td><td>0.900</td><td>0.929</td><td>0.843</td><td>0.886</td></tr><tr><td>AEDNet[87]</td><td>0.941</td><td>0.850</td><td>0.891</td><td>0.938</td><td>0.876</td><td>0.907</td><td>0.928</td><td>0.732</td><td>0.830</td><td>0.925</td><td>0.681</td><td>0.803</td><td>0.922</td><td>0.624</td><td>0.773</td><td>0.923</td><td>0.547</td><td>0.735</td></tr><tr><td>EDformer[66]</td><td>0.981</td><td>0.818</td><td>0.899</td><td>0.962</td><td>0.832</td><td>0.897</td><td>0.924</td><td>0.844</td><td>0.884</td><td>0.894</td><td>0.834</td><td>0.864</td><td>0.876</td><td>0.840</td><td>0.858</td><td>0.839</td><td>0.834</td><td>0.836</td></tr><tr><td>Ours</td><td>0.994</td><td>0.914</td><td>0.954</td><td>0.978</td><td>0.915</td><td>0.947</td><td>0.955</td><td>0.911</td><td>0.933</td><td>0.940</td><td>0.903</td><td>0.922</td><td>0.918</td><td>0.898</td><td>0.908</td><td>0.908</td><td>0.895</td><td>0.902</td></tr></table>
+
+Table 1. Quantitative Comparisons on Synthetic Dataset with Varying Rainfall Intensities. We mark the best and second best results.
+
+![](images/f8003e4dee36b93e9df22d63cafc14667dfc2256819b52094e7ef686a6adb4f4.jpg)  
+Figure 5. Our fixed-point & speed data acquisition system.
+
+namic range) for both rainy and rain-free data. Slight camera vibrations were applied during captures to obtain complete scene details. Utilizing the K-Nearest Neighbors (KNN) algorithm, background events in rainy data were identified via spatiotemporal alignment with rain-free data, enabling background labeling and isolating rain events. The dataset focuses on urban outdoor scenes (buildings, trees, cars, lakes) and covers diverse rainfall intensities across four scenarios (Figure 4(c)).
+
+Real-world Dataset (label-free). To validate event deraining effectiveness and generalization, we collect a real-world event dataset. As illustrated in Figure 5, we designed a data acquisition system with fixed-point and fixed-speed (10 cm/s) cruising. We mounted a Propheesee EVK4 event camera on a motion-controlled vehicle, which traversed a slide rail to capture background and rain data simultaneously under controlled dynamic conditions. The dataset includes diverse viewpoints and rain intensities, covering a wide range of real-world scenarios (Figure 4(d)). Additionally, it incorporates 4K samples of real snow data captured to evaluate the model's generalization capability.
+
+# 5. Experiments
+
+# 5.1. Experiments Setups
+
+Evaluation Metrics. We use event Denoising Accuracy (DA), Signal Retention (SR) and Noise Removal (NR) to measure the event rain removal performance on PRE-Mamba. There is  $\mathrm{DA} = \frac{1}{2} (\mathrm{SR} + \mathrm{NR}) = (\frac{\mathrm{PB}}{\mathrm{TB}} +\frac{\mathrm{PR}}{\mathrm{TR}})$ , where PB, TB, PR, TR are the predicted background, the predicted rain, the real background, and the number of real rain, respectively. DA measures rain removal while preserving real motion (e.g., vehicles, pedestrians) and structural details
+
+(e.g., roads, building profiles). SR and NR analyze signal retention and noise suppression, respectively.
+
+Implementation Details. Our network is implemented in PyTorch and trained using the AdamW optimizer with an initial learning rate of  $4.8 \times 10^{-4}$  and a weight decay coefficient of  $5 \times 10^{-3}$ . The entire training is conducted on six NVIDIA RTX A6000 GPUs for 50 epochs with a batch size of 6. The model processes a sequence of five temporal windows as input, with each window spanning 0.1 seconds. In the multi-spatial-scale module of the MS3M, we employ kernel sizes of (1, 3, 5) to capture features at different spatial resolutions. Network details can be found in the Appendix.
+
+# 5.2. Comparisons with State-of-The-Art Methods
+
+Baselines. As this work represents the first attempt at point-based event deraining, there are no directly comparable baselines in the literature. Alternatively, we adapt state-of-the-art event denoising methods, as deraining can be viewed as a specialized case of noise removal. These include filter-based methods TS [82], DWF [83], Knoise [84], Ynoise [85], RED [86] and learning-based algorithms EDnCNN [86], AEDNet [87], EDformer [66]. Additionally, we qualitatively compare with DistillNet [11], a voxel-based GAN deraining method, but its voxelized processing is incompatible with event point denoising metrics, preventing quantitative evaluation.
+
+Results on Synthetic and Artificial Datasets. In Table 1, we present SR, NR, and DA results on the synthetic dataset with six rain intensities, ranging from light rain to heavy rain. Our method achieves state-of-the-art performance, with an average SR/NR/DA of  $0.95 / 0.91 / 0.93$ . It is notable that learning-based algorithms outperform filtering methods due to their ability to leverage supervised training. When training EDnCNN [86], we replaced its original event probability mask (EPM) labeling framework with our ground truth labels, which improved its performance. In addition, we provide qualitative results in Figure 6. Filtering methods are ineffective for rain removal, as rain streaks exhibit distinct spatiotemporal patterns that deviate from typical noise. Other learning-based methods yield incomplete deraining results, often retaining residual rain streaks. In contrast, our method effectively removes rain streaks and restores details through multi-scale temporal and spatial modeling
+
+![](images/463d8cc9a2b75a892419f8845159a8aaba16a1653feb2e133b02bcedde199891.jpg)  
+Figure 6. Qualitative Comparison on Synthetic and Artificial Datasets. PRE-Mamba demonstrates superior deraining capabilities across diverse structural contours and complex textures with enhanced detail preservation. Please zoom in for details.
+
+<table><tr><td></td><td>GFLOPs</td><td>Parameters</td><td>Inference time (s)</td><td>Relative speed</td></tr><tr><td>TS[82]</td><td>N/A</td><td>N/A</td><td>0.1296</td><td>1.0 ×</td></tr><tr><td>DWF[83]</td><td>N/A</td><td>N/A</td><td>0.0954</td><td>1.36 ×</td></tr><tr><td>KNoise[84]</td><td>N/A</td><td>N/A</td><td>0.0198</td><td>6.55 ×</td></tr><tr><td>YNoise[85]</td><td>N/A</td><td>N/A</td><td>0.0513</td><td>2.53 ×</td></tr><tr><td>RED[86]</td><td>N/A</td><td>N/A</td><td>2.2716</td><td>0.06 ×</td></tr><tr><td>EDnCNN[86]</td><td>234.51</td><td>614.55K</td><td>20.1885</td><td>1.0 ×</td></tr><tr><td>AEDNet[87]</td><td>4400.46</td><td>45.87M</td><td>43.4250</td><td>0.46 ×</td></tr><tr><td>DistillNet[11]</td><td>255.17</td><td>18.96M</td><td>0.2029</td><td>99.50 ×</td></tr><tr><td>EDformer[66]</td><td>8.41</td><td>49.80K</td><td>2.4943</td><td>8.09 ×</td></tr><tr><td>Ours</td><td>6.23</td><td>264.63K</td><td>0.0987</td><td>204.54 ×</td></tr></table>
+
+Table 2. Model complexity comparisons with previous model.
+
+combined with motion and appearance perception.
+
+Generalization on Real-World Dataset. To validate practical applicability, we compare our method against baselines on the real-world data of EventRain-27K. Figure 7 (top four rows) visualizes results for four scenarios, including three ground-level and one low-altitude viewpoints. Our method achieves superior performance compared to competitors. In the low-altitude scenario, our method effectively eliminates rain streaks while maintaining the structural integrity of static objects (e.g., buildings) and preserving the motion patterns of dynamic objects (e.g., vehicles). Filter-based methods treat both background and rain equally, resulting in effects similar to downsampling. While EDnCNN [86], AEDNet [87], and EDformer [66] are trained on synthetic data, their denoising-focused designs fail to generalize well to real rain, further hindered by the synthetic-to-real domain gap. DistillNet [11] removes large rain streaks via voxel-based GANs but loses background details and introduces artifacts. In contrast, our method, designed for point-based event deraining, effectively removes rain while preserving both background details and dynamic motion content, demonstrating robust real-world performance.
+
+Model Complexity and Efficiency Comparison. In Table
+
+<table><tr><td>Model</td><td>STDF</td><td>MS3M</td><td>Lfit</td><td>SR↑</td><td>NR↑</td><td>DA↑</td></tr><tr><td>M1</td><td></td><td></td><td></td><td>0.9123</td><td>0.8394</td><td>0.8759</td></tr><tr><td>M2</td><td>✓</td><td></td><td></td><td>0.8983</td><td>0.8747</td><td>0.8865</td></tr><tr><td>M3</td><td></td><td>✓</td><td></td><td>0.8941</td><td>0.8906</td><td>0.8923</td></tr><tr><td>M4</td><td>✓</td><td>✓</td><td></td><td>0.9046</td><td>0.8954</td><td>0.9000</td></tr><tr><td>Ours</td><td>✓</td><td>✓</td><td>✓</td><td>0.9080</td><td>0.8949</td><td>0.9015</td></tr></table>
+
+Table 3. Ablation about each module in PRE-Mamba on the EventRain-27K validation set.
+
+2, we compare the computational efficiency of our method and baselines in terms of FLOPs, parameters, and inference time for processing 100K events on an NVIDIA RTX A6000 GPU. Filtering-based methods enable fast inference but suffer from poor deraining performance due to lack of task-specific optimization. EDnCNN [86] and AEDNet [87] incur high computational overhead from encoding or sampling spatial-temporal neighborhoods, resulting in high FLOPs and slow inference. Transformer-based EDformer [66] reduces FLOPs and parameters but suffers from high latency due to the quadratic complexity of its attention mechanism. DistillNet [11] achieves a lower inference time by voxelizing events but at the cost of high computational complexity, large model size, and loss of fine-grained temporal precision. In contrast, our method achieves SOTA performance with only 6.23G FLOPs and 0.26M parameters, processing 100K events in 0.0987s and scaling to 1M events in 0.398s. Compared to EDnCNN and AEDNet, our method reduces FLOPs to  $2.66\%$  and  $0.14\%$  of theirs, respectively, while achieving  $204.5\times$  and  $440.0\times$  faster inference. Compared to the lightweight EDformer, our method achieves a  $25.3\times$  speedup with comparable model size, excelling in both accuracy and computational efficiency.
+
+# 5.3. Ablation Studies
+
+Baseline Design. We conduct systematic ablation studies to validate the efficacy of three proposed components:
+
+![](images/df93f306664efcbdfdc35074bb31c3ed3965497afb63573d86e6664442e14785.jpg)  
+Figure 7. Qualitative Comparison on Real-World Datasets. The top four rows highlight our method's superior deraining performance on real rain datasets, while the last two rows showcase its generalization to snow scenes. Please zoom in for details.
+
+<table><tr><td>Window Number</td><td>Num = 3</td><td>Num = 5</td><td>Num = 8</td></tr><tr><td>DA</td><td>0.8268</td><td>0.9015</td><td>0.9330</td></tr><tr><td>Improve (%)</td><td>-</td><td>9.03</td><td>12.84</td></tr><tr><td>Inference time (ms)</td><td>274.02</td><td>394.16</td><td>525.68</td></tr><tr><td>Relative speed</td><td>-</td><td>0.70 ×</td><td>0.52 ×</td></tr></table>
+
+Table 4. Ablation about time window number on validation set. the Spatio-Temporal Decoupling and Fusion (STDF) module, Multi-Scale State Space Model (MS3M), and Frequency Loss  $(\mathcal{L}_{\mathrm{fft}})$ . Our baseline model ("M1") removes all three components, employing naive summation for spatiotemporal feature fusion, reverting to the original MSSM block [70] for temporal modeling, and using only cross-entropy loss. Building upon "M1", "M2" introduces the STDF module to embed structured spatio-temporal features, while "M3" integrates the MS3M architecture for enhanced multi-scale modeling. And "M4" combines both STDF and MS3M components. Finally, we incorporate the Frequency Loss into "M4" to complete our full model.
+
+Quantitative Comparison. From Table 3, we observe that both STDF ("M2") and MS3M ("M3") improve the baseline performance by  $1.21\%$  and  $1.87\%$  in DA, respectively, validating their individual contributions. "M3" achieves greater improvement, demonstrating MS3M's ability to capture deep spatio-temporal features and adapt to diverse rain patterns. By integrating STDF with MS3M, "M4" further enhances performance, demonstrating the complementary nature of these components in modeling multi-scale rain and event characteristics. Finally, our full network outperforms "M4", which verifies the importance of the frequency-domain regularization term in leveraging both amplitude and phase constraints for improved learning.
+
+Effects of Time Windows Number. In Table 4, we evaluate model performance with time windows of 3, 5, and 8
+
+on validation set. Increasing the number of time windows improves the model's ability to capture temporal dependencies, with a  $9.0\%$  improvement in DA from 3 to 5 windows and an additional  $3.8\%$  improvement from 5 to 8. However, the inference speed decreases by  $30\%$  when the window size is increased from 3 to 5, and by  $48\%$  when increased from 3 to 8. Therefore, we set the number of time windows to 5, as further increases in time windows yield diminishing returns, emphasizing the need to balance temporal depth and computational efficiency.
+
+# 5.4. Generalization to Snowy Datasets
+
+While primarily designed for event-based deraining, our method also generalizes well to snow removal, as demonstrated in Figure 7 (last two rows). Without architectural changes, it delivers promising results under snowy conditions, showcasing robustness to diverse weather patterns. Additional visual results are included in the Appendix.
+
+# 6. Conclusion
+
+To the best of our knowledge, PRE-Mamba is the first point-based framework for event camera deraining. It adopts a 4D event cloud to preserve native temporal resolution via inter- and intra-temporal windows, a spatiotemporal decoupling and fusion module for efficient feature extraction, and a multi-scale state space model that captures rain dynamics across spatial and temporal scales with linear complexity. A frequency regularization term and cross-entropy loss guide learning, and we contribute a large-scale dataset with diverse rain scenarios. Extensive experiments demonstrate SOTA performance on synthetic and real-world data. The method is efficient, lightweight, and real-time capable, with potential for robust multi-agent perception [88, 89].
+
+Acknowledgments This paper was supported by the Natural Science Foundation of China under Grant 62371269, Guangdong Innovative and Entrepreneurial Research Team Program (2021ZT09L197), and Meituan Academy of Robotics Shenzhen.
+
+# References
+
+[1] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1  
+[2] Haoyang Wang, Ruishan Guo, Pengtao Ma, Ciyu Ruan, Xinyu Luo, Wenhua Ding, Tianyang Zhong, Jingao Xu, Yunhao Liu, and Xinlei Chen. Towards mobile sensing with event cameras on high-agility resource-constrained devices: A survey. arXiv preprint arXiv:2503.22943, 2025. 1  
+[3] Arren Glover and Chiara Bartolozzi. Robust visual tracking with a freely-moving event camera. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769-3776. IEEE, 2017. 1  
+[4] Xinyu Luo, Haoyang Wang, Ciyu Ruan, Chenxin Liang, Jingao Xu, and Xinlei Chen. Eventtracker: 3d localization and tracking of high-speed object with event and depth fusion. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 1974-1979, 2024. 1  
+[5] Jianhao Jiao, Huaiyang Huang, Liang Li, Zhijian He, Yilong Zhu, and Ming Liu. Comparing representations in tracking for event camera-based slam. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 1369-1376, 2021. 1  
+[6] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate slam? combining events, images, and imu for robust visual slam in hdr and high-speed scenarios. IEEE Robotics and Automation Letters, 3(2):994-1001, 2018.  
+[7] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1  
+[8] Haoyang Wang, Jingao Xu, Xinyu Luo, Xuecheng Chen, Ting Zhang, Ruiyang Duan, Yunhao Liu, and Xinlei Chen. Ultra-high-frequency harmony: mmwave radar and event camera orchestrate accurate drone landing. In Proceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems, pages 15-29, 2025. 1  
+[9] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1  
+[10] Nitin J Sanket, Chethan M Parameshwara, Chahat Deep Singh, Ashwin V Kuruttukulam, Cornelia Fermüller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 10651-10657. IEEE, 2020. 1  
+[11] Ciyu Ruan, Chenyu Zhao, Chenxin Liang, Xinyu Luo, Jingao Xu, and Xinlei Chen. Distill drops into data: Event
+
+based rain-background decomposition network. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 2072-2077, 2024. 1, 2, 6, 7, 8  
+[12] Long Cheng, Ni Liu, Xusen Guo, Yuhao Shen, Zijun Meng, Kai Huang, and Xiaqin Zhang. A novel rain removal approach for outdoor dynamic vision sensor event videos. Frontiers in Neurorobotics, 16:928707, 2022. 1, 2  
+[13] Xinlei Chen, Aveek Purohit, Carlos Ruiz Dominguez, Stefano Carpin, and Pei Zhang. Drunkwalk: Collaborative and adaptive planning for navigation of micro-aerial sensor swarms. In Proceedings of the 13th ACM Conference on Embedded Networked Sensor Systems, pages 295-308, 2015. 1  
+[14] Shuhang Zhang, Hongliang Zhang, Boya Di, and Lingyang Song. Cellular uav-to-x communications: Design and optimization for multi-uav networks. IEEE Transactions on Wireless Communications, 18(2):1346-1359, 2019.  
+[15] Haoyang Wang, Xuecheng Chen, Yuhan Cheng, Chenye Wu, Fan Dang, and Xinlei Chen. H-swarmloc: Efficient scheduling for localization of heterogeneous mav swarm with deep reinforcement learning. In Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems, pages 1148-1154, 2022. 1  
+[16] Zhuozhu Jian, Zejia Liu, Haoyu Shao, Xueqian Wang, Xinlei Chen, and Bin Liang. Path generation for wheeled robots autonomous navigation on vegetated terrain. IEEE Robotics and Automation Letters, 9(2):1764-1771, 2023. 1  
+[17] Kaddour Messaoudi, Abdullah Baz, Omar Sami Oubbati, Abdelrezak Rachedi, Tahar Bendouma, and Mohammed Atiquzzaman. Ugv charging stations for uav-assisted aoi-aware data collection. IEEE Transactions on Cognitive Communications and Networking, 10(6):2325-2343, 2024.  
+[18] jianzhuozhu, Xuran Pu, Jianjie Fang, Zhiyuan Deng, Xueqian Wang, and Xinlei Chen. A large language model-driven heterogeneous air-ground rescue swarm. In ICLR 2025 Workshop on Embodied Intelligence with Large Language Models In Open City Environment, 2025. 1  
+[19] Lei Yu, Bishan Wang, Jingwei He, Gui-Song Xia, and Wen Yang. Single image deraining with continuous rain density estimation. IEEE Transactions on Multimedia, 25:443-456, 2021. 2  
+[20] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5896-5905, 2023. 2  
+[21] Sotiris Karavarsamis, Ioanna Gkika, Vasileios Gkitsas, Konstantinos Konstantoudakis, and Dimitrios Zarpalas. A survey of deep learning-based image restoration methods for enhancing situational awareness at disaster sites: the cases of rain, snow and haze. Sensors, 22(13):4707, 2022.  
+[22] Jiehui Huang, Zhenchao Tang, Xuedong He, Jun Zhou, Defeng Zhou, and Calvin Yu-Chian Chen. Progressive network based on detail scaling and texture extraction: A more general framework for image deraining. Neurocomputing, 568: 127066, 2024. 2  
+[23] Qinyi Wang, Yexin Zhang, Junsong Yuan, and Yilong Lu. Space-time event clouds for gesture recognition: From rgb
+
+cameras to event cameras. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1826-1835. IEEE, 2019. 2  
+[24] Yusuke Sekikawa, Kosuke Hara, and Hideo Saito. Eventnet: Asynchronous recursive event processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3887-3896, 2019. 2  
+[25] Yin Bi, Aaron Chadha, Alhabib Abbas, Eirina Courtssolatze, and Yiannis Andreopoulos. Graph-based spatio-temporal feature learning for neuromorphic vision sensing. IEEE Transactions on Image Processing, 29:9084-9098, 2020. 2  
+[26] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12371-12381, 2022. 2  
+[27] Alexander Kugele, Thomas Pfeil, Michael Pfeiffer, and Elisabetta Chicca. Hybrid snn-ann: Energy-efficient classification and object detection for event-based vision. In DAGM German Conference on Pattern Recognition, pages 297–312. Springer, 2021. 2  
+[28] Yannick Schnider, Stanisław Wozniak, Mathias Gehrig, Jules Lecomte, Axel Von Arnim, Luca Benini, Davide Scaramuzza, and Angeliki Pantazi. Neuromorphic optical flow and real-time implementation with event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4129-4138, 2023.  
+[29] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2641-2651, 2021. 2, 3  
+[30] Alejandro Hernández Díaz, Rebecca Davidson, Steve Eckersley, Christopher Paul Bridges, and Simon J Hadfield. Emamba: Using state-space-models for direct event processing in space situational awareness. In IAA Conference on AI in and for Space (SPAICE), 2024. 2, 3  
+[31] Hongwei Ren, Yue Zhou, Jiadong Zhu, Haotian Fu, Yu long Huang, Xiaopeng Lin, Yuetong Fang, Fei Ma, Hao Yu, and Bojun Cheng. Rethinking efficient and effective point-based networks for event camera classification and regression: Eventmamba. arXiv preprint arXiv:2405.06116, 2024. 2, 4  
+[32] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3  
+[33] Hongzhong Tang, Ling Zhu, Dongbo Zhang, and Xiang Wang. Single image rain removal model using pure rain dictionary learning. IET Image Processing, 13(10):1797-1804, 2019. 2  
+[34] Hong Wang, Qi Xie, Qian Zhao, and Deyu Meng. A model-driven deep neural network for single image rain removal. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3100-3109, 2020.  
+[35] Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE Transactions on Image Processing, 24(9):2658-2670, 2015.
+
+[36] Yi Chang, Luxin Yan, and Sheng Zhong. Transformed low-rank model for line pattern noise removal. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2  
+[37] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. arXiv preprint arXiv:2108.10257, 2021. 2  
+[38] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 2  
+[39] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14821-14831, 2021. 2  
+[40] Yueyi Zhang, Jin Wang, Wenming Weng, Xiaoyan Sun, and Zhiwei Xiong. Egvd: Event-guided video deraining. ArXiv, abs/2309.17239, 2023. 2  
+[41] Shangquan Sun, Wenqi Ren, Jingzhi Li, Kaihao Zhang, Meiyu Liang, and Xiaochun Cao. Event-aware video deraining via multi-patch progressive learning. IEEE Transactions on Image Processing, 32:3040-3053, 2023. 2  
+[42] Chengjie Ge, Xueyang Fu, Peng He, Kunyu Wang, Chengzhi Cao, and Zheng-Jun Zha. Neuromorphic event signal-driven network for video de-raining. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1878-1886, 2024. 2  
+[43] Peiqi Duan, Zihao W Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos K Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 8261-8275, 2021. 2  
+[44] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6), 2020.  
+[45] Yaoyi Chen, Yujie Huang, Feiqiang Li, Xiaoyang Zeng, Wenhong Li, and Mingyu Wang. Denoising method for dynamic vision sensor based on two-dimensional event density. In 2023 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-4, 2023. 2  
+[46] Pei Zhang, Zhou Ge, Li Song, and Edmund Y. Lam. Neuro-morphic imaging with density-based spatiotemporal denoising. IEEE Transactions on Computational Imaging, 9:530-541, 2023. 2  
+[47] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270-12279, 2019. 2  
+[48] Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In
+
+Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022-8031, 2019.  
+[49] Wenhan Yang, Robby T Tan, Shiqi Wang, Yuming Fang, and Jiaying Liu. Single image deraining: From model-based to data-driven and beyond. IEEE Transactions on pattern analysis and machine intelligence, 43(11):4059-4077, 2020.  
+[50] Hong Wang, Zongsheng Yue, Qi Xie, Qian Zhao, Yefeng Zheng, and Deyu Meng. From rain generation to rain removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14791-14801, 2021. 2  
+[51] Jin Wang, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Unsupervised video deraining with an event camera. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10797-10806, 2023. 2  
+[52] Alexander Andreopoulos, Hirak J. Kashyap, Tapan K. Nayak, Arnon Amir, and Myron D. Flickner. A low power, high throughput, fully event-based stereo system. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7532-7542, 2018. 2  
+[53] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3  
+[54] Fernando Cladera, Anthony Bisulco, Daniel Kepple, Volkan Isler, and Daniel D. Lee. On-device event filtering with binary neural networks for pedestrian detection using neuromorphic vision sensors. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3084-3088, 2020. 2  
+[55] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5819-5828, 2024. 3  
+[56] Simone Undri Innocenti, Federico Becattini, Federico Pernici, and Alberto Del Bimbo. Temporal binary representation for event-based action recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 10426-10432, 2021. 3  
+[57] Dayong Ren, Zhe Ma, Yuanpei Chen, Weihang Peng, Xiaode Liu, Yuhan Zhang, and Yufei Guo. Spiking pointnet: Spiking neural networks for point clouds. In Advances in Neural Information Processing Systems, pages 41797-41808. Curran Associates, Inc., 2023. 3  
+[58] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3  
+[59] Daniel Gehrig and Davide Scaramuzza. Low latency automotive vision with event cameras. 2024. 3  
+[60] Jiangrong Shen, Qi Xu, Jian K Liu, Yueming Wang, Gang Pan, and Huajin Tang. Esl-snns: An evolutionary structure learning strategy for spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 86-93, 2023. 3
+
+[61] Zou Zhen, Yu Hu, and Zhao Feng. Freqmamba: Viewing mamba from a frequency perspective for image deraining. arXiv preprint arXiv:2404.09476, 2024. 3, 5  
+[62] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European conference on computer vision, pages 222-241. Springer, 2024.  
+[63] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3  
+[64] Shangquan Sun, Wenqi Ren, Juxiang Zhou, Jianhou Gan, Rui Wang, and Xiaochun Cao. A hybrid transformer-mamba network for single image deraining. arXiv preprint arXiv:2409.00410, 2024. 3  
+[65] Yufeng Li, Chuanlong Xie, and Hongming Chen. Multiscale representation for image deraining with state space model. Signal, Image and Video Processing, 19(1):183, 2025. 3  
+[66] Bin Jiang, Bo Xiong, Bohan Qu, M Salman Asif, You Zhou, and Zhan Ma. Edformer: Transformer-based event denoising across varied noise levels. In European Conference on Computer Vision, pages 200-216. Springer, 2024. 4, 6, 7, 8  
+[67] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 4  
+[68] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 4  
+[69] Jiahao Qin and Feng Liu. Mamba-spike: Enhancing the mamba architecture with a spiking front-end for efficient temporal data processing. arXiv preprint arXiv:2408.11823, 2024. 4  
+[70] Kang Zeng, Hao Shi, Jiacheng Lin, Siyu Li, Jintao Cheng, Kaiwei Wang, Zhiyong Li, and Kailun Yang. Mambamos: Lidar-based 3d moving object segmentation with motion-aware state space model. In ACM International Conference on Multimedia (MM), 2024. 4, 8  
+[71] Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099-9117. PMLR, 2022. 4  
+[72] Haotian Zhang, Keyan Chen, Chenyang Liu, Hao Chen, Zhengxia Zou, and Zhenwei Shi. Cdmamba: Remote sensing image change detection with mamba. arXiv preprint arXiv:2406.04207, 2024. 4  
+[73] Lin Liu, Jianzhuang Liu, Shanxin Yuan, Gregory Slabaugh, Aleš Leonardis, Wengang Zhou, and Qi Tian. Wavelet-based dual-branch network for image demoiring. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16, pages 86–102. Springer, 2020. 5  
+[74] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier learning integration with state space models for image deraining. arXiv preprint arXiv:2405.19450, 2024. 5
+
+[75] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13919-13929, 2021. 5  
+[76] Ji Luo, Zijian Xiao, Zuxin Li, Xuecheng Chen, Chaopeng Hong, Xiao-Ping Zhang, and Xinlei Chen. Smartspr: A physics-informed mobile sprinkler scheduling system for reducing urban particulate matter pollution. IEEE Transactions on Mobile Computing, 2025. 5  
+[77] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5  
+[78] Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6286-6295, 2018. 5  
+[79] Maxime Tremblay, Shirsendu Sukanta Halder, Raoul De Charette, and Jean-François Lalonde. Rain rendering for evaluating and improving robustness to bad weather. International Journal of Computer Vision, 129:341-360, 2021. 5  
+[80] Adobe Inc. Adobe after effects software, 2023. 5  
+[81] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrio, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 5  
+[82] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. A  $128 \times 128$  1.5 IEEE Journal of Solid-State Circuits, 48(3): 827-838, 2013. 6, 7  
+[83] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):785-795, 2023. 6, 7  
+[84] Alireza Khodamoradi and Ryan Kastner.  $o(n)\mathrm{o}(n)$ -space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Transactions on Emerging Topics in Computing, 9(1):15-23, 2021. 6, 7, 8  
+[85] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 2020. 6, 7, 8  
+[86] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. arXiv preprint arXiv:2003.08282, 2020. 6, 7, 8  
+[87] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. 6, 7, 8  
+[88] Xuecheng Chen, Haoyang Wang, Yuhan Cheng, Haohao Fu, Yuxuan Liu, Fan Dang, Yunhao Liu, Jinqiang Cui, and Xinlei Chen. Ddl: Empowering delivery drones with large-scale urban sensing capability. IEEE Journal of Selected Topics in Signal Processing, 2024. 8
+
+[89] Xuecheng Chen, Zijian Xiao, Yuhan Cheng, Chen-Chun Hsia, Haoyang Wang, Jingao Xu, Susu Xu, Fan Dang, Xiao-Ping Zhang, Yunhao Liu, et al. Soscheduler: Toward proactive and adaptive wildfire suppression via multi-uav collaborative scheduling. IEEE Internet of Things Journal, 11(14): 24858-24871, 2024. 8

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84a528a84bc42a0e57d89b963ea65af0ad7a94c1d468a25239172e27d2d58a60
+size 1182527

ICCV/2025/PRE-Mamba_ A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f657feb38c33e289882e83d2e5072ef1150c1428c299a16e231788ec4d5a23b2
+size 440705

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/fd4810f3-d689-4b89-b41b-90b3d1713c37_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84d65c4bc51d3d604c642e4f56976dafa1c439d7c51a349860e3e46b4dd81701
+size 85801

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/fd4810f3-d689-4b89-b41b-90b3d1713c37_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:136d023a4a827e0783e618ed5a7d636666598e95d99746f54b80421d496a8f19
+size 112360

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/fd4810f3-d689-4b89-b41b-90b3d1713c37_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57c468db17e8b0ce6a0b78bd5496d9cd0c71bb814fc9f6b58856820ea188e2fd
+size 1221554

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/full.md

@@ -0,0 +1,345 @@
+# PRIMAL: Physically Reactive and Interactive Motor Model for Avatar Learning
+
+Yan Zhang $^{1}$ , Yao Feng $^{1,3}$ , Alpár Cseke $^{1}$ , Nitin Saini $^{1}$ , Nathan Bajandas $^{1}$ , Nicolas Heron $^{1}$ , Michael J. Black $^{2}$
+
+$^{1}$ Meshcapade,  $^{2}$ Max Planck Institute for Intelligent Systems, Tübingen  $^{3}$ Stanford University
+
+![](images/a90aba84fc4f6148b2040aa2dc13b987e48426289058eac7a53af0ed5154c204.jpg)  
+Figure 1. PRIMAL is a novel generative real-time 3D character animation system that works in Unreal Engine. The avatar reacts to induced impulses promptly and naturally (top). After efficient adaptation, the avatar can be pulled to chase a "magnet" (middle). We also personalize the avatar's movements based on a tiny mocap dataset, captured by Mocapade3.0 [19] from cellphone videos (bottom). As a result, we can control the avatar with discrete commands and continuous signals. Without any control signal, or external perturbations, the avatar moves autonomously in the 3D space without end. PRIMAL is purely data driven; no physical simulation is used.
+
+# Abstract
+
+We formulate the motor system of an interactive avatar as a generative motion model that can drive the body to move through 3D space in a perpetual, realistic, controllable, and responsive manner. Although human motion generation has been extensively studied, many existing methods lack the responsiveness and realism of real human movements. Inspired by recent advances in foundation models, we propose PRIMAL, which is learned with a two-stage paradigm. In the pretraining stage, the model learns body movements from a large number of sub-second motion segments, providing a generative foundation from which more complex motions are built. This training is fully unsupervised without annotations. Given a single-frame initial state during inference, the pretrained model not only generates unbounded, realistic, and controllable motion, but also enables the avatar to be responsive to induced impulses in real time. In the adaptation phase, we employ a novel ControlNet-like adaptor to fine-tune the base model effi
+
+ciently, adapting it to new tasks such as few-shot personalized action generation and spatial target reaching. Evaluations show that our proposed method outperforms state-of-the-art baselines. We leverage the model to create a real-time character animation system in Unreal Engine that feels highly responsive and natural.
+
+# 1. Introduction
+
+There has been rapid progress in 3D human motion generation conditioned on text or existing motion sequences [32, 34, 39, 41, 42, 44, 68, 70, 73, 84, 88]. Despite the realism of such approaches, they are not good models of how real humans behave, in that they are not interactive. Autonomous characters need to respond in real-time to commands and environmental inputs such as external forces. To feel "alive", such an agent must be continuously on-line and
+
+reactive. It must move on its own yet be able to respond to commands. We introduce such an agent using a two-stage approach, in which we first train a low-level motor system from motion capture (mocap) data such that it mimics the dynamics of human motion. This provides a platform upon which to build more complex interactive behaviors in a second stage. This two-stage approach provides a base "motor system" that implicitly models the dynamics of human motion, generalizes beyond the training data, and supports the training of more complex behaviors.
+
+Our approach, called PRIMAL (Physically Reactive and Interactive Motor model for Avatar Learning), models low-level motor control using an autoregressive diffusion model that is trained to capture human motion movement over short time intervals. Without any physics simulation, we observe that this approach is able to model physically realistic movements, including ground contact, even though we never explicitly model or label contact. This provides a data-driven approach to learning avatar motor control that appears physically realistic without physics simulation.
+
+We make two key observations that ground our approach. First, existing autoregressive motion models usually predict a future motion conditioned on a past motion. Since a motion sequence implicitly contains semantic information, this ensures the predicted motion is semantically related to the past. However, it raises the risk that the model overfits to the semantics in the training data when the training samples are limited (as they are with existing mocap datasets). This limits generalization, resulting in a model that does not make a good foundation for general motor control. Second, human motion is dominated by physics over short time intervals and by semantics over long ones. We use different approaches to model each of these. We observe that physics plays a critical role in time intervals on the order of 0.5 sec, or 15 frames at 30fps. On longer time horizons, higher level goals dominate.
+
+In PRIMAL, the first stage models this short-term motion: given an initial condition of the body, including its joint locations and velocities at a single time instant, we predict the motion of the next half second. This assumes that, over a short time horizon, the motion of the future is independent of the past. With this assumption, we have vast amounts of motion capture data with which to train the model; every segment of 15 frames is available for training. We demonstrate that training our generative model in this way mimics the dynamics of human motion without doing any physics simulation. Note that, while we do not explicitly model foot-ground contact, the model learns to produce movements with physically plausible contact.
+
+Motivated by the representation power of diffusion models [26, 66, 67], the first stage is implemented as a diffusion model, which is trained on diverse half-second motion segments in a fully unsupervised manner. The trained model
+
+not only generates realistic and autonomous motions in 3D space, but also makes it straightforward to control the avatar in real-time. For example, we simply reset any joint velocities to new values and the model will adapt its future motion appropriately. We can also drive the avatar to reach desired speeds and directions, using a classifier-based guidance (CBG) [14] approach, for which the CBG gradients are analytically derived.
+
+High-level and long-term behaviors can be composed as a time sequence of consecutive atomic actions. This decouples the learning of movements and semantics, enabling us to learn them individually in a two-stage paradigm with a pretraining stage (motor control) and the adaptation stage (behavior generation). Specifically, we propose a generic adaptation approach based on ControlNet [83] in the second stage. Here, we consider two types of adaptation. One focuses on spatial target reaching and the other on few-shot semantic action generation. To improve the effectiveness of the control signal, we follow [44] to add the control embedding to individual blocks of the transformer.
+
+We demonstrate the benefits of our interactive avatar with several applications. First, we create an interactive 3D environment in Unreal Engine, in which the agent exhibits perpetual motion (see Fig. 1). The agent's movement is controlled with standard game controllers to move in any direction at varied speeds. Unlike traditional game technology, however, all the motions are uniquely generated in real time based on the commands. We introduce external forces ("pulling" with a "magnet" or an impulsive force that "zaps" the character) that dynamically change the agent's behavior (see Supplemental Video). The agent reacts promptly to interventions in a natural way, without any explicit physics-based training, showing that the first stage learns a generic motor model. Second, we demonstrate how our approach enables training a person/style-specific model using a tiny amount of motion capture data captured with a cell phone. The results illustrate how pre-training a foundational movement stage provides a basic "motor system" upon which to build complex behaviors.
+
+In summary, PRIMAL produces naturalistic human motion common to methods trained on mocap data together with the physical plausibility and generalization of physics-based methods. To build a motor system for interactive avatars, we use an autoregressive diffusion model with a two-stage pretrain-and-adaptation learning paradigm. This first stage trains the model to capture human motion over short time scales, providing a reactive model that runs in real time. The second stage uses an efficient and generic approach to adapt the base model to different avatar-related tasks. The interactive nature of PRIMAL effectively "breathes life" into an avatar, making it appropriate for applications in gaming, AR/VR, controllable video generation, and interactive agents.
+
+# 2. Related Work
+
+Offline motion generation. Enabled by large mocap datasets [20, 50, 58], generative motion models have advanced rapidly, including methods focused on text-to-motion [13, 31, 32, 44, 70, 84], human-scene/object interactions [4, 21, 22, 34, 39, 40, 68, 69, 79, 80, 88], spatial control [15, 35, 45, 61, 78, 81, 86], dance generation [42, 43, 75], and more. Specifically, diffusion models [26, 66, 67] are a promising approach for motion generation, due to their representation power and controllability. For example, MDM [70] proposes a diffusion model with the transformer [77] network, which is trained with text-motion paired data [20] and is able to generate high-fidelity motions given text descriptions. OmniControl [81] leverages ControlNet [83] to create a new adaptor, enabling spatial control of text-to-motion. MotionLCM [13] proposes a consistency model for real-time text-to-motion, and implements autoregressive generation as a spatial control task. Most of these approaches are offline methods and have motion length limitations, making them inappropriate for interactive avatar control.
+
+More similar to our approach is OMG [44], which proposes a pretrain-then-finetune strategy for text-to-motion. Its pretraining phase trains foundation models with up to 1 billion parameters only with unlabeled mocap data, and the fine-tune phase leverages a ControlNet-like adaptor with a mixture-of-experts. In contrast to our approach, OMG is offline and their pretraining process focuses on generating motions up to 300 frames long, much longer than our short time horizon. Unlike OMG, our model formulation is autoregressive, which not only generates unbounded and controllable motion in real time, but also makes the avatar reactive to external perturbations.
+
+Physics-based motion generation. A key problem of the data-driven methods above is that they often lack physical plausibility; e.g. feet slide, float, or penetrate the floor, and impacts are not realistic. To address this, many methods have been developed that exploit physical simulation [22, 46, 56, 61, 62, 78]. Hassan et al. [22] employ reinforcement learning (RL) and adversarial imitation learning to produce life-like and physically plausible human-scene interactions. PACER++ [78] generates physically plausible pedestrian locomotion in driving scenarios with fine-grained body part control. Simulation can also be efficiently combined with generative models. For example, PhysDiff [82] projects the generated motion into a physically plausible motion space via physical simulation, which is applied within the reverse diffusion process. CLoSD [71] uses an efficient autoregressive diffusion model, DiP, for planning and applies simulation for tracking, generating unbounded physically plausible motions in 3D. InsActor [62] proposes an hierarchical approach to generate motions given instructions, which contains diffusion-based planning and
+
+simulation-based motion tracking. An advantage of simulation approaches is that they can generalize to adapt to external forces. However, the resulting motions are usually not yet as natural as those from diffusion models. Additionally, simulations tend to learn policies of a narrow set of motor skills, rather than a generic human motion control system.
+
+Autoregressive unbounded motion generation. Autoregressive models are trained to generate future frames conditioned on past frames. GAMMA [86] is a VAE [37] model that generates motion primitives recursively and leverages RL to train a policy for spatial target reaching. WANDR [15] is a conditional VAE that generates motions to reach a 3D target with the wrist joint. DIMOS [88] uses scene-aware policies for motion-primitive generation, enabling the synthesis of continuous human-scene interactions. Recently, diffusion models have been used for autoregressive motion modeling. For example, CAMDM [11] exploits a diffusion model that treats motion style, spatial control, and motion history as separate tokens, producing unbounded motions for real-time character animation. Jiang et al. [33] propose a diffusion model that is conditioned on the past 2 frames, as well as text, scene, etc., to generate human-scene interactions. AMDM [64] uses an autoregressive diffusion model with an MLP denoiser. In addition to generating random motions, policy networks are trained with RL for motion control in various tasks. The DiP model proposed in CLoSD [71] is conditioned on the motion history of 20 frames and text tokens, and generates a 40-frame motion in the future. DART [89] uses an online generation method that has an autoregressive VAE tokenizer and a latent autoregressive diffusion model, modeling motion primitives and performing high-fidelity text-to-motion in real time. Latent space optimization and RL policies are used for motion control on the pre-trained model.
+
+Compared to existing autoregressive diffusion models that generate  $N$  future frames based on  $M$  past frames, our model formulation generates motions based on an initial state of a single frame that includes joint positions and velocities. With this form, PRIMAL learns human movement from unstructured and unlabeled mocap data directly, without the support of physics or annotations. Despite the lack of explicit physics training, we find that visually plausible physical effects are present in the generated motions. In addition, neither motion tokenizers nor scheduled sampling [5] schemes are necessary in our method, leading to a simple and straightforward training workflow.
+
+# 3. Autoregressive Diffusion Motion Models
+
+# 3.1. Preliminaries
+
+SMPL-X [54]. We employ the SMPL-X gender-neutral body model and denote it as  $\mathcal{M}(\beta, r, \varphi, \theta)$ , in which  $\beta \in \mathbb{R}^{16}$  represents the body shape parameters,  $r \in \mathbb{R}^3$  is the
+
+root translation,  $\varphi \in \mathbb{R}^6$  and  $\theta \in \mathbb{R}^{126}$  are the root and joint rotations w.r.t. the continuous 6D representation [91], respectively. Since we focus on the body motion without facial expressions or hand articulations, we extract the 22 joints of the body skeleton and denote their locations and velocities as  $J \in \mathbb{R}^{66}$  and  $\dot{J} \in \mathbb{R}^{66}$ , respectively.
+
+Denoising diffusion probabilistic models (DDPM) [26]. Diffusion models are widely used for tasks like motion generation. In this work, we follow the DDPM formulation [26]. Provided a clean sample  $\mathbf{X}^0\sim p(\mathbf{X}^0)$ , the forward diffusion process gradually adds Gaussian noise to the sample in  $T$  steps, and produces a sequence of samples at different noise levels, i.e.  $\{\mathbf{X}^1,\dots ,\mathbf{X}^t,\dots ,\mathbf{X}^T\}$ . Given  $q(\mathbf{X}^t |\mathbf{X}^{t - 1}) = \mathcal{N}\left(\sqrt{1 - b_t} X^{t - 1},b_t\mathbf{I}\right)$  with  $\{b_{t}\in (0,1)\}_{t = 1}^{T}$  being a set of noise variance scheduler variables, it has
+
+$$
+\mathbf {X} ^ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {X} ^ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {1}
+$$
+
+in which  $\alpha_{t} = 1 - b_{t}$ ,  $\bar{\alpha}_{t} = \prod_{t^{\prime} = 1}^{t}\alpha_{t^{\prime}}$ , and  $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ .
+
+The reverse diffusion process gradually removes noise from  $\mathbf{X}_t$  to recover the original clean sample  $\mathbf{X}_0$ . Following [60, 70], we use a neural network to predict the clean sample in this work, i.e.  $\hat{\mathbf{X}}^0 = G(t,\mathbf{X}^t,\mathbf{c})$ , which takes the diffusion time step, the noisy sample, and a condition as input. This network is trained by minimizing a simple loss [26], given by  $\mathcal{L}_{\mathrm{simple}} =$
+
+$$
+\mathbb {E} _ {\mathbf {X} ^ {0} \sim p \left(\mathbf {X} ^ {0}\right), t \sim U [ 1, T ]} \left[ \left\| \mathbf {X} ^ {0} - G \left(t, \mathbf {X} ^ {t}, \mathbf {c}\right) \right\| ^ {2} \right]. \tag {2}
+$$
+
+During test-time, the reverse diffusion process can be controlled via classifier-based guidance (CBG) [14] and classifier-free guidance (CFG) [25]. With a pre-trained diffusion model and the guidance objective functions, CBG computes the gradients of the objectives at each denoising step, and hence does not support real-time applications in general. On the other hand, CFG trains the conditional and unconditional diffusion models jointly, and guides the reverse diffusion process via weighted averaging, which is usually much more efficient than CBG. With  $G(t,\mathbf{X}^t,\emptyset)$  being the unconditional model, CFG is performed at each denoising step by
+
+$$
+\hat {X} ^ {0} = G (t, \mathbf {X} ^ {t}, \emptyset) + \gamma (G (t, \mathbf {X} ^ {t}, \mathbf {c}) - G (t, \mathbf {X} ^ {t}, \emptyset)), \tag {3}
+$$
+
+in which  $\gamma \in \mathbb{R}^{+}$  is a hyper-parameter. A larger value of  $\gamma$  will increase the influence of the condition. However, invalid results can be produced if it is too large.
+
+# 3.2. Motion Representation
+
+We represent an  $N$ -frame motion segment as a time sequence of states,  $\pmb{X} = \{\pmb{x}_0,\pmb{x}_i,\dots,\pmb{x}_{N - 1}\}_{i = 0}^{N - 1}$ . Each state has  $\pmb{x}_i = (r,\varphi ,\theta ,J,\dot{J})\in \mathbb{R}^{267}$  based on SMPL-X. During training, we use the ground truth body shape  $(\beta)$ . During testing, we can either leverage a ground truth  $\beta$  from the
+
+![](images/e9b3e644ee8bdefa4ea09caa39d05390c699b57816d673141e4d81967f83ce35.jpg)  
+Figure 2. Illustration of our network  $\hat{\pmb{X}}^0 = G(t,\pmb{X}^t,\pmb{x}_0)$ . The tea-green layers contain trainable parameters and the orange blocks are non-learnable operations. The pink squares denote the tokens at individual frames.
+
+test set, or draw it from  $N(0,\mathbf{I})$ , in order to compute the initial state  $\pmb{x}_0$ . During motion generation,  $\beta$  is fixed.
+
+Each motion segment spans 0.5 seconds in this work. This time duration corresponds to the average hang time of human jumps² or nearly half of a walking cycle. This is a time frame during which human motion is dominated more by physical forces than by intentional control.
+
+Following the canonical coordinate system proposed by [86, 87], the body-centric canonical coordinate is located at the pelvis of the first frame, with X,Y, and Z pointing left, up, and forward, respectively. The neural network is trained with canonicalized motion segments.
+
+# 3.3. Neural Networks
+
+Architecture. Following e.g. [11, 64, 71, 89], in this paper the autoregressive diffusion model is defined as a diffusion model that generates  $N$  future frames based on  $M$  past frames. We formulate our denoising network as  $G(t,\mathbf{X}^t,\pmb{x}_0)$  with transformers [55, 77] and illustrate the architecture in Fig. 2. It leverages in-context conditioning, in which the diffusion time embedding and the initial state embedding are first added and then concatenated with the embedding of the noisy motion segment. SiLU [16, 24, 59] is used in the feed-forward layers.
+
+Training. Unlike existing autoregressive diffusion models, e.g. [11, 71, 89], that leverage annotated training data, our model focuses on modeling the body movements and is trained only with unlabeled short mocap segments. Therefore, the data limitation issue is effectively overcome, since we can collect a huge amount of segments from AMASS [50]. In addition, our motion model is directly trained in
+
+the motion space, instead of the latent space as in [32, 89]. No separate tokenizers are trained in our method.
+
+Our training loss is given by
+
+$$
+\mathcal {L} = \mathcal {L} _ {\text {s i m p l e}} + \gamma_ {1} \mathcal {L} _ {F K} + \gamma_ {2} \mathcal {L} _ {F K V}, \tag {4}
+$$
+
+which contains a DDPM simple term (see Eq. (2)), two forward kinematics terms on the joint locations and velocities, respectively, and two weights  $\gamma_{1}$  and  $\gamma_{2}$ . The two forward kinematics losses are given by
+
+$$
+\mathcal {L} _ {F K} = \frac {1}{N} \sum_ {i = 0} ^ {N - 1} \| \mathcal {M} (\beta , \hat {r} _ {i}, \hat {\varphi} _ {i}, \hat {\theta} _ {i}) - J _ {i} \| _ {2} ^ {2}, \tag {5}
+$$
+
+and
+
+$$
+\mathcal {L} _ {F K V} = \frac {1}{N} \sum_ {i = 0} ^ {N - 1} \| \dot {\mathcal {M}} (\beta , \hat {r} _ {i}, \hat {\varphi} _ {i}, \hat {\theta} _ {i}) - \dot {J} _ {i} \| _ {2} ^ {2}, \tag {6}
+$$
+
+in which the predicted motion state at the  $i$ -th frame is denoted by  $\hat{\pmb{x}}_i = (\hat{r}_i, \hat{\varphi}_i, \hat{\theta}_i, \hat{J}_i, \hat{J}_i)$ . The ground truth body shape  $\beta$  is used.
+
+Our model is trained in a fully unsupervised manner, and does not use foot contact labels as in [70, 85]. We think the contact information, which is usually defined by the horizontal velocities of the foot joints, can be automatically learned by minimizing the  $\mathcal{L}_{FKV}$  term. Incorrect annotations can also degrade model performance. Moreover, we find that our model is self-stabilizing after being trained. Therefore, scheduled sampling [5], which can substantially reduce the training speed, is not employed in our training phase. Compared to existing methods such as [11, 45, 64, 70, 71, 86, 88, 89], our training pipeline is simple and straightforward.
+
+# 3.4. Inference
+
+Our motion model learns the distribution of  $p(\mathbf{X}|\mathbf{x}_0)$ . Given an initial state, the model generates a motion segment starting with this initial state, and works repeatedly to produce an arbitrarily long motion sequence. The overall inference approach is summarized in Alg. 1 in Sup. Mat.
+
+In addition, we propose three efficient and effective test-time processing approaches to refine the motion on the fly, i.e. joint re-projection, snapping to the ground, and inertialization blending [8, 27]; see Sup. Mat. for details.
+
+# 3.4.1. Motion Control with Induced Impulses
+
+The model generalizes outside the training data, exhibiting behavior typically seen only with physics simulation. Specifically, we perturb the avatar with novel external impulses by changing the velocity at one or more joints. Using this, we can control the avatar with impulses to generate certain actions in a principled manner. For example, we can generate kicks by giving an upward impulse to the joints on
+
+the leg, or generate runs by pushing the avatars from the back.
+
+Impulse induction is implemented by perturbing the initial velocities before generating a new segment. We find that such impulses can be defined to reliably generate specific actions; see Sup. Mat. Tab. A1. Future work should explore the automated discovery of impulses for controlling behavior.
+
+# 3.4.2. Motion Control with Classifier-based Guidance
+
+Another basic level of controllability needed for an interactive avatar is the ability to move in a given direction at a given speed. To enable that, we propose two intuitive classifier-based guidance (CBG) loss terms to control the movement and the facing direction of the avatar.
+
+The movement guidance loss is based on the average velocity of all 22 joints in the half-second time horizon, which is given by
+
+$$
+\mathcal {L} _ {\text {m o v e}} = \left\| \left(\frac {1}{2 2 N} \sum_ {i = 0} ^ {N - 1} \sum_ {j = 1} ^ {2 2} \hat {J}\right) - \boldsymbol {v} _ {\text {g o a l}} \right\| ^ {2}, \tag {7}
+$$
+
+in which  $\pmb{v}_{\text{goal}} \in \mathbb{R}^3$  denotes the target mean speed.
+
+The facing guidance loss is based on the last frame of the motion segment. Given the avatar's normalized facing directional vector  $\mathbf{z}$  and a target facing unit vector  $\mathbf{r}_{\text{goal}}$ , the guidance loss is given by
+
+$$
+\mathcal {L} _ {\text {f a c i n g}} = \left\| \boldsymbol {z} - \boldsymbol {r} _ {\text {g o a l}} \right\| ^ {2}, \tag {8}
+$$
+
+in which  $z$  is computed from the predicted joint locations.
+
+In order to obtain real-time performance, we derive the analytical gradients of these terms instead of using autodifferentiation. Note that we control motions via manipulating  $J$  or  $\dot{J}$  in the motion state  $\pmb{x}_i$ . The SMPL-X parameters, i.e.  $(r,\varphi ,\theta)$ , are not directly controlled.
+
+# 4. Model Adaptation
+
+Since it takes 3-4 days to train our autoregressive diffusion model to obtain high-quality results, we regard it as a base model for task-specific adaptation; this reduces the costs of data and computation. Therefore, we do not need to train from scratch to solve new tasks. The effectiveness of this approach has been widely reported in recent advances of foundation models [30, 52, 53, 63, 72, 83].
+
+In this work, we take advantage of ControlNet [83] and propose a generic approach for task adaptation, which is illustrated by Fig. 3. For each adaptor, the control embedding and the output from the previous transformer block are added first and then fed into a trainable copy of the current transformer block. The adaptation process only updates the parameters in the adaptors, while keeping the base model parameters frozen. The training loss is identical to Sec. 3.3.
+
+![](images/b9016faf9dca2d73f5e4f3ed04d3d8cc20e575b485045e5999b2166f21271a90.jpg)  
+Figure 3. Illustration of our adaptation method. Given a base model, we introduce a ControlNet-like adaptor at each individual transformer block. The control signal  $y$  is a generic notation. We adapt the base model for semantic action generation and spatial target reaching. The scale operation manipulates the impact of the control signal, corresponding to  $\gamma$  in Eq. (3) of CFG.
+
+This architecture is different from the ControlNets applied in [81, 90], in which the control embedding is just input once to the adaptor. Our transformer block-wise adaptation is similar to OMG [44]. However, their mixture-of-experts module is specifically designed to fuse the text tokens via cross attention, whereas ours is a generic framework that fuses a generic control embedding via addition.
+
+Few-shot semantic action generation. Given a small set of personalized mocap sequences with action labels, we adapt our base model to generate actions given the labels. In this case, the control signal  $y$  shown in Fig. 3 is an integer action label, and an embedding layer maps this label to a latent vector. After adaptation, our model is able to generate infinite personalized movements based on the action labels, and natural transitions between actions, beyond what this small dataset provides. We can control the impact of the action embedding via classifier-free guidance (CFG). Simultaneously, we can also perturb the joint velocities to generate natural reactions, and control the movement continuously via the CBG method proposed in Sec. 3.4.2.
+
+Spatial target reaching. In this scenario, the control signal  $y$  becomes a 2D target location for the avatar to reach. During finetuning, the pelvis XZ coordinate of the last frame of the canonicalized motion segment is treated as the target. Following [15, 69], we re-scale the target location via  $2 \cdot (1 - e^{-\|x\|_2}) \cdot \frac{x}{\|\pmb{x}\|_2}$ , in order to make the model generalize to targets in extreme scenarios. In addition, we use a linear layer to map the scaled target location to a latent variable. After training, target-reaching behavior can be synthe
+
+sized by specifying the target locations at test time. Since the motion is generated autoregressively, one can control the avatar's motion direction by controlling the target.
+
+# 5. Experiments
+
+# 5.1. Evaluations on Body Motion Learning
+
+A model that learns human body motion well should have two capabilities: 1) It can generate high-fidelity, realistic motions, given an arbitrary initial state; 2) The model's responses to inputs should be natural and human-like. Thus, we have two sets of experiments to measure motion realism and responsiveness, respectively. Below, we denote our PRIMAL architecture in Fig. 2 as InContext. In addition, we compare it with another architecture that exploits adaptive layer normalization [55, 57], denoted as AdaLN, which is illustrated in Sup. Mat.
+
+# 5.1.1. Motion Realism
+
+Datasets and metrics. We draw 80 and 160 initial states from the HumanEva [65] and SFU [1] subsets of AMASS, respectively. We use the gender-neutral SMPL-X body and replace the ground truth shapes with random samples from  $\mathcal{N}(0,I)$ , testing generalization to unseen identities. From an initial state, we generate a 240-frame motion, ie. 8s.
+
+We evaluate the motion realism w.r.t. the following terms. 1) Averaged skating ratio (ASR): Following [86-88], we define foot skating as occurring when both foot joints are in contact with the ground (within a tolerance of  $5\mathrm{cm}$ ) and their horizontal velocities are both above  $15\mathrm{cm/s}$ . We report the averaged ratio of skating frames. 2) Averaged non-collision ratio (ANCR): Similar to [82], we compute the lowest body joint at each frame, and define a collision when the lowest body joint is deeper than  $5\mathrm{cm}$  below the ground plane. 3) Averaged neural Riemannian distance (ARD): We employ NRDF [23] to measure the distance of the generated body pose to the pretrained natural pose manifold of AMASS. A smaller value indicates that the generated pose is more natural. 4) Baseline preference ratio (BPR): For each method we render 50 videos and perform perceptual studies to visually compare InContext with individual baselines; we report the percentage of participants who prefer the baseline results. See Sup. Mat. for more details.
+
+Comparisons to state-of-the-art. We compare our methods with two groups of baselines. The off-the-shelf group includes DiP [71], CLoSD [71] and MotionLCM [13]. The "our implemented" group in Tab. 1 includes versions of DiP and DART [89] where the motion representations and pretraining strategies are identical to ours. For details of how we train and fairly compare the methods, see Sup. Mat.
+
+Table 1 shows that all methods avoid ground collision. Our PRIMAL-InContext model outperforms data-driven
+
+![](images/b668a50d43ff8969276a0739155bc58213f1b272a367966435eaf5d35eab8aad.jpg)  
+Figure 4. Illustration of three random motion generations (one per row). Given different initial states, the model is able to generate high-fidelity future motions of arbitrary lengths in real time. The numbers denote the frame indices. See Supplemental Video.
+
+<table><tr><td>Methods</td><td>ASR ↓</td><td>ANCR ↑</td><td>ARD ×10-3↓</td><td>BPR ↓</td></tr><tr><td colspan="5">off-the-shelf</td></tr><tr><td>DiP [71]</td><td>0.067</td><td>1.0</td><td>7.373</td><td>20.81%</td></tr><tr><td>CLoSD [71]</td><td>0.019</td><td>1.0</td><td>1.994</td><td>9.17%</td></tr><tr><td>MotionLCM [13]</td><td>0.149</td><td>1.0</td><td>-</td><td>-</td></tr><tr><td colspan="5">our implemented</td></tr><tr><td>DART [89]</td><td>0.072</td><td>1.0</td><td>0.077</td><td>34.88%</td></tr><tr><td>DiP [71]</td><td>0.065</td><td>1.0</td><td>0.174</td><td>41.15%</td></tr><tr><td colspan="5">our methods</td></tr><tr><td>InContext (ours)</td><td>0.024</td><td>1.0</td><td>0.074</td><td>-</td></tr><tr><td>AdaLN (ours)</td><td>0.036</td><td>1.0</td><td>0.147</td><td>-</td></tr></table>
+
+belines w.r.t. foot skating, despite not explicitly training for this. Note that PRIMAL directly outputs SMPL-X parameters, whereas DiP and CloSD need to fit SMPL-X to joint locations. SMPL-X is needed for the ARD and BPR scores and the fit step could negatively impact their scores. Note that InContext is visually preferred to all baselines (ie. low BPR) suggesting the realism of the generated motions. Figure 4 shows generated motions given different initial states; see the Supplemental Video.
+
+Furthermore, we can see that InContext outperforms AdaLN w.r.t. motion realism. We also find InContext runs slightly faster in our trials. Therefore, we will only use InContext in the following experiments. More details and ablation studies are in Sup. Mat.
+
+# 5.1.2. Motion Responsiveness
+
+The ability to respond to external input is necessary for an avatar to feel "alive" (see Fig. 1). We draw 45 random initial states from SFU [1], and replace the ground truth body shape with a random value as before. We compare our In-Context method with our implemented DiP and DART diffusion models, and PRIMAL-InContext-8f, which has an 8-frame time horizon as in DART.
+
+Table 1. Motion Realism: Comparison with SOTA. Best results are highlighted in boldface. We only compare "InContext" with baselines for the perceptual study (BPR). See text.  
+
+<table><tr><td>Methods</td><td>ASR ↓</td><td>ANCR ↑</td><td>ARD ×10-3↓</td><td>BPR ↓</td></tr><tr><td>DART diffusion [89]</td><td>0.159</td><td>0.961</td><td>2.259</td><td>21.29%</td></tr><tr><td>DiP [71]</td><td>0.126</td><td>0.995</td><td>0.645</td><td>10.50%</td></tr><tr><td>InContext-8f</td><td>0.172</td><td>0.995</td><td>7.903</td><td>30.66%</td></tr><tr><td>InContext</td><td>0.088</td><td>1.0</td><td>0.511</td><td>-</td></tr></table>
+
+Table 2. Results of reactions to induced impulses. The evaluation metrics are the same as in Tab. 1. Best results are in boldface.  
+
+<table><tr><td>Methods</td><td>ASR ↓</td><td>ARD ×10-3↓</td><td>R-precision ↑</td></tr><tr><td>Finetuning</td><td>0.253</td><td>0.060</td><td>0.87</td></tr><tr><td>OmniControlNet [81]</td><td>0.299</td><td>0.063</td><td>0.90</td></tr><tr><td>Ours</td><td>0.215</td><td>0.064</td><td>0.96</td></tr></table>
+
+Table 3. Results of semantic action generation. R-precision indicates classification accuracy. See more results in Sup. Mat.
+
+Reactions to induced impulses. Here, we perturb the velocities for subsets of the joints and quantify the resulting motions generated by the agent. Depending on the initial velocities, we can make the agent generate kicks, forward runs, back flips, and forward rolls; see Sup. Mat. for the intuitive principles. With PRIMAL, we only perturb the first frame, but apply the velocities to all history frames of DiP and DART to improve their responsiveness. To measure the motion realism, we generate a 3-second motion for each starting state, leading to 225 sequences per method, and compute ASR, ANCR, and ARD. In addition, we render 30 videos per action for a perceptual study that tests how well the models align with human intuitions; see the Sup. Mat. for details. Table 2 shows that InContext outperforms the prior art. Figure 5 visualizes some results; see Supplementary Video. We see that the avatars have realistic reactions to the velocity interventions.
+
+CBG-based control. The CBG method proposed in Sec. 3.4.2 provides control of the movement direction and speed of the agent. In Context significantly outperforms the baselines on this task - see Sup. Mat. for more details.
+
+# 5.2. Evaluations on Model Adaptation
+
+We refer the reader to the Sup. Mat. for the model adaptation experiment details, which include few-shot semantic action generation and spatial target reaching. We evaluate two approaches: 1) We add the control embedding (i.e. the embedding of  $y$  in Fig. 3) to the time embedding, and fine-tune the pretrained model directly. 2) We implement the motion ControlNet proposed in [81, 90], in which the control signal is fused to the input of the ControlNet only once. We denote them as finetuning and OmniControlNet, respectively. Overall, we find the ControlNet approach can adapt the PRIMAL base model to novel tasks efficiently and effectively. Compared to OmniControlNet, we find that
+
+![](images/466ccbb2e9aa35ba718d70d8edd0c770d5f410ac7efa544ac5ffa7aaf98afa10.jpg)  
+Figure 5. Avatar reactions to induced impulses. Each row shows a motion. The pale orange arrows in the initial frame denote the perturbed velocities, and bars in the future frames denote the generated joint locations and velocities. Darker color means smaller velocity norms.
+
+![](images/2bc5e34f8ae38a14da5bd83bbf34b42f40204e016a17e028ccb9381d10f951a1.jpg)  
+Figure 6. Snapshots of personalized mocap dataset, captured by MoCapade3.0 [19] from cellphone videos.
+
+our proposed adaptor enables the avatar to respond more quickly to the control signal.
+
+For few-shot semantic action generation, we use Mocapade3.0 [19] to capture 5 classes of actions from video (Fig. 6). Although the dataset is only about 3 minutes in total, it is sufficient to fine-tune the model. Quantitative results are shown in Tab. 3, while visualizations are in Fig. 1. We observe that all these PRIMAL-based methods are effective. The ControlNet approach allows adaptation of the base model to a small set of captured motions, maintaining the controllability and responsiveness of the original model, while being able to mimic the new data.
+
+# 6. Conclusion
+
+To create a motor model for interactive agents, we propose a new form of autoregressive diffusion model that generates
+
+perpetual, controllable, realistic, and responsive motions in real time. Following recent advances in foundation models, our training paradigm has two phases, i.e. 1) unsupervised pretraining to learn body movements, and 2) adaptation to novel tasks. Experimental results show that our model learns motions effectively, adapts to external velocity inputs, and can be efficiently adapted to few-shot personalized action generation and spatial target reaching. We demonstrate the advantages with a real-time character animation system built in Unreal Engine.
+
+Limitations and future work. First, while most motions exhibit perceptually physical realism, non-physical motions sometimes occur, e.g. when getting up after falling, the avatar sometimes "flies" up. The method generalizes to falls that it has never seen, but more training data of this type would be helpful. We clarify that we do not ensure physical realism. Rather we generate motions that generally look physically realistic. Second, PRIMAL has limited types of control inputs. Extending this to have more natural inputs through text or speech should be explored. For example, we can modify our ControlNet to fuse text and motion features, and fine-tune it on text-motion paired datasets, e.g. BA-BEL [58] and HumanML3D [20]. Third, the agent has no awareness of its environment and does not know about obstacles. Learning to condition on the scene is future work. Finally, the model could eventually replace traditional motion graphs and motion matching (cf. [28, 29, 38]) in video games, however, despite running in real time (i.e. computation takes less than 0.5 second), the computational cost is still too much for this application.
+
+Acknowledgments. We sincerely thank Radu Alexandru Rosu, Yu Sun, and Aman Shenoy for the Rust-based SMPL-X rendering. We are also heartily grateful to Guy Tevet for the help on CLoSD [71]. Moreover, we thank fruitful discussions with Muhammed Kocabas and Nikolas Hesse. While MJB is a co-founder and Chief Scientist at Meshcapade, it is acknowledged that his research in this project was performed solely at, and funded solely by, the Max Planck Society. This work was done while YF was an employee of Meshcapade.
+
+# References
+
+[1] SFU motion capture database. http://mocap.cs.sfu.ca/, 2000. 6, 7, 16  
+[2] Advanced Computing Center for the Arts and Design. AC-CAD MoCap Dataset. 18  
+[3] Ijaz Akhter and Michael J. Black. Pose-conditioned joint angle limits for 3D human pose reconstruction. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) 2015, 2015. 18  
+[4] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 3  
+[5] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 3, 5  
+[6] RD Bishop and JG Hay. Basketball: the mechanics of hanging in the air. Medicine and science in sports, 11(3): 274-277, 1979. 4  
+[7] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2017. 18  
+[8] David Bollo. Inertialization: High-performance animation transitions in gears of war, 2018. GDC 2018. 5, 15  
+[9] Carnegie Mellon University. CMU MoCap Dataset. 18  
+[10] Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems, 37:24081-24125, 2025. 17  
+[11] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3, 4, 5, 17, 18  
+[12] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 539-546. IEEE, 2005. 16
+
+[13] Wenxun Dai, Ling-Hao Chen, Jingbo Wang, Jinpeng Liu, Bo Dai, and Yansong Tang. Motionlcm: Real-time controllable motion generation via latent consistency model. In European Conference on Computer Vision, pages 390-408. Springer, 2024. 3, 6, 7, 15  
+[14] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 4  
+[15] Markos Diomataris, Nikos Athanasiou, Omid Taheri, Xi Wang, Otmar Hilliges, and Michael J Black. Wandr: Intention-guided human motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 927-936, 2024. 3, 6  
+[16] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107:3-11, 2018. 4, 14, 18  
+[17] Nima Ghorbani and Michael J. Black. SOMA: Solving optical marker-based mocap automatically. In Proc. International Conference on Computer Vision (ICCV), pages 11117-11126, 2021. 18  
+[18] Saeed Ghorbani, Kimia Mahdaviani, Anne Thaler, Konrad Kording, Douglas James Cook, Gunnar Blohm, and Nikolaus F. Troje. MoVi: A large multipurpose motion and video dataset. arXiv preprint arXiv: 2003.01888, 2020. 18  
+[19] Meshcapade GmbH. Mocapade3.0. https://memeshcapade.com/, 2025. Accessed: March 2025. 1, 8, 16  
+[20] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 3, 8, 15  
+[21] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 3  
+[22] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3  
+[23] Yannan He, Garvita Tiwari, Tolga Birdal, Jan Eric Lenssen, and Gerard Pons-Moll. Nrdf: Neural riemannian distance fields for learning articulated pose priors. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 6  
+[24] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4, 14, 18  
+[25] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4  
+[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 4  
+[27] Daniel Holden. Spring-it-on: The game developer's spring-roll-call, 2021. 5, 15  
+[28] Daniel Holden, Jun Saito, Taku Komura, and Thomas Joyce. Learning motion manifolds with convolutional autoencoders. In SIGGRAPH Asia 2015 Technical Briefs. 2015. 8
+
+[29] Daniel Holden, Ouussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. TOG, 2020. 8  
+[30] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5  
+[31] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing, 2024. 3  
+[32] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 1, 3, 5  
+[33] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3  
+[34] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1737-1747, 2024. 1, 3  
+[35] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 3  
+[36] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16, 18  
+[37] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. In ICLR, 2014. 3  
+[38] Lucas Kovar, Michael Gleicher, and Frédéric Pighin. Motion graphs. ACM Trans. Graph., 21(3):473-482, 2002. 8  
+[39] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. EgoGen: An Egocentric Synthetic Data Generator. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3  
+[40] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 3  
+[41] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. In European Conference on Computer Vision, pages 54-72. Springer, 2024. 1  
+[42] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1524-1534, 2024. 1, 3  
+[43] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and
+
+Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024. 3  
+[44] Han Liang, Jiacheng Bao, Ruichi Zhang, Sihan Ren, Yuecheng Xu, Sibei Yang, Xin Chen, Jingyi Yu, and Lan Xu. Omg: Towards open-vocabulary motion generation via mixture of controllers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482-493, 2024. 1, 2, 3, 6  
+[45] Hung Yu Ling, Fabio Zinno, George Cheng, and Michiel Van De Panne. Character controllers using motion VAEs. ACM Trans. Gr., 2020. 3, 5  
+[46] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 3  
+[47] Matthew M. Loper, Naureen Mahmood, and Michael J. Black. MoSh: Motion and shape capture from sparse markers. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 33(6):220:1-220:13, 2014. 18  
+[48] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 16, 18  
+[49] Eyes JAPAN Co. Ltd. Eyes Japan MoCap Dataset. 18  
+[50] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings International Conference on Computer Vision, pages 5442-5451. IEEE, 2019. 3, 4, 18  
+[51] M. Müller, T. Röder, M. Clausen, B. Eberhardt, B. Krüger, and A. Weber. Documentation mocap database hdm05. Technical Report CG-2007-2, Universität Bonn, 2007. 18  
+[52] NVIDIA, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, Daniel Dworakowski, Jiaojiao Fan, Michele Fenzi, Francesco Ferroni, Sanja Fidler, Dieter Fox, Songwei Ge, Yunhao Ge, Jinwei Gu, Siddharth Gururani, Ethan He, Jiahui Huang, Jacob Huffman, Pooya Jannaty, Jingyi Jin, Seung Wook Kim, Gergely Klár, Grace Lam, Shiyi Lan, Laura Leal-Taixe, Anqi Li, Zhaoshuo Li, Chen-Hsuan Lin, Tsung-Yi Lin, Huan Ling, Ming-Yu Liu, Xian Liu, Alice Luo, Qianli Ma, Hanzi Mao, Kaichun Mo, Arsalan Mousavian, Seungjun Nah, Sriharsha Niverty, David Page, Despoina Paschalidou, Zeeshan Patel, Lindsey Pavao, Morteza Ramezanali, Fitsum Reda, Xiaowei Ren, Vasanth Rao Naik Sabavat, Ed Schmerling, Stella Shi, Bartosz Stefaniak, Shitao Tang, Lyne Tchapmi, Przemek Tredak, Wei-Cheng Tseng, Jibin Varghese, Hao Wang, Haoxiang Wang, Heng Wang, Ting-Chun Wang, Fangyin Wei, Xinyue Wei, Jay Zhangjie Wu, Jiashu Xu, Wei Yang, Lin Yen-Chen, Xiaohui Zeng, Yu Zeng, Jing Zhang, Qinsheng Zhang, Yuxuan Zhang, Qingqing Zhao, and Artur Zolkowski. Cosmos world foundation model platform for physical ai, 2025. 5  
+[53] OpenAI. Chatgpt: Conversational ai model, 2024. Accessed: 2025-02-26. 5  
+[54] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and
+
+Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3  
+[55] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4, 6, 14  
+[56] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. arXiv preprint arXiv:2104.02180, 2021. 3  
+[57] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 6, 14  
+[58] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Maria Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2021), pages 722-731, Piscataway, NJ, 2021. IEEE. 3, 8  
+[59] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. CoRR, abs/1710.05941, 2017. 4, 14, 18  
+[60] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 4  
+[61] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13756-13766, 2023. 3  
+[62] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. NeurIPS, 2023. 3, 13  
+[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5  
+[64] Yi Shi, Jingbo Wang, Xuekun Jiang, Bingkun Lin, Bo Dai, and Xue Bin Peng. Interactive character control with autoregressive motion diffusion models. ACM Trans. Graph., 43, 2024. 3, 4, 5, 18  
+[65] Leonid Sigal, Alexandru O Balan, and Michael J Black. HumanEva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. 2010. 6  
+[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2, 3  
+[67] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2, 3
+
+[68] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In Computer Vision - ECCV 2020, 2020. 1, 3  
+[69] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6  
+[70] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 4, 5  
+[71] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 3, 4, 5, 6, 7, 9, 13, 15, 17, 18  
+[72] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5  
+[73] Shashank Tripathi, Omid Taheri, Christoph Lassner, Michael Black, Daniel Holden, and Carsten Stoll. Humos: Human motion model conditioned on body shape. In European Conference on Computer Vision, pages 133-152. Springer, 2024. 1  
+[74] Nikolaus F. Troje. Decomposing biological motion: A framework for analysis and synthesis of human gait patterns. Journal of Vision, 2(5):2-2, 2002. 18  
+[75] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 3  
+[76] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines, 2024. 17  
+[77] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 4, 14  
+[78] Jingbo Wang, Zhengyi Luo, Ye Yuan, Yixuan Li, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 718-728, 2024. 3  
+[79] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In ECCV, 2022. 3  
+[80] Zhen Wu, Jiaman Li, Pei Xu, and C Karen Liu. Human-object interaction from human-level instructions. arXiv preprint arXiv:2406.17840, 2024. 3  
+[81] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. In The Twelfth International Conference on Learning Representations, 2024. 3, 6, 7, 16, 17
+
+[82] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3, 6  
+[83] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836-3847, 2023. 2, 3, 5  
+[84] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 3  
+[85] Siwei Zhang, Bharat Lal Bhatnagar, Yuanlu Xu, Alexander Winkler, Petr Kadlecek, Siyu Tang, and Federica Bogo. Rohm: Robust human motion reconstruction via diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14606-14617, 2024. 5  
+[86] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 3, 4, 5, 6, 15  
+[87] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In CVPR, 2021. 4  
+[88] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14738-14749, 2023. 1, 3, 5, 6  
+[89] Kaifeng Zhao, Gen Li, and Siyu Tang. DartControl: A diffusion-based autoregressive motion model for real-time text-driven motion control. In The Thirteenth International Conference on Learning Representations (ICLR), 2025. 3, 4, 5, 6, 7, 15, 16, 18  
+[90] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. In European Conference on Computer Vision, pages 405-421. Springer, 2024. 6, 7, 16  
+[91] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:528649420c7f9af8262049286a43edaaa9bbfa60437129f9323c5cbe1e15b7de
+size 363200

ICCV/2025/PRIMAL_ Physically Reactive and Interactive Motor Model for Avatar Learning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cb6f49d6aad4f79312fd811f2ac02f43d8e14aad364eac025204acb5449518c
+size 416205

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/38f64838-a08c-4e4f-9654-04db04907cf0_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8871d5809b182072d228e678ca65c0692039e921c75a828fb60fe7be3db01b0e
+size 69851

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/38f64838-a08c-4e4f-9654-04db04907cf0_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f699e67db801f7264c4f0340c0b9df286da760352bca62c2a9da8356a2f4dcc
+size 85033

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/38f64838-a08c-4e4f-9654-04db04907cf0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a8f7cf7cc534f097e6470775474eacbda3e60f1533a653d5ed62bc3a9b78d4c
+size 760677

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/full.md

@@ -0,0 +1,290 @@
+# PRISM: Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection
+
+Mahdiyar Molahasani*, Azadeh Motamedi*, Michael Greenspan, Il-Min Kim, Ali Etemad Queen's University, Canada
+
+{m.molahasani, 19am43, michael.greenspan, ilmin.kim, ali.etemad}@queensu.ca
+
+# Abstract
+
+We introduce Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM), a new data-free and task-agnostic solution for bias mitigation in VLMs like CLIP. VLMs often inherit and amplify biases in their training data, leading to skewed predictions. PRISM is designed to debias VLMs without relying on predefined bias categories or additional external data. It operates in two stages: first, an LLM is prompted with simple class prompts to generate scene descriptions that contain spurious correlations. Next, PRISM uses our novel contrastive-style debiasing loss to learn a projection that maps the embeddings onto a latent space that minimizes spurious correlations while preserving the alignment between image and text embeddings. Extensive experiments demonstrate that PRISM outperforms current debiasing methods on the commonly used Waterbirds and CelebA datasets. We make our code public at: https://github.com/MahdiyarMM/PRISM.
+
+# 1. Introduction
+
+Vision-Language Models (VLMs) have emerged as a class of machine learning models capable of understanding and generating multi-modal content by jointly processing images and textual descriptions. Prominent VLMs, such as Contrastive Language-Image Pre-training (CLIP) [20], have demonstrated impressive performance in zero-shot classification and cross-modal retrieval. By aligning textual and visual representations in a shared embedding space, these models enable seamless interactions between vision and language tasks. Despite their unprecedented effectiveness, VLMs inherit and propagate systemic biases from their large-scale pretraining data, leading to inaccurate or skewed predictions.
+
+Biases in VLMs manifest in various forms, including social [19, 27, 30, 33] and contextual biases [26, 31], of
+
+![](images/96e971afecb83a4eec7ce8878cf17dc712ef1a8acf11ed45d550611ceca6cb1f.jpg)  
+(a) Waterbirds.
+
+![](images/5fb455f6279471084730491adde45c1205be94b9016d0f0e8e3e76c3ac0f4f97.jpg)  
+(b) CelebA.  
+Figure 1. Performance comparison among data-free CLIP debiasing methods in terms of debiasing performance (i.e., the accuracy of the worst group) vs. the overall accuracy on (a) Waterbirds and (b) CelebA datasets. Our proposed method, PRISM, achieves state-of-the-art debiasing performance (higher worst-group accuracy) without compromising overall accuracy.
+
+ten influenced by 'spurious correlations' in training data. Spurious correlations can stem from associating non-target attributes, such as image backgrounds or contextual features, with labels in unintended ways [5]. For instance, CLIP's zero-shot classifiers frequently associate 'camel' with 'desert' and 'panda' with 'bamboo', reflecting underlying biases in training data. These biases can have real-world implications reducing model robustness in diverse settings. Therefore, mitigating implicit bias in VLMs has become a critical research direction.
+
+Recent research has investigated various strategies to identify, quantify, and reduce spurious correlations in VLMs. These approaches can be classified into three categories: feature-space transformations [7, 29], prompt-based debiasing [1, 5, 6], and adapter design [17]. Methods with a primary focus on the feature space [7, 29] attempt to transform feature embeddings to eliminate spurious correlations. Prompt-based techniques [1, 5, 6] aim to redefine textual inputs to mitigate bias but often struggle with generalization. Meanwhile, adapter-based approaches such as [17] address implicit bias in VLMs by modifying representations at various stages of the model, including but not limited to the final layer.
+
+Despite the effectiveness of these approaches in addressing spurious bias in VLMs, they fall short in many instances. These methods mostly require fixed predefined bias categories makes them difficult to scale to task-agnostic settings [5-7]. Furthermore, these debiasing solutions either compromise the overall performance of the model by removing essential semantic content [1, 5] or rely on external visual data for debiasing [7, 17, 29]. This underscores the need for a new debiasing method without the reliance on predefined bias categories or any external data that can preserve essential semantic content without compromising overall accuracy.
+
+To address these limitations, we propose Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM). PRISM consists of two key stages. First, we prompt an LLM with labels to return a list of potential spurious connections corresponding to these labels and dynamically generate biased scene disruptions based on the extracted spurious attributes. Next, we find a projection of the embedding space that can effectively remove the impact of spurious attributes using the generated biased scene descriptions. To this end, we propose a novel loss named Latent space Debiasing loss (LD) to learn the debiasing projection in the embedding space effectively. Our proposed LD loss is a contrastive-style debiasing loss that aligns text and image embeddings based on semantic meaning rather than spurious correlations. It ensures that the text embeddings of the same class but different spurious attributes remain similar to each other, and the text embeddings of different classes but the same spurious attribute remain distinct.
+
+By removing the impact of bias attributes from the embedding space, PRISM generates representations with no or reduced statistical dependence on the bias. To measure this dependency, previously proposed methods mostly employ a metric capable of capturing statistical dependencies, such as Mutual Information [22], Hilbert-Schmidt Independence Criterion (HSIC) [4], or bias scores [7]. In contrast, our proposed method eliminates the need for direct dependency measurement by utilizing an LLM to generate bias-aware descriptions. These descriptions capture common bias attributes in the dataset using only the labels. Our approach facilitates debiasing without the need for predefined bias categories, external text or image datasets, or lossy data conversions to identify bias features. This makes PRISM an effective, task-agnostic, and data-free solution for debiasing VLMs.
+
+We assess the performance of PRISM on CLIP using different benchmarks, including Waterbirds and CelebA datasets, and compare its performance against various baselines. The results highlight the capability of the proposed PRISM in mitigating spurious biases in VLMs without compromising the overall performance. Figure 1 illustrates
+
+that our proposed PRISM achieves state-of-the-art debiasing performance without compromising the total accuracy on both Waterbirds and CelebA benchmarks. Our key contributions in this paper can be summarized as follows:
+
+- We leverage LLMs to generate biased textual descriptions that reflect common spurious correlations found in datasets. This provides a dynamic and adaptable way to identify biases without relying on predefined bias categories.  
+- We propose a novel task-agnostic data-free debiasing method for VLMs that learns a projection of the embedding space, removing the impact of spurious attributes using a novel contrastive style loss function.  
+- We demonstrate that our proposed method achieves state-of-the-art debiasing performance without compromising the model's overall accuracy. To allow rapid reproducibility and contribute to the community, we make our code public at: https://github.com/MahdiyarMM/PRISM.
+
+# 2. Related Works
+
+Bias identification in VLMs. Recent research has placed considerable emphasis on identifying and assessing bias in foundation models [9], including VLMs [2, 25]. The initial step in debiasing involves recognizing and evaluating the sources of bias. Different frameworks proposed in [31] and [11, 21] comprehensively explored various types of biases in vision and text datasets, respectively. Previous works [2, 25, 28] investigated bias in CLIP's representation space and proposed different methods to mitigate it. These biases cause unfair outcomes in zero-shot classification and reduce performance in certain underrepresented and overlooked categories [25].
+
+Debiasing VLMs. The methods proposed in [3, 19, 31] apply mathematical transformations on data to isolate bias in datasets before feeding them into models. However, such transformations may result in information loss. In the field of feature space transformation, FairerCLIP [7] introduced a mathematically rigorous approach to reduce the bias in CLIP's representations by leveraging kernel Hilbert spaces. SFID [15] proposed a selective feature imputation for debiasing, which combines feature pruning and low-confidence imputation to eliminate biases across tasks and modalities. These methods are grounded in rigorous mathematical principles for transformation but can be computationally expensive and risk removing crucial semantic information which compromises the overall accuracy of the model.
+
+In the field of prompt engineering, VisualDistiller [6] focused on refining visual representations to counter dataset biases. The solutions proposed in [5], namely Orth-cali and Orth-proj, efficiently remove the bias by projecting out the biased components from CLIP's text embeddings using predefined biased prompts. However, because these bi
+
+ased prompts are fixed, the method may struggle to generalize to biases that weren't anticipated in the prompts. In the field of adapter designing, DFR [17] suggested debiasing by retraining only the model's final layer. On the other hand, contrastive adapter [32] proposed to insert lightweight adapter modules into the pre-trained model architecture. These adapters are trained using a contrastive loss designed to pull representations from similar groups closer while pushing those from different groups apart.
+
+In a recent study, RoboShot [1] and BendVLM [10] also leveraged language models to extract insights from task descriptions. The primary focus of RoboShot was on improving robustness in underrepresented and adversarial data distributions rather than bias mitigation. BendVLM was proposed to tackle the bias in VLMs by learning task-specific projections via labeled images, with part of the projections computed dynamically during inference. In contrast, our PRISM explicitly targets bias mitigation in CLIP by leveraging LLMs for adaptive bias identification and embedding transformation. PRISM eliminates the need for LLM calls during inference and removes spurious correlations through learning a debiasing projection of the embedding space, ensuring effective bias removal without relying on predefined bias categories or external visual data. Furthermore, unlike prompt-based debiasing techniques that modify only text embeddings, our approach preserves the alignment between text and image representations, ensuring debiased zero-shot classification without compromising the overall accuracy.
+
+# 3. Method
+
+# 3.1. Problem formulation
+
+Let us consider a multimodal foundation model, such as CLIP [20], which maps images and natural language onto a shared  $d$ -dimensional space using two encoders: an image encoder  $\phi_{\mathrm{I}}(\cdot)$  and a text encoder  $\phi_{\mathrm{T}}(\cdot)$ . In the zero-shot setting, given image  $x$  and a set of text prompts  $\{t_y\}_{y=1}^K$  corresponding to class labels  $y \in \{1, \dots, K\}$ , the prediction is made via
+
+$$
+\hat {y} = \arg \max  _ {y} \left\langle \phi_ {\mathrm {I}} (x), \phi_ {\mathrm {T}} \left(t _ {y}\right) \right\rangle , \tag {1}
+$$
+
+where  $\langle \cdot ,\cdot \rangle$  represents the vector inner product. Here, the text prompts are descriptions of the object in each class as:  $t_y = "A$  photo of a  $<  class y>$
+
+This formulation presumes that  $\phi_I(x)$  predominantly captures the core, task-relevant features of  $x$ . However, real-world data often contain spurious correlations that can cause the encoder to also capture irrelevant cues. The image embedding can be modeled as a (normally, unknown) function  $f$  of a core component and a spurious component:
+
+$$
+\phi_ {\mathrm {I}} (x) = f \left(h _ {\text {c o r e}} (x), h _ {\text {s p u}} (x)\right), \tag {2}
+$$
+
+where  $h_{\mathrm{core}}(x)$  represents the features essential for correctly predicting the label  $y$ , and  $h_{\mathrm{spu}}(x)$  encodes extraneous information (e.g., background elements) that is spuriously correlated with  $y$ .
+
+Assume that the spurious attribute is denoted by  $a$ . Accordingly, we define the group identity of each sample as
+
+$$
+g = (a, y), \quad \text {w i t h} \quad g \in \mathcal {G} = A \times Y, \tag {3}
+$$
+
+where  $A$  is the set of possible spurious attributes  $a$ , and  $Y = \{1, \dots, K\}$  corresponds to the set of true class labels. For example, in a bird classification task,  $x$  is an image of a bird, and  $y \in \{0, 1\}$  represents the true class where 0 denotes Landbird and 1 denotes Waterbird. The spurious attribute  $a$  belongs to the set  $A = \{\text{land background}, \text{water background}\}$ , which indicates the type of background present in the image. Each combination  $(a, y)$  defines a distinct group within  $G$ , resulting in a total of four groups in this example.
+
+When the model undesirably relies on the spurious component  $h_{\mathrm{spu}}(x)$ , it may assign  $x$  to the incorrect group, leading to misclassification. To mitigate this bias, the objective is to minimize the worst-case error over all groups. Formally, we aim to solve
+
+$$
+\min  _ {\theta \in \Theta} \max  _ {g \in \mathcal {G}} \mathbb {E} \left[ \ell \left(\left\langle \phi_ {\mathrm {I}} (x), \phi_ {\mathrm {T}} \left(t _ {y}\right) \right\rangle , y\right) | g \right], \tag {4}
+$$
+
+where  $\ell (\cdot ,\cdot)$  denotes the classification loss.
+
+There are several challenges associated with the optimization of Eq. 4. First, optimizing over the parameter space of  $\theta$  is costly due to the number of parameters of the model, CLIP. This is also in contrast to the original purpose of CLIP which is to serve as a universal and general zero-shot classifier (Eq. 1). Second, the identification of spurious attributes  $A$  for an arbitrary set of text prompts is challenging and requires prior knowledge about the task, or an external dataset, which is not always accessible. Moreover, we hypothesize that due to the nature of spurious correlations, similar biases should also be present in the text domain. Hence, spurious correlations identified in the text modality can also be leveraged to debias the image modality. We will empirically verify this hypothesis in Section 5.
+
+# 3.2. PRISM
+
+To address the challenges mentioned in Section 3.1, we propose Projection-based Reduction of Implicit Spurious bias in vision-language Models (PRISM), which consists of two main steps: (i) PRISM detects potential spurious correlations  $A$  by prompting an LLM  $\mathcal{F}$  with a set of inputs  $\{t_y\}_{y=1}^K$  and instructs the model to generate a set of scene descriptions  $\mathcal{T}$ , corresponding to different groups  $\mathcal{G}$ . (ii) To preserve CLIP's generality and to avoid the need to fine-tune the network, we then introduce a novel method for
+
+![](images/9aa0dfa0d9ce2da6cb9b198418b790081f47838bc7e21384a2a03d26e37d21b2.jpg)  
+Figure 2. An overview of the proposed PRISM. In Stage 1, by passing text prompts as 'A photo of  $< y >$  to an LLM, a set of scene descriptions with different labels and spurious attributes is generated. Then, in Stage 2, these scene descriptions are used for learning a debiasing projection of the embedding space,  $P$ , using LD loss. Finally, zero-shot classification is performed in the projected debiased embedding space of the VLM.
+
+finding an embedding projector  $P$  that maps the embedding space onto another subspace where the influence of spurious attributes  $A$  is minimized. Figure 2 presents a high-level illustration of our proposed method.
+
+Stage 1: Leveraging LLMs to identify spurious correlations. To propose a fully task-agnostic debiasing method without relying on any external data or predefined bias categories, we propose using pre-trained LLM for bias category discovery. For this purpose, an LLM is prompted with the labels and instructed to return potential spurious correlations. The LLMs' effectiveness in the discovery of spurious correlations stems from their training scheme. These foundation models are trained over a massive corpus  $\mathcal{D}$  tokenized to  $w_{i}$  by minimizing the following:
+
+$$
+\frac {1}{N} \sum_ {\left(w _ {1}, \dots , w _ {n}\right) \in \mathcal {D}} \sum_ {i = 1} ^ {n} - \log P _ {\theta} \left(w _ {i} \mid w _ {<   i}\right), \tag {5}
+$$
+
+thereby learning the joint distribution  $P_{\theta}(w_1, \ldots, w_n)$ . Whenever a concept  $\alpha$  (e.g., 'camel') frequently co-occurs with another concept  $\beta$  (e.g., 'desert') in  $\mathcal{D}$ , the conditional probability  $P_{\theta}(\beta \mid \alpha)$  becomes large, even if  $\alpha$  and  $\beta$  are merely spuriously correlated rather than causally connected. Hence, for any set of text prompts  $\{t_y\}_{y=1}^K$ , the LLM's learned parameters encode high-probability transitions from  $\{t_y\}_{y=1}^K$  to  $A$  solely due to their co-occurrences in text [24]. This statistical encoding of spurious correlations arises because the model optimizes the likelihood over  $\mathcal{D}$  without discriminating between causal and non-causal
+
+patterns. Consequently, prompting the LLM with queries about  $\{t_y\}_{y = 1}^K$  produces  $A$ , as follows:
+
+$$
+A = \mathcal {F} \left(\left\{t _ {y} \right\} _ {k = 1} ^ {K}\right). \tag {6}
+$$
+
+Then, we use the returned spurious attributes to generate a set of controlled scene descriptions  $\mathcal{T}_{a,y}$  correspond to each group  $g = (a,y)$  in  $\mathcal{G}$ . The full prompts used in this stage of PRISM are provided in Appendix B.
+
+Stage 2: Learning a debiasing projection of the embedding space. Our method finds a projection  $P$  to minimize the impact of  $A$  in the embedding space. This is then followed by zero-shot classification in the debiased projected embedding space. To this end, we propose a novel loss, which we call Latent space Debiasing Loss (LD), formulated as:
+
+![](images/47910dc8426f3474669018278c6c21a0500ee997d13024c023e685e06452f594.jpg)
+
+where  $m$  is the margin hyperparameter. We propose  $\mathcal{L}_{\mathrm{LD}}$  to learn a projection that robustly mitigates spurious correlations in the model's joint embedding space. For each iteration of optimization over  $P$ , we use a set of scene descriptions  $\mathcal{T}$ , generated by the LLM for each group,  $g = (a,y)$ . The first component of the loss ((a) Intra-class, inter-attribute term) encourages invariance by minimizing the difference between embeddings of scene descriptions from the same class but differing spurious attributes. Specifically, for each pair  $(a,a')$  with  $a \neq a'$ , the term  $1 - \langle \phi_{\mathrm{T}}((a,y)), \phi_{\mathrm{T}}((a',y')) \rangle$  is minimized. This term ensures that the core class-specific features are consistently represented, independent of the confounding spurious cues. Meanwhile, the second component of the loss ((b) Inter-class, intra-attribute term) employs a margin-based penalty that enforces a minimum separation between embeddings from the same spurious attribute but different classes. For any  $y' \neq y$ , the penalty  $\max \{0, \langle \phi_{\mathrm{T}}((a,y)), \phi_{\mathrm{T}}((a',y')) \rangle - m\}$  is applied, ensuring that even if the spurious context is shared, class-discriminative features remain distinct.
+
+The design of  $\mathcal{L}_{\mathrm{LD}}$  is inspired by contrastive loss in self-supervised learning, where the objective is to draw semantically similar pairs closer and push dissimilar ones apart. By jointly optimizing for intra-class invariance and inter-class separation, our loss effectively penalizes deviations from the desired embedding structure. Importantly, this approach removes the need to fine-tune the entire CLIP model, as the learnable projection  $P$  is optimized independently. Empirical evaluations confirm that minimizing  $\mathcal{L}_{\mathrm{LD}}$  not only enhances the robustness of the zero-shot classifier by reducing reliance on spurious cues but also improves worst-group performance and leads to disentanglement of the embedding space.
+
+Finally, leveraging the learned projection,  $P$ , by minimizing Eq. 7, we perform zero-shot classification on the projected embedding space as follows:
+
+$$
+\hat {y} = \arg \max  _ {k} \langle P \left(\phi_ {\mathrm {I}} (x)\right), P \left(\phi_ {\mathrm {T}} \left(t _ {k}\right)\right) \rangle . \tag {8}
+$$
+
+PRISM-mini. Here, we propose a variation of our method called PRISM-mini, which is mostly different from PRISM in stage 2. Unlike PRISM, which uses scene descriptions for finding  $P$  by optimizing over LD loss, PRISM-mini relies on orthogonalizing the embedding space again spurious attributes. Hence, PRISM-mini is less computationally complex and is more suitable for use-cases where computation resources are extremely limited.
+
+In the first stage, we pass the class prompts to the LLM and obtain a set of spurious attributes rather than full scene descriptions. Then, all these spurious attributes are encoded using CLIP's text encoder as follows:
+
+$$
+\mathcal {A} = \phi_ {\mathrm {T}} \left(\mathcal {F} \left(\left\{t _ {k} \right\} _ {k = 1} ^ {K}\right)\right), \tag {9}
+$$
+
+Algorithm 1 PRISM  
+Require: image  $x$ $\{t_y\}_{y = 1}^K$  : Text prompts for  $K$  classes  $\phi_{\mathrm{I}},\phi_{\mathrm{T}}$  : CLIP's image and text encoders  $\mathcal{F}$  LLM   
+1: Stage 1: Identify Spurious Directions   
+2:  $A = \mathcal{F}(\{t_y\}_{k = 1}^K)$  Extract spurious attributes   
+3: Stage 2 (PRISM): Learned Projection   
+4: Generate scene descriptions  $\mathcal{T}_{a,y}$  via  $\mathcal{F}$  and  $A$    
+5: Initialize  $P$    
+6: repeat   
+7: Compute  $\mathcal{L}_{\mathrm{LD}}$  Eq.7 on  $\{(a,y)\}$    
+8: Update  $P$  via gradient descent   
+9: until converged   
+10: Stage 2 (PRISM-mini): Simple Orthogonal Projection   
+11:  $\mathcal{A} = \phi_{\mathrm{T}}(A)$    
+12:  $P\gets I - \mathcal{A}(\mathcal{A}^{\top}\mathcal{A})^{-1}\mathcal{A}^{\top}$    
+13: Stage 3: Zero-shot Inference   
+14:  $\mathbf{z}_{\mathrm{I}}\gets P(\phi_{\mathrm{I}}(x)),\quad \mathbf{z}_{\mathrm{T},k}\gets P(\phi_{\mathrm{T}}(t_k))$    
+15:  $\hat{y}\gets \arg \max_k\langle \mathbf{z}_\mathrm{I},\mathbf{z}_{\mathrm{T},k}\rangle$
+
+where  $\mathcal{A}$  is a matrix whose columns are the embeddings of bias attributes in  $A$ . Inspired by [5], here we seek to reduce the impact of these spurious attributes on the embedding space by orthogonalizing it against  $\mathcal{A}$ . To this end, we calculate the projection  $P$  as follows:
+
+$$
+P = I - \mathcal {A} \left(\mathcal {A} ^ {\top} \mathcal {A}\right) ^ {- 1} \mathcal {A} ^ {\top}. \tag {10}
+$$
+
+The rest of the method is essentially the same as PRISM. PRISM-mini can effectively reduce the impact of spurious correlations in CLIP without the need for any optimization for finding  $P$ . As a result, this method is very cost-effective. However, compared to PRISM, it might not be capable of fully mitigating the spurious biases in the embedding space due to its simpler projection mechanism. Algorithm 1 summarizes both variations of the proposed method.
+
+# 4. Experiment setup
+
+Datasets. To evaluate the effectiveness of our method, we conduct various experiments on Waterbirds [23] and CelebA [18] datasets, following prior works such as [7]. The Waterbirds dataset consists of images where bird species are paired with specific backgrounds, deliberately inducing spurious correlations between the foreground object and the environmental context. The CelebA dataset includes 200,000 facial images of celebrities annotated with 40 attributes, such as gender, hair color, face shape, etc.
+
+Evaluation. For performance evaluation, we use Five metrics: Accuracy (Acc), Worst-Group accuracy (WG), which is the lowest accuracy of all subgroups,  $\Delta$ WG, which indicates the worst group accuracy improvement compared to
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Waterbirds</td><td colspan="5">CelebA</td></tr><tr><td>WG↑</td><td>Acc↑</td><td>ΔWG↑</td><td>ΔAcc↑</td><td>Gap↓</td><td>WG↑</td><td>Acc↑</td><td>ΔWG↑</td><td>ΔAcc↑</td><td>Gap↓</td></tr><tr><td colspan="11">Baseline</td></tr><tr><td>Zero-shot</td><td>36.4%</td><td>89.3%</td><td>—</td><td>—</td><td>52.9%</td><td>72.8%</td><td>87.6%</td><td>—</td><td>—</td><td>14.8%</td></tr><tr><td colspan="11">Methods using images for debiasing</td></tr><tr><td>WiSE-FT [29]</td><td>65.9%</td><td>97.6%</td><td>29.5%</td><td>8.3%</td><td>31.7%</td><td>80.0%</td><td>87.4%</td><td>7.2%</td><td>-0.2%</td><td>7.4%</td></tr><tr><td>DFR (Sub) [17]</td><td>51.9%</td><td>95.7%</td><td>15.5%</td><td>6.4%</td><td>43.8%</td><td>76.3%</td><td>92.1%</td><td>3.5%</td><td>4.5%</td><td>15.8%</td></tr><tr><td>DFR (Up) [17]</td><td>65.9%</td><td>96.1%</td><td>29.5%</td><td>6.8%</td><td>30.2%</td><td>83.7%</td><td>91.2%</td><td>10.9%</td><td>3.6%</td><td>7.5%</td></tr><tr><td>FairerCLIP [7]</td><td>78.1%</td><td>85.1%</td><td>41.7%</td><td>-4.2%</td><td>7.0%</td><td>86.1%</td><td>88.0%</td><td>13.3%</td><td>0.4%</td><td>1.9%</td></tr><tr><td colspan="11">Data-free methods</td></tr><tr><td>VisualDistiller [6]</td><td>42.7%</td><td>90.6%</td><td>6.3%</td><td>1.3%</td><td>47.9%</td><td>—</td><td>—</td><td>—</td><td>—</td><td>—</td></tr><tr><td>Orth-Proj [5]</td><td>45.3%</td><td>86.4%</td><td>8.9%</td><td>-2.9%</td><td>41.1%</td><td>71.1%</td><td>87.0%</td><td>-1.7%</td><td>-0.6%</td><td>15.9%</td></tr><tr><td>Orth-Cali [5]</td><td>68.8%</td><td>84.5%</td><td>32.4%</td><td>-4.3%</td><td>15.7%</td><td>76.1%</td><td>86.2%</td><td>3.3%</td><td>-1.4%</td><td>10.1%</td></tr><tr><td>RoboShot [1]</td><td>45.2%</td><td>79.2%</td><td>8.8%</td><td>-10.1%</td><td>34.0%</td><td>82.6%</td><td>85.5%</td><td>9.8%</td><td>-2.1%</td><td>2.9%</td></tr><tr><td>PRISM-mini (ours)</td><td>69.5%</td><td>92.6%</td><td>33.1%</td><td>3.3%</td><td>23.1%</td><td>82.6%</td><td>84.4%</td><td>9.8%</td><td>-3.2%</td><td>1.8%</td></tr><tr><td>PRISM (ours)</td><td>84.2%</td><td>93.6%</td><td>47.8%</td><td>4.3%</td><td>9.4%</td><td>84.0%</td><td>86.9%</td><td>11.2%</td><td>-0.7%</td><td>2.9%</td></tr></table>
+
+Table 1. The performance of PRISM compared to baselines on CLIP-ViT-L/14 on Waterbirds and CelebA datasets. The top performance in each column is in **bold**, and the second-best is **underlined**.
+
+the zero-shot clip,  $\Delta$ Acc, which indicates the overall accuracy improvement compared to the zero-shot performance of CLIP, and the Gap between Acc and WG.
+
+Baselines. We compare our method against two categories of baselines: the methods that use auxiliary images for debiasing and data-free methods. The first category employs a set of labeled or unlabeled images in the training process to reduce bias in CLIP. These images are used to keep the alignment of text and image embeddings intact throughout the debiasing process. Three state-of-the-art methods from this category are compared to ours, including WiSE-FT [29], DFR [17], and FairerCLIP [7]. The second category consists of data-free models which do not require any images for debiasing. These methods are inherently task-agnostic as they do not rely on any task-specific visual data to achieve effective debiasing. VisualDistiller [6], Orth [5], and RoboShot [1] are used in the following experiments as state-of-the-art methods in this category. Note that both variations of our method fall into this category.
+
+Implementation details. All experiments were conducted using PyTorch on an NVIDIA GeForce RTX 3090 GPU with 24 GB VRAM and 64 GB RAM. Each experiment was run three times with different initialization seeds, and the average performance of the models was reported. Adam optimizer [16] with the learning rate of 0.1 (Waterbirds) and 0.01 (CelebA) is used for performing optimization over  $P$ . The batch size is set to 64, and the training for the projection step is only performed for a single epoch.
+
+# 5. Results
+
+Performance. We evaluate the performance of our method and compare it against prior baselines on two datasets (Waterbirds and CelebA) using the CLIP-ViT-L/14 backbone in Table 1. The results demonstrate that PRISM outperforms prior works, setting a new state-of-the-art over data-free debiasing methods. Specifically, we observe that our proposed PRISM achieves the highest WG in both datasets, meaning it effectively reduces bias without needing additional image data. The effectiveness of our method is further demonstrated as PRISM outperforms the models using auxiliary data in the Waterbirds benchmark and achieve competitive results in the CelebA benchmark. FairerCLIP is the best among image-based methods, improving WG at the cost of some accuracy loss. WiSE-FT and DFR achieve high accuracies but have limited bias mitigation, indicating that fine-tuning might not be a good option to mitigate bias. Orth and RoboShot perform moderately well but struggle in WG gains, demonstrating that although prompt tuning is a viable method for guiding CLIP's outputs, it is not a robust solution for bias removal. Moreover, the efficient variation of our method, PRISM-mini, also achieves the second-best performance in debiasing across all data-free methods while avoiding any optimization, making it a strong candidate for scenarios where resources are very limited.
+
+Per-group accuracy. In Figure 3, each subfigure shows bar plots of per-group accuracy for three different methods, i.e., Zero-shot, PRISM-mini, and PRISM, on Waterbirds (Figure 3a) and CelebA (Figure 3b) datasets. The groups on the  $x$ -axis ( $g(0,0)$ ,  $g(0,1)$ ,  $g(1,0)$ ,  $g(1,1)$ ) represent different
+
+![](images/91ab97243691c57379bb763cfad9926c74855b0d1b94ee05f972581e05e63b6f.jpg)  
+(a) Waterbirds.
+
+![](images/21cd3ea3b7c95741a91e493834e0c14d6f68b0771424737b2f924f2b37cd98a6.jpg)  
+(b) CelebA.
+
+![](images/c562215afb88cd8020d8e3ed6967eda3088aed225bb2c2ce198a499035b9912f.jpg)  
+Figure 3. Per-group accuracy of our method on (a) Waterbirds and (b) CelebA datasets.  
+(a) CLIP.  
+Figure 4. The representation map at the output of CLIP (a) without and (b) with PRISM.
+
+![](images/677e3b85db3f9a71c3da94dd5967841626d122c926314766b5e5a6bbaeb42597.jpg)  
+(b) CLIP w/ PRISM.
+
+subpopulations within each dataset, and the  $y$ -axis shows the accuracy achieved by each method on those subpopulations. The dashed horizontal lines indicate the WG accuracy achieved by each method, i.e., the lowest accuracy among the four subgroups. By looking at where each bar stands relative to the dashed lines, we can see how each method handles the most challenging subgroup. Across both datasets, PRISM improves accuracy over the zero-shot baseline in each subgroup and tends to raise the WG performance. PRISM-mini also provides gains over the zero-shot baseline, but it may not always reach the same level of WG performance as PRISM. In essence, the figure illustrates that both PRISM variants can help reduce performance disparities across different groups compared to a purely zero-shot approach, with PRISM typically offering the strongest and the most balanced improvements.
+
+Bias text. We previously hypothesized that since CLIP aligns image and text embeddings, the inherent nature of spurious bias should cause similar biases to emerge in the text domain alone [24]. Consequently, spurious correlations extracted from the text domain can also be applicable to images. To empirically evaluate this hypothesis, we generate a set of scene descriptions featuring different bird species in various locations. We then perform zero-shot classifica
+
+<table><tr><td>Zero shot</td><td>WG ↑</td><td>Acc ↑</td><td>g(0,0)</td><td>g(0,1)</td><td>g(1,0)</td><td>g(1,1)</td></tr><tr><td>Image</td><td>38.3%</td><td>92.8%</td><td>99.0%</td><td>87.6%</td><td>38.3%</td><td>59.8%</td></tr><tr><td>Scene description</td><td>33.1%</td><td>96.0%</td><td>100%</td><td>33.1%</td><td>41.9%</td><td>96.8%</td></tr></table>
+
+Table 2. Bias similarity in text and image modalities.
+
+tion on the embeddings of these descriptions and present the results in Table 2, demonstrating that similar biases can indeed be observed in the text domain.
+
+Trade-off analysis. Figures 1a and 1b compare different data-free CLIP debiasing methods in terms of accuracy and debiasing performance, measured through WG, on Waterbirds and CelebA datasets. Although some methods exhibit high debiasing performances, their overall performance is drastically degraded. PRISM is the most effective among all compared methods, achieving the best trade-off between these two metrics.
+
+Feature disentanglement. Figure 4 compares the representation distributions of zero-shot CLIP with PRISM on the Waterbirds dataset, where each color represents a group  $(a,y) \in \mathcal{G}$ . In CLIP's representation space (Figure 4a), clusters are highly mixed, indicating that the model heavily relies on spurious background correlations. In contrast, PRISM's representation space (Figure 4b) shows more structured and well-disentangled clusters, demonstrating that it effectively reduces the influence of background information. This analysis confirms that PRISM successfully mitigates bias rather than being influenced by spurious background correlations.
+
+Number of scene descriptions. Figure 5 shows the effect of varying the number of scene descriptions used during training the projection  $P$ . In these experiments, we evaluate how increasing or decreasing the volume of scene descriptions, which serve as bias-aware prompts generated by the LLM, affects the model's ability to mitigate spurious correlations. For both the Waterbirds and CelebA datasets, the results illustrate that there is an optimal range for the number of descriptions: too few may provide insufficient guid
+
+![](images/8f90683f4ebad2d893afd0b7013cac472dcc7b991ec1d37fba9974922548d660.jpg)  
+(a) Waterbirds.
+
+![](images/021fe02f605ed7b87ea1f7302646cc1b755c0982316a2cf7a97d8c3845c5ceeb.jpg)  
+(b)CelebA.
+
+![](images/3ed9d37a47260abf02854e6b5b360b0cb1fa1f3ea78f96414479983334b61b12.jpg)  
+Figure 5. Worst-group performance of PRISM for different numbers of scene descriptions used in training for (a) Waterbirds and (b) CelebA datasets.  
+(a) Waterbirds.  
+Figure 6. Worst-group performance of PRISM for different values of margin hyperparameter  $m$  for (a) Waterbirds and (b) CelebA datasets.
+
+![](images/606e3260c8ba1da1a03190ebef9f8b46d44e04e2b7cc4967a05e519dd24b8742.jpg)  
+(b)CelebA.
+
+ance for debiasing, while too many might undermine the discriminative signal, ultimately impacting the worst-group accuracy.
+
+$m$  in LD loss. Figure 6 shows the sensitivity of PRISM's performance to the hyperparameter  $m$  in the LD loss function, which determines the margin in the contrastive component of the loss. This parameter influences how aggressively the model penalizes similarities between embeddings of different classes with the same spurious attribute. The results, demonstrate that for both datasets, an  $m$  of 0.6 maximizes WG performance by separating the representations across different groups, achieving an optimal balance between preserving essential semantic information and removing spurious correlations.
+
+Role of LLMs. Table 3 shows that the effectiveness of PRISM in reducing bias, measured by WG and Acc, can be significantly influenced by the choice of LLM. Selecting a more capable LLM for bias identification can improve debiasing without severely compromising the overall classification performance. The table compares several LLMs that vary in size and capability, where we see that GPT-40 achieves the best WG performance with score of  $84.0\%$  and an overall accuracy of  $86.9\%$ .
+
+Limitations. While PRISM demonstrates strong performance in mitigating spurious biases in VLMs, it has some
+
+<table><tr><td>LLM</td><td>WG ↑</td><td>Acc ↑</td></tr><tr><td>Llama 3.2 (70B) [8]</td><td>58.9%</td><td>90.1%</td></tr><tr><td>DeepSeek r1 (8B) [12]</td><td>68.3%</td><td>89.2%</td></tr><tr><td>GPT 4o mini (~8B) [13]</td><td>81.6%</td><td>86.6%</td></tr><tr><td>GPT o1 (~300B) [14]</td><td>77.9%</td><td>83.9%</td></tr><tr><td>GPT 4o (~1.7T) [13]</td><td>84.0%</td><td>86.9%</td></tr></table>
+
+Table 3. The performance of PRISM on CelebA dataset using different LLMs.
+
+limitations. First, PRISM relies on LLMs to generate spurious attribute descriptions, and the quality of these descriptions can vary depending on the specific LLM used. While our results show that more capable LLMs yield better biasing performance, less powerful models may struggle to extract meaningful spurious correlations, potentially reducing effectiveness. Additionally, our approach focuses on linear projections to remove bias in the embedding space, which may not always capture highly complex or non-linear biases present in real-world data.
+
+# 6. Conclusion
+
+We propose PRISM, a data-free debiasing framework for vision-language models that mitigates spurious correlations in CLIP's embedding space without requiring any weight updates. PRISM has two stages. In the first stage, an LLM is leveraged with simple class prompts to generate scene descriptions that unravel common spurious correlations in the training data. In the second stage, these generated descriptions guide the learning of an optimized projection mapping, which is carried out via our novel LD loss to systematically reduce spurious biases while preserving the semantic alignment between image and text embeddings. Through extensive evaluations on benchmark datasets (Waterbirds and CelebA), we demonstrated that PRISM significantly improves worst-group accuracy while maintaining overall classification performance. Beyond its empirical effectiveness, PRISM offers a data-free and task-agnostic solution to debiasing multimodal representations. Future research directions include exploring non-linear projections for embedding space debiasing, which could potentially better capture complex spurious correlations beyond linear transformations.
+
+# Acknowledgment
+
+We would like to thank Geotab Inc., the City of Kingston, and NSERC for their invaluable and continued support of this work.
+
+# References
+
+[1] Dyah Adila, Changho Shin, Linrong Cai, and Frederic Sala. Zero-shot robustification of zero-shot models. In International Conference on Learning Representations, 2024. 1, 2, 3, 6  
+[2] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021. 2  
+[3] Ibrahim Alabdulmohsin, Xiao Wang, Andreas Peter Steiner, Priya Goyal, Alexander D'Amour, and Xiaohua Zhai. CLIP the bias: How useful is balancing data in multimodal learning? In International Conference on Learning Representations, 2024. 2  
+[4] Hyojin Bahng, Sanghyuk Chun, Sangdoo Yun, Jaegul Choo, and Seong Joon Oh. Learning de-biased representations with biased representations. In International conference on machine learning, pages 528-539. PMLR, 2020. 2  
+[5] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 1, 2, 5, 6  
+[6] Haocheng Dai and Sarang Joshi. Refining skewed perceptions in vision-language models through visual representations. arXiv preprint arXiv:2405.14030, 2024. 1, 2, 6  
+[7] Sepehr Dehdashtian, Lan Wang, and Vishnu Naresh Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. arXiv preprint arXiv:2403.15593, 2024. 1, 2, 5, 6  
+[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8  
+[9] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 2  
+[10] Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas K Sharma, Tom Hartvigsen, and Marzyeh Ghassemi. Bendvlm: Test-time debiasing of vision-language embeddings. NeurIPS, 37:62480-62502, 2024. 3  
+[11] Priyanshul Govil, Hemang Jain, Vamshi Krishna Bonagiri, Aman Chadha, Ponnurangam Kumaraguru, Manas Gaur, and Sanorita Dey. Cobias: Contextual reliability in bias assessment. arXiv preprint arXiv:2402.14889, 2024. 2  
+[12] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 8  
+[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024.8
+
+[14] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 8  
+[15] Hoin Jung, Taeuk Jang, and Xiaogian Wang. A unified debi-aising approach for vision-language models across modalities and tasks. Advances in Neural Information Processing Systems, 37:21034-21058, 2025. 2  
+[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6  
+[17] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In International Conference on Learning Representations, 2023. 1, 2, 3, 6  
+[18] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5  
+[19] Nicole Meister, Dora Zhao, Angelina Wang, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Gender artifacts in visual datasets. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4837-4848, 2023. 1, 2  
+[20] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3  
+[21] Shaina Raza, Mizanur Rahman, and Michael R Zhang. Beads: Bias evaluation across domains. arXiv preprint arXiv:2406.04220, 2024. 2  
+[22] Christian Reimers, Paul Bodesheim, Jakob Runge, and Joachim Denzler. Towards learning an unbiased classifier from biased data via conditional adversarial debiasing. arXiv preprint arXiv:2103.06179, 2021. 2  
+[23] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. International Conference on Learning Representations, 2020. 5  
+[24] Pritam Sarkar, Sayna Ebrahimi, Ali Etemad, Ahmad Beirami, Sercan Ö Arık, and Tomas Pfister. Data-augmented phrase-level alignment for mitigating object hallucination. International Conference on Learning Representations, 2025. 4, 7  
+[25] Jie-Jing Shao, Jiang-Xin Shi, Xiao-Wen Yang, Lan-Zhe Guo, and Yu-Feng Li. Investigating the limitation of clip models: The worst-performing categories. CoRR, 2023. 2  
+[26] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman, Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and Olga Russakovsky. Revise: A tool for measuring and mitigating bias in visual datasets. International Journal of Computer Vision, 130(7):1790-1810, 2022. 1  
+[27] Jialu Wang, Yang Liu, and Xin Wang. Are gender-neutral queries really gender-neutral? mitigating gender bias in im
+
+age search. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. 1  
+[28] Junyang Wang, Yi Zhang, and Jitao Sang. Fairclip: Social bias elimination based on attribute prototype learning and representation neutralization. arXiv preprint arXiv:2210.14562, 2022. 2  
+[29] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959-7971, 2022. 1, 2, 6  
+[30] Kaiyu Yang, Clint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparency, pages 547-558, 2020. 1  
+[31] Boya Zeng, Yida Yin, and Zhuang Liu. Understanding bias in large-scale visual datasets. arXiv preprint arXiv:2412.01876, 2024. 1, 2  
+[32] Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. Advances in Neural Information Processing Systems, 35:21682-21697, 2022. 3  
+[33] Kankan Zhou, Yibin LAI, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In Proceedings of the Association for Computational Linguistics. International Joint Conference on Natural Language Processing, 2022. 1

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2c4d10161c6159628366229edb248fd23cdd396b96eef615f19f40c9f6e0979
+size 453215

ICCV/2025/PRISM_ Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:423718181b0b871f7a7700950a04bcef6d2a37b96024f9be834c087f6f8085bd
+size 365395

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/376be5fb-cff2-44aa-87ae-beb1972ff68e_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc935885718251294e78b0b178e325e93c91edb38aaed88f07d15b7c4ce9284b
+size 75093

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/376be5fb-cff2-44aa-87ae-beb1972ff68e_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f39de67a37cfdcb0f8535a8aefc02171b4032b5e8bd8ee249aece42ff9de582
+size 93100

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/376be5fb-cff2-44aa-87ae-beb1972ff68e_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f967fdece584f9cde002e924971996ed6b822a4ad8f2b2a403cdf892bf3afea
+size 14834407

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/full.md

@@ -0,0 +1,306 @@
+# PRM: Photometric Stereo based Large Reconstruction Model
+
+Wenhang Ge $^{1*}$   
+Jiantao Lin $^{1*}$   
+Guibao Shen $^{1*}$   
+Jiawei Feng $^{1}$   
+Tao Hu $^{3}$   
+Xinli Xu $^{1}$   
+Ying-Cong Chen $^{1,2\dagger}$ $^{1}$ HKUST-GZ  $^{2}$  HKUST  $^{3}$  Pico, Bytedance
+
+# Abstract
+
+We propose PRM, a novel photometric stereo based large reconstruction model to reconstruct high-quality meshes with fine-grained details. Previous large reconstruction models typically prepare training images under fixed and simple lighting, offering minimal photometric cues for precise reconstruction. Furthermore, images containing specular surfaces are treated as out-of-distribution samples, resulting in degraded reconstruction quality. To handle these challenges, PRM renders images by varying materials and lighting, which not only improves the local details by providing rich photometric cues but also increases the model's robustness to variations in the appearance of input images. To offer enhanced flexibility, we incorporate a real-time physically-based rendering (PBR) method and mesh rasterization for ground-truth rendering. By using an explicit mesh as 3D representation, PRM ensures the application of differentiable PBR for predicted rendering. This approach models specular color more accurately for images with varying materials and illumination than previous neural rendering methods and supports multiple supervisions for geometry optimization. Extensive experiments demonstrate that PRM significantly outperforms other models.
+
+# 1. Introduction
+
+Recent advancements in generative models [12, 27, 38] have spurred notable progress in 2D content creation, driven by fast growth in data volumes. In contrast, the development in 3D field remains encumbered due to limited 3D assets, which are essential for diverse applications including game modeling [8], computer animation [23, 32], and virtual reality [35]. Traditional approaches to generating 3D assets have utilized optimization-based techniques from multi-view posed images [42, 49, 50] or have harnessed SDS-based distillation methods from 2D diffusion models [25, 26, 34]. Despite their effectiveness, these methods often require increased computational costs, making them less
+
+suitable for rapid deployment in real world.
+
+Feed-forward 3D generative models [13, 15, 51] have been developed to address the limitations of per-scene optimization by training a generalizable model on large-scale 3D assets. Notably, the Large Reconstruction Model (LRM) [13] has demonstrated promising results, exhibiting exceptional reconstruction speeds. The subsequent LRM series [13, 39, 43, 46, 48] utilizes a Transformer-based architecture to encode either single or multi-view images, and decoding them into 3D representations, such as triplanes [3], Flexicubes [36] or 3D Gaussians [39]. These 3D representations enable differentiable neural rendering from arbitrary viewpoints, which is crucial for calculating multi-view reconstruction loss for optimization.
+
+While the LRM series effectively reconstructs globally coherent 3D assets, it struggles with capturing fine-grained details. This limitation stems from its reliance on images rendered under fixed, simple lighting, which provide insufficient photometric information for detailed surface reconstruction. Additionally, the LRM series is sensitive to variations in input image appearance, particularly with glossy surfaces, as these are out-of-distribution samples that rarely occur in the training data. Moreover, the use of neural rendering for predicted rendering exacerbates this issue [41], since the view-dependent effect is not considered.
+
+To address the challenges, we introduce PRM, a photometric stereo based large reconstruction model. This model is adept at capturing fine-grained local details and ensures robustness against the complex appearances of input images. We achieve these objectives by leveraging images with varying materials and illumination [10]. Specifically, we render ground-truth images by varying camera pose, materials (i.e., metallic and roughness), and lighting. However, rendering these images is not trivial since there are infinite possible combinations of camera pose, materials and lighting. In the recent LRM series, images are typically rendered offline using Blender's Cycles engine [11]. While this approach produces high-quality, noise-free images, it requires numerous samples of lighting directions, significantly increasing the time cost and making it expensive to maximize the training sample distribution.
+
+![](images/0896bd3c4bd1244de8b9f0605c559b4e24f14442940c62e1e6199170301c255d.jpg)  
+Figure 1. Top left: PRM is capable of reconstructing high-quality meshes with fine-grained local details even under complex image appearances, such as specular highlights and dark appearances. Right: We demonstrate a scene comprising diverse 3D objects generated by our models. Bottom left: A zoomed-in visualization of the scene highlights these details more clearly.
+
+![](images/08daf036044bc9b59b4ed0ed7531072423011e51f36f56ee31454133e3c4bbec.jpg)
+
+To handle this issue, we incorporate a real-time, physically based rendering technique known as split-sum approximation [19], along with mesh rasterization for online rendering. This approach offers greater flexibility compared to traditional offline methods. We discuss two training strategies in the Supplement. Images with varying materials and illumination offer two distinct advantages. First, they provide additional photometric cues, thereby enhancing the capacity to recover fine-grained local details. Second, the PRM model demonstrates remarkable robustness to variations in the appearances of inputs. For instance, it is capable of accurately reconstructing the geometry of images with glossy surfaces. Moreover, by utilizing mesh as 3D representation, we can employ differentiable PBR for predicted rendering, producing intermediate shading variables such as albedo, specular and diffuse light maps, along with geometric cues like normals and depth. These variables provide multiple supervisions, including photometric and geometric supervision for high-quality reconstruction. Furthermore, differentiable PBR can better disentangle the specular component, making the geometry be correctly recovered when the images are characterized with glossy surfaces.
+
+To summarize, our contributions are listed as follows.
+
+- We introduce PRM, a model that is capable of reconstructing geometry with fine-grained local details and robust to variations in the appearance of input image by utilizing images with varying materials and illumination.  
+- We employ differentiable PBR for predicted rendering by utilizing mesh as the 3D representation. This approach is advantageous for modeling reflective components and enables the incorporation of multiple supervisions for high-quality geometry reconstruction.
+
+- Extensive experiments on multiple datasets demonstrate the effectiveness of the proposed framework, which significantly outperforms previous methods, particularly on glossy objects.
+
+# 2. Related Work
+
+# 2.1. Feed-forward 3D Generative Models
+
+Large-scale 3D assets [5] facilitate the training of generalizable reconstruction models. Recent works have focused on generating 3D objects using feed-forward models [13, 14, 24, 39, 43, 47, 48, 51], demonstrating impressive results in terms of speed and quality. Specifically, Clay [51] utilizes occupancy for direct supervision. X-ray [15] explores novel 3D representations by converting a 3D object into a series of surface frames at different layers. The LRM series [13, 24, 43, 46, 48] shows that a transformer backbone can effectively map image tokens to 3D triplanes, benefiting from multi-view supervision. Instant3D [24] employs multi-view images to provide additional 3D information for triplane prediction, yielding promising outcomes. CRM [43] and InstantMesh [46] opt for an explicit mesh representation, supporting mesh rasterization and rendering additional geometric cues for supervision. Despite these achievements, the existing LRM series typically render low-frequency images under fixed and simple lighting, which compromises the model's adaptability to complexity in the appearance of input images and the model's capability to recover local details due to limited photometric cues. In response, we render images with varying materials and illumination that significantly enhance the photometric cues necessary for the recovery of fine-grained local details.
+
+# 2.2. Photometric Stereo
+
+Photometric stereo (PS) is a technique for recovering surface normals from the appearance of an object under varying lighting conditions. Traditional methods, inspired by the seminal work [44], assume calibrated, directional lighting. Recently, uncalibrated photometric stereo methods have emerged, which assume Lambertian integrable surfaces and aim to resolve the General Bas-Relief ambiguity [9] between light and geometry. However, these methods are still constrained to single directional lighting. More contemporary research [16, 17, 29] has shifted focus towards natural lighting conditions. Despite the significant progress, these approaches generally concentrate on single-view photometric stereo, relying solely on photometric cues and neglecting multi-view information, which is crucial for accurately reasoning geometric features. Some studies [10, 20–22, 33, 53] leverage both photometric and geometric cues for reconstruction. These cues are complementary: photometric stereo provides precise local details, while multiview information yields accurate global shapes [53]. For example, UA-MVPS [20] utilizes complementary strengths of PS and multi-view stereo for geometry reconstruction. NeRF-MVPS [21] utilizes surface normal estimated from images with varying materials and illumination to enhance the reconstruction performance of NeRF. Our approach integrates the principles of photometric stereo into LRM, aiming to harness the strengths of photometric cues for enhanced reconstruction accuracy.
+
+# 2.3. Physically-based Rendering
+
+Physically based rendering (PBR) is a computer graphics approach that renders photo-realistic images. PBR offers a physically plausible approach to modeling radiance by simulating the interaction between lighting and materials. PBR has proven to be effective in improving the geometry for multi-view reconstruction task. For example, incorporating the principles of PBR into volume rendering significantly improves the accuracy [7, 28, 41], especially for glossy surfaces. Since geometry and predicted radiance are closely entangled, improving radiance modeling can also enhance geometry reconstruction. Besides, PBR is also widely used in inverse rendering task [1, 4, 31]. The task aims at decomposing image appearance into intrinsic properties. Unlike previous LRM methods that predict radiance without explicitly considering the interactions between materials and lighting, we leverage advancements from the multi-view reconstruction field and employ PBR for improved radiance modeling and geometry reconstruction. To this end, we predict albedo instead of color, which is more reasonable as albedo is view-independent. The final color is derived using the predicted albedo and the ground truth metallic, roughness, and lighting.
+
+# 3. Method
+
+We begin with a succinct overview of large reconstruction model, physically based rendering and photometric stereo in Section 3.1. Then, we introduce how to prepare images with varying materials and illumination in Section 3.2. Subsequently, we introduce PRM in Section 3.3, with our proposed comprehensive objectives and applications. An overview of our framework is provided in Figure 2.
+
+# 3.1. Preliminaries
+
+Large Reconstruction Model aims to reconstruct 3D assets given a single image. LRM first utilizes a pre-trained visual transformer [2], to encode the images into image tokens. Subsequently, it employs an image-to-triplane transformer decoder that projects these 2D image tokens onto a 3D triplane using cross-attention [13]. Following this, images can be differentiable rendered from any viewpoint, supporting photometric supervision and optimization.
+
+Physically-based Rendering aims to produce 2D images using specified geometry, materials, and lighting. Central to this process is the rendering equation [18] formulated by
+
+$$
+C (\boldsymbol {x}, \omega_ {o}) = \int_ {\Omega} f (\boldsymbol {x}, \omega_ {o}, \omega_ {i}) L _ {i} (\boldsymbol {x}, \omega_ {i}) (\omega_ {i} \cdot \mathbf {n}) d \omega_ {i}, \tag {1}
+$$
+
+where  $\omega_{o}$  is the viewing direction of the outgoing light,  $L_{i}$  is the incident light of direction  $\omega_{i}$  sampled from the upper hemisphere  $\Omega$  of the surface point  $\mathbf{x}$ , and  $\mathbf{n}$  is the surface normal.  $f$  is the BRDF properties. The function  $f$  consists of a diffused and a specular component
+
+$$
+f (\boldsymbol {x}, \boldsymbol {\omega} _ {o}, \boldsymbol {\omega} _ {i}) = (1 - m) \frac {\boldsymbol {a}}{\pi} + \frac {D F G}{4 (\boldsymbol {\omega} _ {i} \times \mathbf {n}) (\boldsymbol {\omega} _ {o} \times \mathbf {n})}, \quad (2)
+$$
+
+where  $m\in [0,1]$  is the metallic,  $\pmb {a}\in [0,1]^3$  is the albedo. We detail the expression of  $D,F$  and  $G$  in the Supplement. With Eq.(1) and Eq.(2), the outgoing radiance is given by
+
+$$
+\boldsymbol {C} (\boldsymbol {x}, \boldsymbol {\omega} _ {o}) = \boldsymbol {C} _ {\mathrm {d}} (\boldsymbol {x}, \boldsymbol {\omega} _ {o}) + \boldsymbol {C} _ {\mathrm {s}} (\boldsymbol {x}, \boldsymbol {\omega} _ {o}), \tag {3}
+$$
+
+$$
+\boldsymbol {C} _ {\mathrm {d}} (\boldsymbol {x}, \boldsymbol {\omega} _ {\boldsymbol {o}}) = (1 - m) \boldsymbol {a} \int_ {\Omega} L _ {i} (\boldsymbol {x}, \boldsymbol {\omega} _ {\boldsymbol {i}}) \frac {(\boldsymbol {\omega} _ {\boldsymbol {i}} \cdot \mathbf {n})}{\pi} d \boldsymbol {\omega} _ {\boldsymbol {i}}, \tag {4}
+$$
+
+$$
+C _ {\mathrm {s}} (\boldsymbol {x}, \omega_ {o}) = \int_ {\Omega} \frac {D F G}{4 (\boldsymbol {\omega} _ {i} \times \mathbf {n}) (\boldsymbol {\omega} _ {o} \times \mathbf {n})} L _ {i} (\boldsymbol {x}, \omega_ {i}) (\boldsymbol {\omega} _ {i} \cdot \mathbf {n}) d \boldsymbol {\omega} _ {i}, \tag {5}
+$$
+
+$C_{\mathrm{s}}$  and  $C_\mathrm{d}$  are specular and diffuse color, respectively.
+
+Photometric Stereo aims to estimate the surface normals by observing an object under varying lightings [44]. When considering a Lambertian surface illuminated by a single
+
+![](images/e7f63656e2a252119f11e3f00b82d4f7c491a7e03d7c40e6190c9dbf115d4d4e.jpg)  
+Figure 2. Overview of our framework. During training, images with varying materials and illumination are rendered using PBR with randomly varied materials, lighting, and camera poses, along with depth, normal, albedo, and lighting maps. Images are encoded as a mesh through the network. All associated maps are used for supervision. During inference, an optional multi-view diffusion model outputs multi-view images, which are then fed into the network for mesh prediction. Relighting and material editing functionalities are also supported.
+
+point-like, distant light, the color is determined solely by the diffuse term, which can be formulated by
+
+$$
+\boldsymbol {C} (\boldsymbol {x}) = \boldsymbol {a} (\boldsymbol {L} \cdot \boldsymbol {n}), \tag {6}
+$$
+
+where  $L = L\cdot \omega$ $\omega$  and  $L$  are the direction and the intensity of the lighting, respectively. When we observe the object under different lighting conditions  $L_{i}$  ,we have multiple such equations. We assume that both the direction and intensity of the lighting are known, and the shading colors  $C(x)$  is also observed. Under these conditions, determining the surface normals  $\pmb{n}$  and the albedo  $\pmb{a}$  is effectively equivalent to solving a system of equations derived from these observations. More lighting indicates more equations, which effectively constrain the solution space.
+
+# 3.2. images with varying materials and illumination
+
+Previous methods typically prepared training data by rendering multi-view images with fixed, simple lighting and materials in Blender [11]. This approach resulted in images with low-frequency appearances, offering limited photometric cues. Consequently, these methods struggled to reconstruct geometry with precise local details. Additionally, they often failed when processing images with glossy surfaces, as these are out-of-distribution samples for the model.
+
+We provide some examples in the Supplement.
+
+In contrast, we prepare images with varying materials and illumination by varying materials and lighting. A naive solution is to prepare these images offline, as with previous methods, but this approach poses significant challenges due to the infinite number of potential combinations of materials and lighting. Moreover, rendering high-quality images requires large sample counts, making traditional data preparation methods infeasible. To overcome these issues, we incorporate a real-time rendering method known as split-sum approximation [19] along with mesh rasterization, which facilitates rapid rendering. This method enables online data preparation and significantly enhances flexibility.
+
+Split-sum approximation. High-quality estimation of physically based rendering typically requires Monte Carlo sampling to approximate the integral in Eq.(1). However, this process demands large sample counts, making it time-consuming. Instead, we employ a real-time rendering method known as the split-sum approximation [19]. According to the split-sum approximation, the specular component in Eq.(5) can be rewritten as:
+
+$$
+\boldsymbol {C} _ {\mathrm {s}} (\boldsymbol {x}, \boldsymbol {\omega} _ {o}) \approx \int_ {\Omega} \frac {D F G}{4 (\boldsymbol {\omega} _ {o} \cdot \boldsymbol {n})} d \boldsymbol {\omega} _ {i} \int_ {\Omega} L (\boldsymbol {x}, \boldsymbol {\omega} _ {i}) D (\hat {\boldsymbol {d}}, \rho) d \boldsymbol {\omega} _ {i}, \tag {7}
+$$
+
+The first term is the integral of the BRDF, which is approximated by specular albedo  $\pmb{a}_{\mathrm{s}} = ((1 - m)*0.04 + m*\pmb {a})*F_1 + F_2$ , where  $F_{1}$  and  $F_{2}$  are pre-computed scalars and stored in a 2D lookup texture related to  $\rho ,n$  and  $\omega_{o}$ . The second term is the integral of lights on the normal distribution function  $D(\hat{\pmb{d}},\rho)$ , which can also be pre-computed and stored as mipmaps  $M_{\mathrm{s}}$ .  $\hat{\pmb{d}} = 2(-\omega_o\cdot n)\pmb {n} + \omega_o$  is the reflective direction. Then the Eq.(7) is modified as
+
+$$
+\boldsymbol {C} _ {\mathrm {s}} (\boldsymbol {x}, \omega_ {o}) = \boldsymbol {a} _ {\mathrm {s}} (\boldsymbol {a}, m, \boldsymbol {n}, \omega_ {o}) \boldsymbol {L} _ {\text {s p e c}} (\boldsymbol {x}, \boldsymbol {n}, \omega_ {o}, \rho , \boldsymbol {M} _ {\mathrm {s}}), \tag {8}
+$$
+
+where  $L_{\mathrm{spec}} = \mathrm{Tex\_Sample}(\pmb{x},\hat{\pmb{d}},\rho ,M_{\mathrm{s}})$ , Tex Sample indicates texture sampling based on different levels of roughness  $\rho$  in pre-computed lighting map  $M_{\mathrm{s}}$ . A low-resolution map  $M_{\mathrm{d}}$  is also created to represent low-frequency diffuse lighting and the diffuse part in Eq.(4) is simplified as
+
+$$
+\boldsymbol {C} _ {\mathrm {d}} (\boldsymbol {x}, \omega_ {o}) = \boldsymbol {a} _ {\mathrm {d}} (\boldsymbol {a}, m) \boldsymbol {L} _ {\text {d i f f}} (\boldsymbol {x}, \boldsymbol {n}, \boldsymbol {M} _ {\mathrm {d}}), \tag {9}
+$$
+
+where  $\mathbf{a}_{\mathrm{d}} = (1 - m)\mathbf{a}$  indicates the diffuse albedo and  $L_{\mathrm{diff}} = \mathrm{Tex\_Sample}(\mathbf{x},\mathbf{n},M_{\mathrm{d}})$ . We show some rendered examples in the Supplement.
+
+Discussion. We render images with varying materials and illumination using varied camera poses, materials, and lighting conditions rather than solely changing the lighting. Please refer to the Supplement for more details. This approach offers two distinct advantages. Firstly, multi-view images provide more geometric cues than single-view images, which are crucial for reconstructing globally reasonable geometry [20-22]. Secondly, by varying materials, we can create images with glossy appearances, particularly when the metallic component is high and roughness is low. These varied images serve as inputs, enhancing the model's robustness to variations in appearance. Furthermore, the core principles of photometric stereo remain applicable. In equations Eq.(8) and Eq.(9), the observation direction  $\omega_{o}$ , metallic value  $m$ , roughness  $\rho$ , and mipmaps  $M_{\mathrm{d}}$  and  $M_{\mathrm{s}}$  are all known. Predicting the surface normals  $\pmb{n}$  and the albedo  $\pmb{a}$  still equates to solving these equations. Moreover, compared to merely changing lighting, altering the metallic and roughness values allows for diverse shading color rendering, which produces a richer set of equations.
+
+Mesh Rasterization Rendering. Given an object with explicit mesh  $O$ , rasterization is utilized to determine surface points  $x$ , along with corresponding depth  $d$ , surface normals  $n$ , and mask  $m$ . After obtaining the surface points  $x$  and their surface normals  $n$  along with selected camera pose, materials and lighting, we leverage split-sum approximation to estimate the specular and diffuse color as Eq.(8) and Eq.(9), respectively. During the process, besides the shading color, we can also render albedo, specular light, and diffuse light maps. The entire process can be formulated as
+
+$$
+\left\{C, \boldsymbol {n}, \boldsymbol {d}, \boldsymbol {m}, \boldsymbol {a}, L _ {\text {s p e c}}, L _ {\text {d i f f}} \right\} = \operatorname {P B R} (\text {R a s t e r i z a t i o n} (O)), \tag {10}
+$$
+
+where  $C$ ,  $n$ ,  $d$ ,  $m$ ,  $a$ ,  $L_{\mathrm{spec}}$ , and  $L_{\mathrm{diff}}$  are the rendered color, normal, depth, mask, albedo, specular light, and diffuse light maps, respectively.
+
+# 3.3. PRM
+
+Mesh as 3D Representation. The previous LRM series typically integrate triplane as 3D representation. In contrast, we opt for an explicit representation using mesh as our 3D format, which enables the use of differentiable PBR method for better radiance modeling. As a result, specular and diffuse lighting maps are also renderable, providing extra photometric cues that are only related to surface normals. Moreover, PBR can effectively model the specular component, leading to improved geometry reconstruction results [7, 41]. Specifically, we leverage differentiable isosurface extraction module, namely FlexiCubes [36].
+
+Two Stage Optimization. Inspired by InstantMesh [46], we have similarly designed a two-stage optimization framework. The first stage mirrors Instantmesh, using triplane and volume rendering for optimization with offline rendered data. In the second stage, FlexiCubes is used as the 3D representation. To reuse the knowledge in the first stage, we load the pretrained model as in InstantMesh [46]. The original color MLP is repurposed as an albedo MLP to incorporate color priors. Since our focus is on reconstruction, we directly utilize the ground-truth metallic and roughness during training. Given that an explicit mesh serves as our 3D representation, novel views can be rendered as described in Eq.(10). The difference is that the mesh  $\hat{O}$  is extracted using the dual marching cubes algorithm [30], which utilizes predicted SDF values, deformation, and weights derived from the triplane formulated by
+
+$$
+\{\hat {\boldsymbol {C}}, \hat {\boldsymbol {n}}, \hat {\boldsymbol {d}}, \hat {\boldsymbol {m}}, \hat {\boldsymbol {a}}, \hat {\boldsymbol {L}} _ {\text {s p e c}}, \hat {\boldsymbol {L}} _ {\text {d i f f}} \} = \mathrm {P B R} (\text {R a s t e r i z a t i o n} (\hat {\boldsymbol {O}})). \tag {11}
+$$
+
+Optimization. During the training process, our total loss function is
+
+$$
+\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\mathrm {M S E}} (\boldsymbol {C}, \hat {\boldsymbol {C}}) + \lambda_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} (\boldsymbol {C}, \hat {\boldsymbol {C}}) \\ + \mathcal {L} _ {\mathrm {M S E}} (\boldsymbol {a}, \hat {\boldsymbol {a}}) + \lambda_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} (\boldsymbol {a}, \hat {\boldsymbol {a}}) \\ + \mathcal {L} _ {\mathrm {M S E}} \left(\boldsymbol {l} _ {*}, \hat {\boldsymbol {l}} _ {*}\right) + \lambda_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} \left(\boldsymbol {l} _ {*}, \hat {\boldsymbol {l}} _ {*}\right) \tag {12} \\ + \lambda_ {\text {n o r m a l}} \hat {\boldsymbol {m}} \otimes (1 - \boldsymbol {n} \cdot \hat {\boldsymbol {n}}) + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} \\ + \lambda_ {\text {d e p t h}} \hat {\boldsymbol {m}} \otimes \| \boldsymbol {d} - \hat {\boldsymbol {d}} \| _ {1} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {M S E}} (\boldsymbol {m}, \hat {\boldsymbol {m}}), \\ \end{array}
+$$
+
+where  $\mathcal{L}_{\mathrm{MSE}}$  and  $\mathcal{L}_{\mathrm{LPIPS}}$  indicates the mean square error loss and LPISP loss [52], respectively.  $* \in \{\mathrm{spec}, \mathrm{diff}\}$  denotes specular light and diffuse light maps, respectively.  $\mathcal{L}_{\mathrm{reg}}$  is the regularization terms used in FlexiCubes [36]. During training, we set  $\lambda_{\mathrm{LPIPS}} = 2.0$ ,  $\lambda_{\mathrm{normal}} = 0.2$ ,  $\lambda_{\mathrm{depth}} = 0.5$ ,  $\lambda_{\mathrm{mask}} = 1.0$  and  $\lambda_{\mathrm{reg}} = 0.01$ .
+
+For both the specular  $L_{\mathrm{spec}}$  and the diffuse lighting map  $L_{\mathrm{diff}}$ , which are directly influenced by surface normals, effectively optimizing these light maps significantly enhances
+
+![](images/ce0665adb94175989846b24499ac8558cc829d3c4a3df972d2d436a0e45788bb.jpg)  
+Figure 3. Qualitative comparisons with state-of-the-art methods and ground truth for single-view reconstruction task. PRM reconstructs the highest quality 3D mesh and provides a more accurate texture prediction from input photographs compared to the others.
+
+the surface normals, thereby refining the precision of local details. This process is analogous to photometric stereo. The key difference is that these maps are exclusively related to surface normals without considering albedo, as demonstrated in Eq.(8) and Eq.(9), in contrast to shading color.
+
+Applications. PRM achieves high-quality geometry reconstruction with predicted albedo. This capability enables us to render the object under novel lighting conditions and also to modify its material properties. Examples of these applications are provided in the Supplement.
+
+# 4. Experiments
+
+# 4.1. Datasets and Evaluation Protocol
+
+Datasets. For training, we used Objaverse [5], a dataset comprised of synthetic 3D assets, that allows us to control the lighting, geometry, and material properties. We first filter Objaverse to get a high-quality subset for training. The filtering process aims to exclude objects lacking texture maps or those of inferior quality, such as those with low-poly properties. Texture maps are essential as they provide detailed albedo maps; in their absence, vertex colors are used as a substitute for albedo. Moreover, low-poly
+
+meshes result in uneven surface lighting. During rendering, we maintained the original albedo unchanged and randomly selected a material combination from a total of 121 possibilities, which were derived by varying the metallic and roughness properties from 0 to 1 in increments of 0.1. For lighting, we utilized environment maps sampled from a collection of 679 maps available on Polyhaven.com, thereby ensuring a diverse range of lighting conditions.
+
+For evaluation, We performed quantitative comparisons using two public datasets, including Google Scanned Objects (GSO) [6] and OmniObject3D (Omni3D) [45]. We randomly picked out 300 objects as the evaluation set both datasets, respectively. To show the robust capabilities of our model on appearance variations, we rendered the input view of each object with a randomly sampled combination of materials and lighting. We also report extra comparison results, following previous methods that utilized fixed lighting and did not change materials.
+
+Evaluation Protocol. We evaluated both the 2D visual quality and the 3D geometric quality. For the 2D visual evaluation, we rendered novel views from the reconstructed 3D mesh and compared them with the ground truth views, using PSNR, SSIM, and LPIPS. Since other methods do not
+
+Table 1. Quantitative comparison with methods on the GSO and Omni3D datasets, showcasing 3D reconstruction and 2D rendering metrics.  
+
+<table><tr><td>Dataset</td><td colspan="5">GSO</td><td colspan="5">OmniObject3D</td></tr><tr><td>Metric</td><td>CD↓</td><td>FS@0.1↑</td><td>PSNR↑</td><td>SSIM↑</td><td>LPIPS↓</td><td>CD↓</td><td>FS@0.1↑</td><td>PSNR↑</td><td>SSIM↑</td><td>LPIPS↓</td></tr><tr><td>TripoSR</td><td>0.109</td><td>0.872</td><td>15.247</td><td>0.859</td><td>0.197</td><td>0.083</td><td>0.940</td><td>15.237</td><td>0.865</td><td>0.176</td></tr><tr><td>CRM</td><td>0.202</td><td>0.707</td><td>17.293</td><td>0.854</td><td>0.142</td><td>0.088</td><td>0.907</td><td>18.293</td><td>0.894</td><td>0.112</td></tr><tr><td>LGM</td><td>0.144</td><td>0.800</td><td>17.643</td><td>0.869</td><td>0.158</td><td>0.138</td><td>0.823</td><td>17.893</td><td>0.884</td><td>0.139</td></tr><tr><td>InstantMesh</td><td>0.076</td><td>0.931</td><td>19.988</td><td>0.901</td><td>0.096</td><td>0.096</td><td>0.892</td><td>18.608</td><td>0.903</td><td>0.096</td></tr><tr><td>PRM</td><td>0.050</td><td>0.981</td><td>25.125</td><td>0.929</td><td>0.061</td><td>0.053</td><td>0.979</td><td>25.063</td><td>0.932</td><td>0.063</td></tr></table>
+
+predict albedo, we compared their shading color in novel views. For our method, we rendered the shading color based on the predicted mesh and albedo, then performed the metric calculations. For the 3D geometric evaluation, we first aligned the coordinate systems of the reconstructed meshes with those of the ground truth meshes. Subsequently, we repositioned and rescaled all meshes into a cube of size  $[-1,1]^3$ . We reported the Chamfer Distance (CD) and F-Score (FS) at a threshold of 0.1, which were computed by uniformly sampling 16K points from the mesh surfaces.
+
+# 4.2. Implementation Details
+
+Our model was developed based on InstantMesh [46]. The architecture of the Transformer encoder, triplane transformer, and the FlexiCubes decoder mirrors that of InstantMesh. Our model underwent training for 7 days and 3 days on 32 NVIDIA RTX A800 GPUs for the first stage and second stage, respectively. During inference, we used a single RGB image to generate six fixed views by Zero123++ [37] as the input of PRM. For more details, please see our Supplement.
+
+# 4.3. Comparison with State-of-the-Art Methods
+
+We compared the proposed PRM with four baselines. These include TripoSR [40], an open-source LRM implementation renowned for its superior single-view reconstruction performance; CRM [43], a UNet-based Convolutional Reconstruction Model that reconstructs 3D meshes from generated multi-view images and canonical coordinate maps (CCMs); LGM [39], a unet-based Large Gaussian Model
+
+Table 2. Comparison on GSO dataset. "Random m&r" indicates whether materials and lighting were randomly changed.  
+
+<table><tr><td>Method</td><td>Random m&amp;R</td><td>CD↓</td><td>FS@0.1↑</td><td>PSNR↑</td></tr><tr><td>InstantMesh</td><td>✓</td><td>0.061</td><td>0.934</td><td>21.115</td></tr><tr><td>InstantMesh</td><td>×</td><td>0.048</td><td>0.972</td><td>23.644</td></tr><tr><td>PRM</td><td>✓</td><td>0.053</td><td>0.982</td><td>24.602</td></tr><tr><td>PRM</td><td>×</td><td>0.043</td><td>0.991</td><td>26.377</td></tr></table>
+
+Table 3. Quantitative results of the ablation study.  
+
+<table><tr><td>Metric</td><td>CD↓</td><td>FS@0.1↑</td><td>PSNR↑</td></tr><tr><td>Only albedo supervision</td><td>0.099</td><td>0.887</td><td>17.532</td></tr><tr><td>w/o PBR</td><td>0.089</td><td>0.909</td><td>19.143</td></tr><tr><td>w/o lighting supervision</td><td>0.073</td><td>0.919</td><td>20.114</td></tr><tr><td>w/o change materials</td><td>0.089</td><td>0.894</td><td>19.662</td></tr><tr><td>Full model</td><td>0.066</td><td>0.948</td><td>20.992</td></tr></table>
+
+that reconstructs Gaussians from generated multi-view images; and InstantMesh [46], a transformer-based LRM that employs a two-stage training strategy for direct 3D mesh reconstruction. We reported both quantitative and qualitative comparative results for a complete comparison analysis.
+
+Quantitative Results. We reported quantitative results with randomly selected lighting and materials in two different datasets in Table 1. We also report quantitative results without changing materials and using fixed lighting as previous methods compared with cutting-edge method Instantmesh on GSO in Table 2, where ground-truth rendered multi-view images were used as input.
+
+For 3D reconstruction metrics, PRM achieves significant improvements over all state-of-the-art methods on both datasets, as shown in Table 1. The qualitative comparison with other methods is presented in Figure 3. We attribute these improvements to our use of images with varying materials and illumination for both input and supervision. This approach not only enables the model to learn fine-grained geometric details by providing photometric cues but also enhances the model's robustness to variations in image appearance. Further validation of these results in the test setting without material changes and using fixed lighting is shown in Table 2.
+
+For 2D visual metrics, our approach effectively mitigates the impact of lighting variations to accurately restore the original colors of objects, as shown in Table 1. We surpass all current methods across all metrics.
+
+![](images/cbd830ded9bdfbf34f137a9240560ddb58c33b76372fe94827013d2f808de2c3.jpg)  
+Figure 4. Visualization of our ablation study to validate the effectiveness of each component in our framework.
+
+![](images/26b18e524b265c93477290a05c04e49db406dc525dd20d33d08cd485dbd9818d.jpg)  
+Figure 5. Shading color offers significant photometric cues for perceiving geometry, whereas albedo lacks this information.
+
+Qualitative Results. For the qualitative comparison, we randomly selected four images from the GSO dataset to serve as inputs for 3D model reconstruction. For each reconstructed mesh, we visualized both the albedo (ours) and the shading color (others), as well as the pure geometry.
+
+As shown in Figure 3, the results reconstructed by PRM exhibit more accurate geometry and appearance. Our model can reconstruct precise geometry and accurately predict albedo from images with specular highlights, whereas other methods fail. For instance, InstantMesh often predicts uneven geometric surfaces and tends to reconstruct incorrect geometry. Similarly, both CRM and LGM struggle to produce satisfactory results, falling short in both geometry accuracy and texture prediction. This underscores the robustness and superior performance of our PRM method in handling complex lighting conditions and intricate surface details, making it a more reliable choice for high-quality 3D model reconstruction.
+
+# 4.4. Ablation Study
+
+We conducted the ablation study to validate the effectiveness of each component in our framework. We reported quantitative results of our ablation study in Table 3. For ablations, we reduce the training set from  $120\mathrm{K}$  to  $30\mathrm{K}$  for faster experiments due to the heavy cost.
+
+The effectiveness of PBR. PBR plays a crucial role in improving geometry, especially for glossy surfaces. To validate the effectiveness of PBR, we conducted experiments by directly predicting shading colors using an MLP, rather than deriving results with PBR. However, direct prediction of shading color, without accounting for view-dependent appearances, struggles to model such effects and thus fails to accurately reconstruct geometry. We presented these results as "w/o PBR" in Table 3 and Figure 4.
+
+Albedo supervision. To avoid the interference caused by specular color on the surface, an intuitive approach is to directly use albedo instead of shading color for supervision. However, this method proves ineffective in practice, as albedo contains few photometric cues, thereby hindering geometry reconstruction. For example, a concave surface
+
+![](images/092277a6cd64b7bea80cdeaa773606fc0566fcf1b11005f97e6618f590fc4fb5.jpg)  
+Figure 6. Ablation study of the effect of changing materials.
+
+with uniform albedo may appear planar without the presence of cast shadows, while shading color provides significant photometric cues critical for accurately perceiving geometry, as illustrated in Figure 5. We conducted an ablation study within our framework by excluding shading color and light maps from supervision. The qualitative comparison is shown in and Table 3 and Figure 4 denoted as "Only albedo supervision". The results demonstrate that photometric cues are crucial for accurate geometry reconstruction.
+
+Lighting maps supervision. We conducted the ablation study within our framework by excluding the lighting maps loss. The results, denoted as "w/o lighting supervision", are displayed in Table 3 and Figure 4. Since lighting maps are exclusively related to surface normals and do not include albedo, optimizing these maps proves particularly beneficial for enhancing the optimization of fine-grained local details.
+
+Variations in materials. We conducted an ablation study to illustrate the effectiveness of varying materials during training. The results are shown in Table 3 and Figure 6. Without changing the materials, some details are lost. Moreover, the model has lost the capability to reconstruct glossy surfaces.
+
+# 5. Conclusion and Limitation
+
+Limitation. Despite the high-quality results achieved in this work, there remain several limitations for future research to explore: 1) Firstly, the reconstructed 3D model is sensitive to the quality of multi-view images. 2) Secondly, the accuracy of the estimated albedo appears to be somewhat entangled with the lighting conditions.
+
+Conclusion. In this work, we introduce PRM, a novel feedforward framework designed to reconstructed high-quality 3D assets with fine-grained local details, even amidst complex image appearances. To achieve this goal, we utilize images with varying materials and illumination as both input and supervision, providing sufficient photometric cues for fine-grained geometry recovery and enhancing the model's robustness to variations in image appearance. Using a mesh as our 3D representation, we employ differentiable PBR for predicted rendering, underpinning the utilization of multiple supervisions for optimization. Experiments on public datasets validate that PRM surpasses other methods.
+
+# 6. Acknowledgement
+
+This research is supported by Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (No.2023B1212010007).
+
+# References
+
+[1] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 2014. 3  
+[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), 2021. 3  
+[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2022. 1  
+[4] Zhifei Chen, Tianshuo Xu, Wenhang Ge, Leyi Wu, Dongyu Yan, Jing He, Luozhou Wang, Lu Zeng, Shunsi Zhang, and Yingcong Chen. Uni-renderer: Unifying rendering and inverse rendering via dual stream diffusion. arXiv preprint arXiv:2412.15050, 2024.3  
+[5] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6  
+[6] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), 2022. 6  
+[7] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3, 5  
+[8] Jason Gregory. Game engine architecture. AK Peters/CRC Press, 2018. 1  
+[9] Hideki Hayakawa. Photometric stereo under a light source with arbitrary motion. JOSA A, 1994. 3  
+[10] Carlos Hernandez, George Vogiatzis, and Roberto Cipolla. Multiview photometric stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2008. 1, 3  
+[11] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. Routledge, 2013. 1, 4  
+[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1
+
+[13] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 2, 3  
+[14] Tao Hu, Wenhang Ge, Yuyang Zhao, and Gim Hee Lee. Xray: A sequential 3d representation for generation, 2024. 2  
+[15] Tao Hu, Wenhang Ge, Yuyang Zhao, and Gim Hee Lee. Xray: A sequential 3d representation for generation. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2  
+[16] Satoshi Ikehata. Universal photometric stereo network using global lighting contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3  
+[17] Satoshi Ikehata. Scalable, detailed and mask-free universal photometric stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3  
+[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 3  
+[19] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 2013. 2, 4  
+[20] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Uncertainty-aware deep multi-view photometric stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 5  
+[21] Berk Kaya, Suryansh Kumar, Francesco Sarno, Vittorio Ferrari, and Luc Van Gool. Neural radiance fields approach to deep multi-view photometric stereo. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (WACV), 2022. 3  
+[22] Berk Kaya, Suryansh Kumar, Carlos Oliveira, Vittorio Ferrari, and Luc Van Gool. Multi-view photometric stereo revisited. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 3, 5  
+[23] John Lasseter. Principles of traditional animation applied to 3d computer animation. In Proceedings of the 14th annual conference on Computer graphics and interactive techniques, 1987. 1  
+[24] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2  
+[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. arXiv preprint arXiv:2311.11284, 2023. 1  
+[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1
+
+[27] Jianman Lin, Haojie Li, Chunmei Qing, Zhijing Yang, Liang Lin, and Tianshui Chen. Geometry-editable and appearance-preserving object compositon. arXiv preprint arXiv:2505.20914, 2025. 1  
+[28] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 3  
+[29] Zhipeng Mo, Boxin Shi, Feng Lu, Sai-Kit Yeung, and Yasuyuki Matsushita. Uncalibrated photometric stereo under natural illumination. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3  
+[30] Gregory M Nielson. Dual marching cubes. In IEEE visualization 2004, 2004. 5  
+[31] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 2019. 3  
+[32] Rick Parent. Computer animation: algorithms and techniques. Newnes, 2012. 1  
+[33] Jaesik Park, Sudipta N Sinha, Yasuyuki Matsushita, Yu-Wing Tai, and In So Kweon. Multiview photometric stereo using planar mesh parameterization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2013. 3  
+[34] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1  
+[35] Martijn J. Schuemie, Peter Van Der Straaten, Merel Krijn, and Charles APG Van Der Mast. Research on presence in virtual reality: A survey. Cyberpsychology & behavior, 2001. 1  
+[36] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Trans. Graph., 2023. 1, 5  
+[37] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 7  
+[38] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1  
+[39] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 1, 2, 7  
+[40] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 7  
+[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields.
+
+In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3, 5  
+[42] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1  
+[43] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 1, 2, 7  
+[44] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 1980. 3  
+[45] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6  
+[46] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 1, 2, 5, 7  
+[47] Xinli Xu, Wenhang Ge, Jiantao Lin, Jiawei Feng, Lie Xu, HanFeng Zhao, Shunsi Zhang, and Ying-Cong Chen. Flexgen: Flexible multi-view generation from text and image inputs. arXiv preprint arXiv:2410.10745, 2024. 2  
+[48] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217, 2023. 1, 2  
+[49] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1  
+[50] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems (NeurIPS), 2021. 1  
+[51] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. arXiv preprint arXiv:2406.13897, 2024. 1, 2  
+[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018. 5  
+[53] Dongxu Zhao, Daniel Lichy, Pierre-Nicolas Perrin, Jan-Michael Frahm, and Soumyadip Sengupta. Mvpsnet: Fast generalizable multi-view photometric stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2023. 3

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b727a3eadfccd7f380bd1300082ef958115fd8dcd72c0702c93a182f7e07e15
+size 544775

ICCV/2025/PRM_ Photometric Stereo based Large Reconstruction Model/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c8f1315ccd98fa75631719392ad316510e3556d525aaefa3f61969e14ba4ebe
+size 354250

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/705f3e43-5f3f-4db6-89b2-0f1e38fc75e9_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:406581c7feec6c56d4e05a33f89855f1bce02958597a01b68bc232faabd4fa42
+size 92690

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/705f3e43-5f3f-4db6-89b2-0f1e38fc75e9_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1b03785bb86297a690dcafeb2316bdd474ea8089ecef401e274c8ce6a2e6cb7
+size 116247

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/705f3e43-5f3f-4db6-89b2-0f1e38fc75e9_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69e0bb3e06edfa6f1f37540e9edae658980135fe06dc2af9fdd6e2aa169b595c
+size 3461750

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/full.md

@@ -0,0 +1,372 @@
+# PRO-VPT: Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation
+
+Chikai Shang $^{1,2}$  Mengke Li $^{3,5}$  Yiqun Zhang $^{4,5}$  Zhen Chen $^{6}$  Jinlin Wu $^{6,7}$  Fangqing Gu $^{4,5*}$  Yang Lu $^{1,2,5*}$  Yiu-ming Cheung $^{5}$
+
+<sup>1</sup>Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University
+
+$^{2}$ School of Informatics, Xiamen University  $^{3}$ Shenzhen University  $^{4}$ Guangdong University of Technology
+
+$^{5}$ Hong Kong Baptist University  $^{6}$ CAIR, HKISI, CAS  $^{7}$ MAIS, CASIA
+
+ckshang12@gmail.com fqgu@gdut.edu.cn luyang@xmu.edu.cn
+
+# Abstract
+
+Visual prompt tuning (VPT), i.e., fine-tuning some lightweight prompt tokens, provides an efficient and effective approach for adapting pre-trained models to various downstream tasks. However, most prior art indiscriminately uses a fixed prompt distribution across different tasks, neglecting the importance of each block varying depending on the task. In this paper, we introduce adaptive distribution optimization (ADO) by tackling two key questions: ① How to appropriately and formally define ADO, and ② How to design an adaptive distribution strategy guided by this definition? Through empirical analysis, we first confirm that properly adjusting the distribution significantly improves VPT performance, and further uncover a key insight that a nested relationship exists between ADO and VPT. Based on these findings, we propose a new VPT framework, termed PRO-VPT (iterative Prompt Relocation-based VPT), which adaptively adjusts the distribution built upon a nested optimization formulation. Specifically, we develop a prompt relocation strategy derived from this formulation, comprising two steps: pruning idle prompts from prompt-saturated blocks, followed by allocating these prompts to the most prompt-needed blocks. By iteratively performing prompt relocation and VPT, our proposal can adaptively learn the optimal prompt distribution in a nested optimization-based manner, thereby unlocking the full potential of VPT. Extensive experiments demonstrate that our proposal significantly outperforms advanced VPT methods, e.g., PRO-VPT surpasses VPT by 1.6 pp and 2.0 pp average accuracy, leading prompt-based methods to state-of-the-art performance on VTAB-1k and FGVC benchmarks. The code is available at https://github.com/ckshang/PRO-VPT.
+
+# 1. Introduction
+
+Fine-tuning pre-trained vision models (PVMs) [26] has proven remarkably effective in adapting to a variety of
+
+![](images/827bc2189f20a003217341ba80cce94f37759915d6021c21de68ab428c0edcff.jpg)  
+Figure 1. PRO-VPT (ours) vs. prior art in VPT. Existing VPT approaches typically insert trainable prompts into the PVM with a pre-specific static distribution, whether shallow or deep, and optimize these prompts to drive the PVM to conduct downstream tasks. Compared to the prior art, our proposal (PRO-VPT) adaptively adjusts prompt distribution by treating it as an optimization objective and coupling the distribution optimization with prompt tuning.
+
+downstream tasks [14, 30, 65]. However, the computational and storage costs associated with full fine-tuning are prohibitively high, particularly as model sizes continue to grow [1, 40, 57]. To overcome these challenges, parameter-efficient fine-tuning (PEFT) methods [2, 20, 61, 62], such as visual prompt tuning (VPT) [22], have emerged as more promising alternatives, garnering significant attention.
+
+Instead of updating all model parameters directly, VPT inserts lightweight learnable prompt tokens into the input space of Transformer blocks, with either a shallow distribution (inserted only in the first block) or a deep distribution (uniformly inserted across all blocks), while keeping the PVM frozen. This approach significantly reduces the number of parameters that need to be fine-tuned, thereby lowering the computational and memory overhead [14]. Follow-up
+
+research further enhances this pipeline by improving aspects such as prompting architecture [13, 50, 52], prompt initialization [58], and prompt propagation structure [5, 63, 71].
+
+Despite these advancements, most of the prior art still relies on a preset fixed prompt distribution, either shallow or deep, across various downstream tasks, as illustrated in Fig. 1. However, recent findings reveal that the importance of each pre-trained block varies significantly across different tasks [12, 28, 43, 49, 60], underscoring the need to dynamically tailor the fine-tuning intensity for each block to improve task performance [11, 25, 29, 45, 53]. This implies that indiscriminately using a fixed prompt distribution may not fully exploit the potential of VPT, motivating us to adaptively adjust the distribution to better accommodate each task.
+
+In this paper, we naturally consider prompt distribution as one of the optimization objectives, which we refer to as adaptive distribution optimization (ADO). As the focus pivots towards the ADO problem, new questions arise. First, since the underlying nature of distribution optimization is still unexplored, the definition of ADO remains unclear. For instance, defining it as a one-shot process or an iterative one leads to a critical consideration, namely whether the effectiveness of distribution adjustments is influenced by ongoing prompt updates. This leads us to our first key question: ① How can we appropriately and formally define ADO? (§ 3.2) Second, once the definition of ADO is established, a more essential question emerges: ② How can we achieve adaptive distribution derived from this definition? (§ 3.3)
+
+To address these questions, we first empirically analyze the ADO problem. Our findings confirm the significance of ADO and further reveal two key insights: the effectiveness of distribution adjustments is indeed impacted by prompt updates, and a nested relationship exists between ADO and VPT. Based on these insights, we propose PRO-VPT (iterative Prompt RelOcation-based VPT), a novel iterative framework derived from a nested optimization definition. Specifically, building on this definition, we develop a prompt relocation (PR) strategy for ADO, which comprises two optimization steps: pruning and allocating. The pruning step first identifies and releases idle prompts from prompt-saturated blocks, while the subsequent allocation step determines the most prompt-needed blocks for redistributing these prompts. By iteratively performing PR and VPT, PRO-VPT can adaptively learn the optimal prompt distribution for each task in a nested optimization manner, thereby unlocking the full potential of VPT. Our main contributions are fourfold.
+
+- We explore the ADO problem within VPT and reveal the nested relationship between ADO and VPT.  
+- Building upon our insights into prompt distribution, we propose the first ADO-VPT co-design framework, termed PRO-VPT, with a nested optimization definition.  
+- Derived from the definition, we present a PR strategy for ADO, which introduces an idleness score to prune under
+
+utilized prompts from prompt-saturated blocks and then allocates them to prompt-needed blocks using reinforcement learning (RL).
+
+- Extensive experiments show that PRO-VPT exhibits superior performance, e.g., it attains state-of-the-art average accuracies of  $78.0\%$  on VTAB-1k and  $91.7\%$  on FGVC, with 1.6 pp and 2.0 pp advantages over the VPT baseline.
+
+# 2. Related Work
+
+Visual Prompt Tuning. VPT [22], one of the leading techniques in PEFT, draws inspiration from language-domain prompting works [19, 32, 37, 38] to fine-tune a small portion of prompts while keeping the pre-trained model frozen. Subsequent research has improved VPT in three main aspects: (1) Extending prompts to other Transformer components (e.g., key and value matrices or the patch embedding layer) to better suit the Transformer architecture [13, 50, 52]; (2) Initializing prompts with prototypes of image tokens to accelerate convergence [58]; (3) Refining the propagation structures of prompts to strengthen cross-layer interactions [5, 63, 71]. However, most existing methods depend on a fixed prompt distribution for all tasks, resulting in less effective prompting performance. In contrast, our approach adaptively adjusts the prompt distribution for each task, enabling prompt-based methods to achieve state-of-the-art performance.
+
+Layer-wise Fine-Tuning. Recent research has highlighted the importance of treating each pre-trained block differently during downstream fine-tuning, emphasizing the need for layer-wise adaptive strategies to calibrate the fine-tuning intensity for each block [12, 25, 31, 43, 54]. For instance, [29] first revealed that some pre-trained blocks may already be near-optimal for specific tasks, indicating that fine-tuning demands of each block vary significantly for different tasks. This work then proposes freezing these near-optimal blocks and selectively fine-tuning the remaining ones to improve performance. Similarly, [45, 49] proposes to adaptively assign different learning rates to each block, tailoring the fine-tuning intensity at a layer level to achieve more fine-grained layer-wise adaptation. Inspired by these findings, we hypothesize that the prompting intensity for each block, i.e., the overall distribution of prompts, may also play a crucial role in task adaptation, leading us to focus on adaptively adjusting the prompt distribution within VPT.
+
+# 3. Methodology
+
+# 3.1. Notation
+
+Visual prompt tuning (VPT) aims to learn lightweight tunable prompts  $\mathbf{P}$  that are inserted into the vision Transformer (ViT)  $f(\cdot)$  with a specific distribution  $\mathcal{D}$ . In particular, for a set of  $N$  prompts  $\mathbf{P} = \{\mathbf{p}_k\}_{k=1}^N$  and their distributed blocks  $\mathcal{D} = \{d_k\}_{k=1}^N$  where  $d_k \in \{0,1,\dots,L\}$  ( $d_k = 0$  for undistributed), the prompted ViT model is denoted as  $f_{\mathbf{P},\mathcal{D}}(\cdot)$ . Given a training set  $\mathcal{T}_{tr}$  of input-label pairs  $(\mathbf{x},y)$ , the model
+
+![](images/fabb879534ef035f291e9250b930bf54e09c5b6477a2229003a717896d609d24.jpg)  
+Figure 2. Performance gaps from distribution adjustments using prompts from epochs 25, 50, and 75. Adjusting prompt distribution appropriately leads to enhanced performance; however, effective adjustments vary significantly across different epochs.
+
+is then trained by minimizing the task loss  $\mathcal{L}(f_{\mathbf{P},\mathcal{D}}(\mathbf{x}),y)$ . For convenience in later sections, we define the loss difference as  $\Delta \mathcal{L}(f_1,f_2) = \mathcal{L}(f_1(\mathbf{x}),y) - \mathcal{L}(f_2(\mathbf{x}),y)$ .
+
+# 3.2. Analysis and Formulation of ADO
+
+We first explore the underlying nature of ADO and its formal definition, in response to question ①.
+
+Intuitively, ADO can be defined in two distinct ways: one-shot (independently adjusting the distribution) or iterative (integrating with prompt tuning through iteratively adjusting the distribution). These definitions are distinct in whether the distribution is adjusted concurrently with prompt tuning. Specifically, an iterative process is more appropriate when distribution adjustments are affected by ongoing changes in prompts, as independent optimization may lead to suboptimal results by failing to account for these dynamics. Otherwise, one-shot adjustments may be sufficient.
+
+To investigate which definition better matches the underlying nature, we experiment to examine how distribution adjustments behave when applied to prompts from different epochs. We use two intuitive adjustment strategies: incrementally adding to a small initial prompt set and repositioning existing prompts. As shown in Fig. 2, adding or repositioning prompts to specific blocks can effectively improve performance; however, the effective adjustments differ significantly when prompts are updated. Comprehensive results with detailed analysis are provided in Appendix C. In conclusion, we summarize our key findings as follows:
+
+Finding 1. Properly adjusting the prompt distribution significantly enhances performance, confirming the necessity to reconsider its optimization.
+
+Finding 2. Distribution adjustments are sensitive not only to different blocks but also to the updated prompts themselves, highlighting the need for an iterative process that continuously adjusts the distribution during training.
+
+Finding 3. The effectiveness of distribution adjustments can
+
+only be properly evaluated after prompt tuning, indicating that the prompt tuning process should be nested within the distribution optimization process (see Appendix C).
+
+Building on these insights, we propose integrating ADO with VPT into an iterative optimization framework. Formally, we present the definition of this ADO-VPT co-design framework through the lens of nested optimization, where the objectives of ADO and VPT are expressed as:
+
+$$
+\mathcal {D} ^ {*} = \underset {\mathcal {D}} {\arg \min } \mathbb {E} _ {(\mathbf {x}, y) \in \mathcal {T} _ {t r}} \left[ \mathcal {L} \left(f _ {\mathbf {P} ^ {*}, \mathcal {D}} (\mathbf {x}), y\right) \right], \tag {1}
+$$
+
+$$
+\mathbf {P} ^ {*} = \underset {\mathbf {P}} {\arg \min } \mathbb {E} _ {(\mathbf {x}, y) \in \mathcal {T} _ {t r}} \left[ \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D} ^ {*}} (\mathbf {x}), y\right) \right]. \tag {2}
+$$
+
+This formulation clearly illustrates the intertwined relationship between visual prompts  $\mathbf{P}$  and their distribution  $\mathcal{D}$ , with its framework presented in Fig. iii. In this framework, prompt tuning is integrated into the ADO process, forming a closed-loop workflow of 'distribution adjustment  $\rightarrow$  prompt tuning  $\rightarrow$  adjustment evaluation  $\rightarrow$  new adjustment', thereby achieving a co-design process aligning with the observed iterative and nested nature.
+
+Remark. This nested optimization framework offers a universal approach capable of incorporating various distribution adjustment strategies, e.g., adding or repositioning. However, we empirically observe that adding-based adjustments often lead to unstable performance (see Appendix D). Therefore, in the following section, we focus on developing a prompt relocation (PR) strategy based on this formulation to effectively tackle the ADO problem.
+
+# 3.3. Iterative Prompt Relocation-based VPT
+
+Following the relocation objective defined in Eq. (1), we introduce the PR strategy derived from this definition and provide a detailed overview of the PRO-VPT framework, in response to question ②.
+
+Building on the aforementioned findings, the discrete PR objective is further complicated by the non-stationary nature of effective relocation (Finding 2) and the necessity to predict the effectiveness of relocation operations (Finding 3). Nonetheless, these three properties make a RL framework [24, 33] a natural and well-suited approach for optimizing the PR objective. However, a naive RL implementation needs to model all possible relocation operations as actions, resulting in  $L^2$  possible arrangements (e.g., 144 actions for ViT-B). Such a large action space necessitates numerous iterations to achieve convergence (cf. § 4.3).
+
+To address this, we propose decomposing the PR process into two sequential steps: pruning and allocating. The pruning step first identifies and releases idle prompts for relocation, while the allocating step then employs RL to determine optimal blocks for these freed prompts. By focusing solely on allocatable blocks instead of all block arrangements, this approach effectively reduces the action space from  $L^2$  to  $L$ . More importantly, this approach has an intuitive explanation: idle prompts typically exist within the prompt-saturated
+
+![](images/99e613aac9f769b9798fb836d99510999807b463e1f5b0720ecc8488a0c954b9.jpg)  
+Figure 3. Overview of our proposed PRO-VPT. Left: The streamlined workflow of PRO-VPT. Right: Illustration of the PR process.
+
+![](images/e334a4a16adad44f6c9b4b11adb3492cac7637121733737b24297a63bf388b85.jpg)
+
+blocks, thus relocating them to prompt-needed blocks may enhance the effectiveness of prompt distribution. Decomposed from the relocation objective in Eq. (1), the pruning and allocation steps<sup>1</sup> can be formalized as:
+
+$$
+k ^ {*} = \underset {k} {\arg \max } \mathbb {E} _ {(\mathbf {x}, y) \in \mathcal {T} _ {t r}} \left[ \Delta \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D}}, f _ {\mathbf {P}, \mathcal {D} \mid d _ {k} = 0}\right) \right], \tag {3}
+$$
+
+$$
+a ^ {*} = \underset {a} {\arg \max } \mathbb {E} _ {(\mathbf {x}, y) \in \mathcal {T} _ {t r}} \left[ \Delta \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D} ^ {-}}, f _ {\mathbf {P} ^ {\prime}, \mathcal {D} | d _ {k ^ {*}} = a}\right) \right], \tag {4}
+$$
+
+where  $\mathcal{D}^{-} = \{d_{k}\in \mathcal{D}|d_{k^{*}} = 0\}$  and  $\mathcal{D}^{+} = \{d_{k}\in \mathcal{D}|d_{k^{*}} = a^{*}\}$  denote the prompt distributions after pruning and allocation with  $a\in [L]$  as the allocation block, and the allocation step uses the tuned prompts  $\mathbf{P}'$  for evaluation following Finding 3. These two equations are specifically designed to identify the  $k^*$ -th prompt suitable for pruning and to determine the optimal block  $a^*$  for allocating.
+
+Following this, we utilize Eqs. (3) and (4) to approach the pruning and allocation stages, respectively, with the overall PRO-VPT framework illustrated in Fig. 3 and Algo. i:
+
+Idleness Score-based Pruning. To measure whether a prompt is idle for its current block and suitable for relocation, we take Eq. (3) as the idleness score:
+
+$$
+\mathcal {I} _ {k} = \Delta \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D}}, f _ {\mathbf {P}, \mathcal {D} \mid d _ {k} = 0}\right). \tag {5}
+$$
+
+Intuitive Interpretation. If  $\mathcal{I}_k = \Delta \mathcal{L} > 0$ , pruning the prompt  $\mathbf{p}_k$  from its current block can enhance performance. This indicates that  $\mathbf{p}_k$  is negative for its current block, which tends to be prompt-saturated and therefore suitable for its relocation to a more prompt-needed block. Conversely,  $\mathcal{I}_k \leq 0 \forall k$  suggests that no blocks are saturated and thus the prompt distribution is near-optimal.
+
+Pruning Implementation. Since computing  $\mathcal{I}_k$  for each individual prompt  $\mathbf{p}_k$  is significantly inefficient, we approximate  $\mathcal{I}_k$  using its first-order Taylor expansion (see Appendix F for detailed derivation):
+
+$$
+\widehat {\mathcal {I}} _ {k} \approx \mathbf {g} _ {k} ^ {T} \mathbf {p} _ {k}, \tag {6}
+$$
+
+where  $\mathbf{g}_k = \frac{\nabla\mathcal{L}}{\nabla\mathbf{p}_k}$  represents the elements of the gradient  $\mathbf{g}$ .
+
+To this end, by calculating the estimated scores as in Eq. (6), we can effectively identify and prune the idle prompt  $\mathbf{p}_{k^*}$  with the maximum  $\widehat{\mathcal{I}}_{k^*}$  from the prompt-saturated block (when  $\max \widehat{\mathcal{I}}_k > 0$ ), thereby fulfilling Eq. (3) and forming the pruned distribution  $\mathcal{D}^{-}$ .
+
+RL-based Allocation. After pruning an idle prompt, the following key step is to leverage RL to determine its optimal allocation to the prompt-needed block. Specifically, we take Eq. (4) as the optimization objective for the RL problem, i.e., the allocation reward, as follows:
+
+$$
+r = \Delta \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D} ^ {-}}, f _ {\mathbf {P} ^ {\prime}, \mathcal {D} | d _ {k ^ {*} = a}}\right). \tag {7}
+$$
+
+Intuitive Interpretation. The optimal allocation block  $a^*$ , which achieves the maximum  $r^* = \max \Delta \mathcal{L}$ , is identified as the block where assigning a prompt leads to the highest performance improvement, implying that this block is the most prompt-needed.
+
+Allocation Implementation. We detail the RL-based allocation under a Markov decision process framework, where the main components are:
+
+(1) State. We leverage all available information to represent the state of the current prompt distribution. Specifically, the state incorporates the block-wise idleness scores, the current distribution, and the pruned prompt's location, formally defined as  $s = \{\sum_{d_k = i}\widehat{\mathcal{I}}_k\}_{i = 1}^L\bigcup \mathcal{D}^{-}\bigcup \mathrm{OneHot}(k^*)$ . After PR and VPT, the state transitions to  $s'$ .  
+(2) Action. The action determines which block the pruned prompt will be allocated to, represented as  $a \in [L]$ .  
+(3) Reward. Although the allocation reward  $r$  is defined in Eq. (7), calculating the intermediate loss of  $f_{\mathbf{P},\mathcal{D}^{-}}$  is computationally inefficient. To address this, we approximate  $r$  by incorporating the idleness score as:
+
+$$
+\hat {r} \approx \Delta \mathcal {L} \left(f _ {\mathbf {P}, \mathcal {D}}, f _ {\mathbf {P} ^ {\prime}, \mathcal {D} | d _ {k ^ {*} = a}}\right) - \widehat {\mathcal {I}} _ {k ^ {*}}, \tag {8}
+$$
+
+where the idleness score  $\widehat{\mathcal{I}}_{k^*} \approx \Delta \mathcal{L}(f_{\mathbf{P},\mathcal{D}}, f_{\mathbf{P},\mathcal{D}^-})$  provides an approximation of the intermediate loss of  $f_{\mathbf{P},\mathcal{D}^-}$ .
+
+To this end, by leveraging RL to maximize the estimated reward as in Eq. (8), we can efficiently determine the optimal allocation to the prompt-needed block  $a^{*}$  even within the non-stationary environment, thereby fulfilling Eq. (4)
+
+and forming the relocated distribution  $\mathcal{D}^+$ . Particularly, we accomplish this by employing proximal policy optimization (PPO) [48], a state-of-the-art RL algorithm.
+
+Remark. Although the relocation process is decomposed into two-step sequential steps, the overall optimization objective, whether based on the expected objectives in Eqs. (3) and (4) or the estimated objectives in Eqs. (6) and (8), can be expressed as  $\max_{k,a} \mathbb{E}_{(\mathbf{x},y) \in \mathcal{T}_{tr}}[\Delta \mathcal{L}(f_{\mathbf{P},\mathcal{D}}, f_{\mathbf{P}'', \mathcal{D}|d_k = a})]$ . This formulation remains equivalent to the original PR objective (i.e., the ADO objective) defined in Eq. (1). Consequently, leveraging the nested optimization formulation, the proposed PR strategy can effectively address the ADO problem and adaptively construct the optimal distribution.
+
+# 3.4. Discussion
+
+We note that some prior works have indirectly achieved adaptive distribution [13, 27, 67]. Specifically, pruning-based techniques [13, 27] introduce mask variables to iteratively prune redundant prompts from the original distribution, and NOAA [67] uses an evolutionary algorithm to search for an optimal sub-distribution within the original distribution in a one-shot manner. PRO-VPT is distinct and advantageous compared to these methods as:
+
+Distinct Goals. These methods aim to enhance performance through the lottery ticket hypothesis [8, 41, 47], focusing on searching for a more effective sub-distribution within a given hyper-distribution by circumventing negative prompts and retaining essential ones. However, they are limited to pruning prompts from prompt-saturated blocks and fail to tackle prompt-deficient blocks (cf. § 4.5). In contrast, our approach actively learns an optimal distribution of all prompts, adaptively relocating them from prompt-saturated blocks to prompt-needed blocks through our PR strategy, thereby maximizing the effectiveness of the entire prompt distribution. This fundamental distinction sets our approach apart from the methods discussed above.
+
+Efficient. While effective, we argue that these approaches typically incur considerable unnecessary training costs to identify an optimal sub-distribution, as numerous trained prompts are pruned and left unused (cf. § 4.3). This fundamentally contradicts the goal of parameter efficiency. Conversely, our work directly adjusts the distribution of the entire prompt set through prompt relocation, thereby minimizing extra costs and improving efficiency.
+
+Effective. These methods straightforwardly apply one-shot or iterative processes without considering the underlying distribution dynamics. For instance, NOAA [67] adopts a one-shot approach post-training, leading to suboptimal results (cf. § 4.2). Conversely, our work conducts empirical analyses and accordingly proposes a nested optimization-based method, enabling prompt-based methods to reach state-of-the-art performance.
+
+# 4. Experiments
+
+# 4.1. Experiment Setups
+
+Datasets. We evaluated our proposed PRO-VPT on a series of datasets grouped into three categories: (1) VTAB-1k. The visual task adaptation benchmark (VTAB) [66] encompasses 19 classification tasks categorized into three group: Natural, Specialized, and Structured, aiming to assess adaptation capabilities under low-shot training conditions. (2) FGVC. Fine-grained visual classification (FGVC) includes 5 classification tasks: CUB-200-2011 [56], NABirds [55], Oxford Flowers [42], Stanford Dogs [6], and Stanford Cars [10], aiming to assess task adaptation in large-scale, fine-grained settings. (3) COCO and ADE20k. These widely-used benchmarks for detection and segmentation are employed to assess the generalizability of our method across diverse downstream visual tasks.
+
+Implementation Details. We mainly use ViT-B/16 [7] supervised pre-trained on ImageNet-21k [46] as the initialization. We employ the SGD optimizer with a batch size of 64 and fine-tune models for 100 epochs. The number of prompt tokens per task in our method is set in accordance with [22]. Following [51], we evaluate our method using Inception normalization instead of ImageNet normalization, and for fair comparisons, we present the results of other counterparts using both normalization techniques. Notably, unlike most approaches [9, 22, 34, 68], our method employs a fixed learning rate instead of using a learning rate schedule, as we necessitate ensuring timely rewinding after prompt relocation [44]. More details about the hyper-parameters are provided in Appendices G.
+
+# 4.2. Main Results
+
+VTAB-1k. We evaluate the proposed PRO-VPT on VTAB-1k and compare it with other methods. We provide results for full fine-tuning and linear probing as baselines of traditional fine-tuning. As competing advanced adapter- or reparameterization-based methods, we include SSF [34], SPT-Adapter [16], Adapter+ [51], LoRA [20], FacT-TK [23], and Consolidator [15]. Furthermore, we benchmark against state-of-the-art prompt-based methods including VPT [22], NOAA [67], SPT [58], and iVPT [71].
+
+Tab. 1 presents results on VTAB-1k with the best-performing normalization techniques (i.e., ImageNet or Inception) for each method (see Appendix H for complete results). Among all prompt-based methods, the previous best performance of  $77.1\%$  average accuracy achieved by iVPT falls short of the advanced adapter-based method Adapter+, which reaches  $77.6\%$ . Nevertheless, with the introduction of the optimization-based adaptive distribution, the proposed PRO-VPT sets a new state-of-the-art performance among all evaluated methods with an average accuracy of  $78.0\%$ , highlighting the significant potential of prompt-based approaches. Additionally, compared to the one-shot optimization-based
+
+Table 1. Detailed results on the VTAB-1k datasets. Performance results are reported as the highest of ImageNet normalization (○) or Inception normalization (●), presented in % after a complete training schedule with ViT-B/16 supervised pre-trained on ImageNet-21k. The best results of prompt-based methods and other PEFT approaches are highlighted in bold. f: Early-stopping based on the test set. †: Lack of complete code or hyperparameter configurations for the method, hence results are reported as presented in the original paper. ¹Average across the average accuracies of the VTAB-1k groups, following previous work.  
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Param (M)</td><td colspan="7">Natural</td><td colspan="5">Specialized</td><td colspan="7">Structured</td><td>Global Avg.1</td><td></td><td rowspan="2"></td><td></td></tr><tr><td>Cifar100</td><td>Caltech101</td><td>DTD</td><td>Flower102</td><td>Pets</td><td>SVHN</td><td>Sun397</td><td colspan="4">Group Avg.</td><td>Group Avg.</td><td>Clev-Count</td><td>Clev-Dist.</td><td>DMLab</td><td>KITTI-Dist.</td><td>dSpr-Loc.</td><td>dSpr-Ori.</td><td>sNORB-Azi.</td><td>sNORB-Ele.</td><td>Group Avg.</td><td></td></tr><tr><td>Full •</td><td>85.8</td><td>73.2</td><td>92.6</td><td>70.4</td><td>97.9</td><td>86.2</td><td>90.6</td><td>39.6</td><td>78.6</td><td>87.1</td><td>96.6</td><td>87.5</td><td>74.0</td><td>86.3</td><td>66.6</td><td>61.0</td><td>49.8</td><td>79.7</td><td>82.6</td><td>51.9</td><td>33.5</td><td>37.0</td><td>57.8</td><td>74.2</td></tr><tr><td>Linear •</td><td>0.04</td><td>78.1</td><td>88.1</td><td>69.0</td><td>99.1</td><td>90.0</td><td>36.0</td><td>56.9</td><td>73.9</td><td>79.8</td><td>90.7</td><td>73.7</td><td>73.7</td><td>79.5</td><td>32.4</td><td>30.5</td><td>35.9</td><td>61.9</td><td>11.2</td><td>26.2</td><td>14.3</td><td>24.5</td><td>29.6</td><td>61.0</td></tr><tr><td>LoRA • [20]</td><td>0.29</td><td>83.0</td><td>91.7</td><td>71.6</td><td>99.2</td><td>90.9</td><td>83.8</td><td>56.7</td><td>82.4</td><td>86.2</td><td>95.7</td><td>83.5</td><td>71.9</td><td>84.3</td><td>77.7</td><td>62.3</td><td>49.0</td><td>80.2</td><td>82.2</td><td>51.7</td><td>31.0</td><td>47.0</td><td>60.1</td><td>75.6</td></tr><tr><td>FacT-TKs • [23]</td><td>0.05</td><td>74.9</td><td>92.7</td><td>73.7</td><td>99.1</td><td>91.3</td><td>85.5</td><td>57.7</td><td>82.1</td><td>86.8</td><td>94.9</td><td>84.1</td><td>70.9</td><td>84.2</td><td>81.9</td><td>64.1</td><td>49.2</td><td>77.2</td><td>83.8</td><td>53.1</td><td>28.2</td><td>44.7</td><td>60.3</td><td>75.5</td></tr><tr><td>FacT-TK≤32 • [23]</td><td>0.10</td><td>74.6</td><td>93.7</td><td>73.6</td><td>99.3</td><td>90.6</td><td>88.7</td><td>57.5</td><td>82.6</td><td>87.6</td><td>95.4</td><td>85.5</td><td>70.4</td><td>84.7</td><td>84.3</td><td>62.6</td><td>51.9</td><td>79.2</td><td>85.5</td><td>52.0</td><td>36.4</td><td>46.6</td><td>62.3</td><td>76.5</td></tr><tr><td>Consolidator † [15]</td><td>0.30</td><td>74.2</td><td>90.9</td><td>73.9</td><td>99.4</td><td>91.6</td><td>91.5</td><td>55.5</td><td>82.4</td><td>86.9</td><td>95.7</td><td>86.6</td><td>75.9</td><td>86.3</td><td>81.2</td><td>68.2</td><td>51.6</td><td>83.5</td><td>79.8</td><td>52.3</td><td>31.9</td><td>38.5</td><td>60.9</td><td>76.5</td></tr><tr><td>SSF f o [34]</td><td>0.24</td><td>69.0</td><td>92.6</td><td>75.1</td><td>99.4</td><td>91.8</td><td>90.2</td><td>52.9</td><td>81.6</td><td>87.4</td><td>95.9</td><td>87.4</td><td>75.5</td><td>86.6</td><td>75.9</td><td>62.3</td><td>53.3</td><td>80.6</td><td>77.3</td><td>54.9</td><td>29.5</td><td>37.9</td><td>59.0</td><td>75.7</td></tr><tr><td>SPT-Adapter f o [16]</td><td>0.23</td><td>72.9</td><td>93.2</td><td>72.5</td><td>99.3</td><td>91.4</td><td>84.6</td><td>55.2</td><td>81.3</td><td>85.3</td><td>96.0</td><td>84.3</td><td>75.5</td><td>85.3</td><td>82.2</td><td>68.0</td><td>49.3</td><td>80.0</td><td>82.4</td><td>51.9</td><td>31.7</td><td>41.2</td><td>60.8</td><td>75.8</td></tr><tr><td>SPT-Adapter f o [16]</td><td>0.43</td><td>72.9</td><td>93.2</td><td>72.5</td><td>99.3</td><td>91.4</td><td>88.8</td><td>55.8</td><td>82.0</td><td>86.2</td><td>96.1</td><td>85.5</td><td>75.5</td><td>85.8</td><td>83.0</td><td>68.0</td><td>51.9</td><td>81.2</td><td>82.4</td><td>51.9</td><td>31.7</td><td>41.2</td><td>61.4</td><td>76.4</td></tr><tr><td>Adapter+ r=16 • [51]</td><td>0.35</td><td>83.7</td><td>94.2</td><td>71.5</td><td>99.3</td><td>90.6</td><td>88.2</td><td>55.8</td><td>83.3</td><td>87.5</td><td>97.0</td><td>87.4</td><td>72.9</td><td>86.2</td><td>82.9</td><td>60.9</td><td>53.7</td><td>80.8</td><td>88.4</td><td>55.2</td><td>37.3</td><td>46.9</td><td>63.3</td><td>77.6</td></tr><tr><td colspan="25">Prompt-based Methods:</td></tr><tr><td>VPT-Deep • [22]</td><td>0.60</td><td>83.0</td><td>93.0</td><td>71.2</td><td>99.0</td><td>91.3</td><td>84.1</td><td>56.0</td><td>82.5</td><td>84.9</td><td>96.6</td><td>82.5</td><td>74.5</td><td>84.6</td><td>77.5</td><td>58.7</td><td>49.7</td><td>79.6</td><td>86.2</td><td>56.1</td><td>37.9</td><td>50.7</td><td>62.1</td><td>76.4</td></tr><tr><td>NOAH f o [67]</td><td>0.43</td><td>69.6</td><td>92.7</td><td>70.2</td><td>99.1</td><td>90.4</td><td>86.1</td><td>53.7</td><td>80.2</td><td>84.4</td><td>95.4</td><td>83.9</td><td>75.8</td><td>84.9</td><td>82.8</td><td>68.9</td><td>49.9</td><td>81.7</td><td>81.8</td><td>48.3</td><td>32.8</td><td>44.2</td><td>61.3</td><td>75.5</td></tr><tr><td>SPT-Dept † [58]</td><td>0.22</td><td>79.3</td><td>92.6</td><td>73.2</td><td>99.5</td><td>91.0</td><td>89.1</td><td>51.2</td><td>82.3</td><td>85.4</td><td>96.8</td><td>84.9</td><td>74.8</td><td>85.5</td><td>70.3</td><td>64.8</td><td>54.2</td><td>75.2</td><td>79.3</td><td>49.5</td><td>36.5</td><td>41.5</td><td>58.9</td><td>75.6</td></tr><tr><td>iVPT † [71]</td><td>0.60</td><td>82.7</td><td>94.2</td><td>72.0</td><td>99.1</td><td>91.8</td><td>88.1</td><td>56.6</td><td>83.5</td><td>87.7</td><td>96.1</td><td>87.1</td><td>77.6</td><td>87.1</td><td>77.1</td><td>62.6</td><td>49.4</td><td>80.6</td><td>82.1</td><td>55.3</td><td>31.8</td><td>47.6</td><td>60.8</td><td>77.1</td></tr><tr><td>PRO-VPT (ours)</td><td>0.61</td><td>84.5</td><td>94.1</td><td>73.2</td><td>99.4</td><td>91.8</td><td>88.2</td><td>57.2</td><td>84.1</td><td>87.7</td><td>96.8</td><td>86.6</td><td>75.5</td><td>86.7</td><td>78.8</td><td>61.0</td><td>50.6</td><td>81.3</td><td>86.7</td><td>56.4</td><td>38.1</td><td>51.7</td><td>63.1</td><td>78.0</td></tr></table>
+
+![](images/48083588549ffa71739fc7376c8b3ff95e05f0f7cbf7b0e0a42432cfd540e51f.jpg)  
+Figure 4. Performance gains achieved by VPT w/ ADO (PRO-VPT) compared to VPT w/o ADO (VPT-Deep). PRO-VPT consistently outperforms VPT-Deep.
+
+NOAH, our nested optimization-based PRO-VPT achieves an average accuracy gain of 2.5 pp, validating the effectiveness of our proposed nested optimization framework.
+
+Furthermore, Fig. 4 illustrates the performance differences between VPT w/ ADO (PRO-VPT) and w/o ADO (VPT-Deep) on the VTAB-1k datasets. The accuracies of PRO-VPT are consistently higher than those of VPT-Deep, yielding an average accuracy improvement of  $1.6\mathrm{pp}$ . This emphasizes the significance of optimizing the prompt distribution to unlock the full potential of VPT.
+
+FGVC. Next, we evaluate PRO-VPT on the FGVC benchmark. Among the aforementioned contenders, only a portion have conducted evaluations on FGVC and released the code and configurations.
+
+Tab. 2 presents the FGVC results with the best-performing normalization (see Appendix H for complete results). PRO-VPT achieves the best average accuracy of  $91.7\%$  over all five FGVC datasets, surpassing the second-best performance of  $91.4\%$  achieved by Adapter+ and SPT-Deep. This demonstrates that PRO-VPT also delivers state-of-the-art results for task adaptation when training data is large-scale.
+
+Table 2. Detailed results on the FGVC datasets. Performance results are reported as the highest of ImageNet normalization (○) or Inception normalization (●), presented in % after a complete training schedule with ViT-B/16 supervised pre-trained on ImageNet-21k. The best results of prompt-based methods and other PEFT approaches are highlighted in bold. †: Same as Tab. 1.  
+
+<table><tr><td></td><td>Param (M)</td><td>CUB200</td><td>NABirds</td><td>Oxford Flowers</td><td>Stanford Dogs</td><td>Stanford Cars</td><td>Global Avg.</td></tr><tr><td>Full •</td><td>86.0</td><td>88.0</td><td>81.5</td><td>99.2</td><td>85.6</td><td>90.6</td><td>89.0</td></tr><tr><td>Linear •</td><td>0.18</td><td>88.9</td><td>81.8</td><td>99.5</td><td>92.6</td><td>52.8</td><td>83.1</td></tr><tr><td>SSF o [34]</td><td>0.39</td><td>89.5</td><td>85.7</td><td>99.6</td><td>89.6</td><td>89.2</td><td>90.7</td></tr><tr><td>SPT-Adapter † [16]</td><td>0.40</td><td>89.1</td><td>83.3</td><td>99.2</td><td>91.1</td><td>86.2</td><td>89.8</td></tr><tr><td>SPT-LoRA † [16]</td><td>0.52</td><td>88.6</td><td>83.4</td><td>99.5</td><td>91.4</td><td>87.3</td><td>90.1</td></tr><tr><td>Adapter+ • [51]</td><td>0.34</td><td>90.4</td><td>85.0</td><td>99.7</td><td>92.6</td><td>89.1</td><td>91.4</td></tr><tr><td colspan="8">Prompt-based Methods:</td></tr><tr><td>VPT-Deep • [22]</td><td>0.85</td><td>90.1</td><td>83.3</td><td>99.6</td><td>90.3</td><td>85.0</td><td>89.7</td></tr><tr><td>SPT-Deep † [58]</td><td>0.36</td><td>90.6</td><td>87.6</td><td>99.8</td><td>89.8</td><td>89.2</td><td>91.4</td></tr><tr><td>iVPT † [71]</td><td>0.41</td><td>89.1</td><td>84.5</td><td>99.5</td><td>90.8</td><td>85.6</td><td>89.9</td></tr><tr><td>PRO-VPT (ours)</td><td>0.86</td><td>90.6</td><td>86.7</td><td>99.7</td><td>91.8</td><td>89.6</td><td>91.7</td></tr></table>
+
+# 4.3. Ablation Study
+
+To comprehensively assess the advantages of our proposed PR strategy, we compare it with the following alternative strategies, as shown in Tab. 3:
+
+Strategies (A), (B) represent alternatives for extracting a sub-distribution from a hyper-distribution (see § 3.4). We utilize the pruning-based technique as in [13], applying pruning percentages ranging from  $10\%$  to  $90\%$  with  $10~\mathrm{pp}$  intervals. (B) prunes from a huge hyper-distribution to obtain a sub-distribution with the same final parameter count as the VPT
+
+Table 3. Ablation study on various alternative ADO strategies. Results are presented using ViT-B/16 on two instances: Natural SVHN and Specialized Resisc45. The best and second-best results are highlighted in bold and underline. *Our proposed PR strategy.  
+
+<table><tr><td rowspan="2">Alternatives for the ADO Problem</td><td colspan="2">SVHN</td><td colspan="2">Resisc45</td></tr><tr><td>Param (M)</td><td>Acc (%)</td><td>Param (M)</td><td>Acc (%)</td></tr><tr><td>VPT Baseline</td><td>0.468</td><td>84.10</td><td>0.127</td><td>82.50</td></tr><tr><td>(A) Pruning</td><td>0.468</td><td>83.76</td><td>0.127</td><td>83.81</td></tr><tr><td>(B) Pruning</td><td>2.543</td><td>84.17</td><td>0.541</td><td>84.86</td></tr><tr><td>(C) Naive RL</td><td>0.482</td><td>83.51</td><td>0.140</td><td>83.02</td></tr><tr><td>(D) Prn1 &amp; Alc1</td><td>0.468</td><td>84.88</td><td>0.127</td><td>84.06</td></tr><tr><td>(E) Prn1 &amp; Alc2</td><td>0.482</td><td>86.99</td><td>0.140</td><td>85.68</td></tr><tr><td>(F) Prn2 &amp; Alc1</td><td>0.468</td><td>85.82</td><td>0.127</td><td>84.76</td></tr><tr><td>(G) Prn2 &amp; Alc2*</td><td>0.482</td><td>88.19</td><td>0.140</td><td>86.56</td></tr></table>
+
+baseline, and (A) prunes from a smaller hyper-distribution to match the number of trained parameters in the VPT baseline.
+
+Strategies (C)-(G) represent alternatives of prompt relocating, each incorporating different components. Particularly, (C) frames the PR process as a naive RL problem (see § 3.3) and utilizes the PPO algorithm. (D)-(G) are various combinations of the following components:
+
+Prn1. Randomly pruning prompts.
+
+Prn2. Pruning underutilized prompts based on the idleness score.
+
+Alc1. Modeling the allocation task as a multi-armed bandit problem, which is well-suited for stationary discrete decision-making, and then adopting the widely used Thompson sampling for prompt allocation.
+
+Alc2. Modeling the allocation task as a vanilla RL problem, which is more suited to non-stationary decision-making, and then using the popular PPO algorithm for prompt allocation.
+
+In summary, the key observations are as follows:
+
+Observation 1. Relocation-based methods are more efficient and effective compared to pruning-based methods. Pruning strategies (A) and (B) typically underperform relocation strategies (D)-(G), indicating that pruning is less effective in adaptive distribution. Specific to pruning, while (B) outperforms (A), it incurs high storage costs (e.g., 2.075M trained but unused parameters in SVHN) and still lags behind our proposed PR strategy. In contrast, RL-based relocation strategies (E) and (G) require only 0.014M extra parameters for PPO's policy networks, which is relatively negligible.
+
+Observation 2. The decomposition for  $RL$  is remarkably effective. Strategy (C) significantly underperforms the two-step PR strategies (D)-(G), underscoring the benefits of decomposing the PR process for simplifying decision-making.
+
+Observation 3. Both the pruning and allocation components in PRO-VPT exhibit significantly effective, especially RL-based allocation. Comparing strategies (D) vs. (F) and (E) vs. (G),  $Prn2$  indicates superior performance improvements (averaging 1.07 pp for SVHN and 0.79 pp for Resisc45), highlighting the effectiveness of the proposed idleness score. Moreover, strategies (G) and (E) achieve the best results, with Alc2 delivering remarkable gains over Alc1 (2.24 pp and 1.71 pp), highlighting the advantage of framing the allocation task as a non-stationary RL problem.
+
+Table 4. Generalizability study on distinct backbones. Results are presented in % on two instances: VTAB-1k Natural Cifar100 and Structured DMLab.  
+
+<table><tr><td>Backbone</td><td></td><td>Cifar100</td><td>DMLab</td></tr><tr><td rowspan="2">ViT-B/16</td><td>VPT-Deep</td><td>83.00</td><td>49.70</td></tr><tr><td>PRO-VPT</td><td>84.49 1.49 ↑</td><td>50.64 0.94 ↑</td></tr><tr><td rowspan="2">ViT-L/16</td><td>VPT-Deep</td><td>85.82</td><td>45.98</td></tr><tr><td>PRO-VPT</td><td>87.07 1.25 ↑</td><td>46.57 0.59 ↑</td></tr><tr><td rowspan="2">ViT-H/14</td><td>VPT-Deep</td><td>78.49</td><td>44.12</td></tr><tr><td>PRO-VPT</td><td>79.73 1.24 ↑</td><td>45.17 1.05 ↑</td></tr><tr><td rowspan="2">Swin-B</td><td>VPT-Deep</td><td>81.33</td><td>49.86</td></tr><tr><td>PRO-VPT</td><td>82.67 1.34 ↑</td><td>50.78 0.92 ↑</td></tr></table>
+
+Table 5. Generalizability study on different pre-training strategies. Results are presented in % on two instances: VTAB-1k Natural Cifar100 and Structured DMLab.  
+
+<table><tr><td colspan="2">Pre-Training Strategy</td><td>Cifar100</td><td>DMLab</td></tr><tr><td rowspan="2">Supervised</td><td>VPT-Deep</td><td>83.00</td><td>49.70</td></tr><tr><td>PRO-VPT</td><td>84.49 1.49 ↑</td><td>50.64 0.94 ↑</td></tr><tr><td colspan="4">Self-Supervised:</td></tr><tr><td rowspan="2">MAE</td><td>VPT-Deep</td><td>32.65</td><td>43.58</td></tr><tr><td>PRO-VPT</td><td>34.36 1.71 ↑</td><td>46.07 2.49 ↑</td></tr><tr><td rowspan="2">MoCo-v3</td><td>VPT-Deep</td><td>72.48</td><td>46.56</td></tr><tr><td>PRO-VPT</td><td>73.62 1.14 ↑</td><td>47.63 1.07 ↑</td></tr></table>
+
+# 4.4. Generalizability Analysis
+
+We evaluate the generalizability of PRO-VPT across different backbones and pre-training strategies, as well as in detection and segmentation tasks.
+
+**Backbone.** We experiment with three different scales of the ViT backbone: ViT-Base (ViT-B/16), ViT-Large (ViT-L/16), and ViT-Huge (ViT-H/14) [7], along with the hierarchical Swin Transformer (Swin-B) [39]. All backbones are supervised pre-trained on ImageNet-21k [46]. As shown in Tab. 4, PRO-VPT delivers higher accuracy compared to VPT with various scales and architectures.
+
+Pre-Training Strategy. In addition to the supervised pretraining strategy, we experiment with two self-supervised strategies: the masked image modeling method (MAE) [18] and the contrastive self-supervised method (MoCo-v3) [3]. As illustrated in Tab. 5, PRO-VPT consistently outperforms VPT regardless of the pre-training strategy, highlighting its strong generalizability.
+
+Detection and Segmentation. We also experiment on detection and segmentation tasks in Appendix H, where PRO-VPT still maintains advantages over VPT.
+
+# 4.5. Observations on PRO-VPT
+
+We design several experiments to thoroughly understand PRO-VPT. The key insights are concluded as follows:
+
+Observation 1. PRO-VPT effectively learns task-specific prompt distributions. We visualize the learned distributions of PRO-VPT under different total numbers of prompts, as shown in Fig. 5. Notably, the distributions remain consistent
+
+![](images/5210786e215a241ce9998933cff03dc5a05d0c511ae83ec4110d1be70a6431c8.jpg)  
+Figure 5. Visualization of prompt distributions learned by PRO-VPT on VTAB-1k Natural CIFar100, Specialized Resisc45, and Structured Clevr-Count with varying numbers of prompts. PRO-VPT effectively learns task-specific distributions.
+
+![](images/30bccd057695e84caebe13024780a8c5b855acb7b95984d5d40994a0a88538b0.jpg)
+
+![](images/955516fcd9c5d540bddb9a28ae547496d8766a0ee8f9de080880e16c0d8f9958.jpg)
+
+![](images/a4fc96948910094b2bd06bd2c3a423b90cddc75a5797876e39f93a50d7235e70.jpg)  
+Figure 7. Cosine similarity of feature changes and the corresponding test accuracy. The PR strategy has a minimal impact on features, and PRO-VPT demonstrates better stability than VPT in both representation and performance.
+
+![](images/4ba24fb2c9b6780513312ea743b264265ff8ea61fb98e7ce4551706b02147fc2.jpg)
+
+![](images/9b718002cedb2c59eb33918d97dc728e1b0724a30c6cebf24258af4d44d208fa.jpg)
+
+![](images/f4f4a7b6b2bd491c54185d3f5f5a068a32f4b07dc74dc39847242dbe53316be5.jpg)
+
+across different prompt numbers but exhibit significant variations across different tasks. This indicates that the prompting importance of each block indeed varies by task, and PROVPT effectively captures and adapts to these task-specific distributions to better accommodate each task.
+
+Observation 2. PRO-VPT effectively meets the block-specific prompt requirements via prompt relocation. We visualize the attention weights from representation features (i.e., [CLS] embeddings) to prompts across different blocks on Cifar100. Each attention weight represents the importance of a given prompt for its corresponding block. As depicted in Fig. 6, the attention weights in VPT reveal phenomena of prompt saturation (e.g., blocks 2 and 5) and prompt deficiency (e.g., blocks 4 and 12), confirming that different blocks exhibit significantly varying prompt demands. However, due to its fixed prompt distribution, VPT lacks the flexibility to adapt to these diverse needs. Even pruning-based methods can only prune negative prompts from prompt-
+
+saturated blocks while failing to address prompt-deficient blocks. In contrast, PRO-VPT effectively meets these block-specific prompt demands through adaptive prompt relocation. Such a fine-grained tuning strategy maximizes the effectiveness of the prompt distribution, thereby enhancing downstream task performance.
+
+![](images/045489051476a645121b46227c8a7c99de60ed81c57f7d2eff71c88f2519a626.jpg)  
+Figure 6. Visualization of prompt attention weights.
+
+Observation 3. PRO-VPT demonstrates enhanced stability and robustness. One might be skeptical that runtime modifi
+
+cations of the prompt distribution could compromise stability. To investigate this, we evaluate both the representation stability and performance stability in PRO-VPT and VPT. Specific to feature stability, we measure cosine similarity between representations (i.e., [CLS] embeddings) from the previous epoch and after each stage of pruning, allocation, and tuning. This similarity reflects the extent of feature changes resulting from each stage. Fig. 7 illustrates the similarity at different blocks and the corresponding test accuracy on Cifar100. Surprisingly, the orange and green lines indicate that the PR strategy has little impact on the learned features, as the features remain nearly unchanged after pruning and allocation. Moreover, PRO-VPT exhibits superior stability compared to VPT, both in terms of representation stability and performance stability. Additionally, as shown in Fig. vi, PRO-VPT demonstrates remarkable robustness to variations in prompt quantity compared to VPT, whereas the performance of VPT fluctuates significantly while PRO-VPT remains consistent. Overall, we argue that these benefits stem from our proposed optimization-based approach, which effectively identifies the optimal road for relocation, thereby minimizing its impact on stability and enhancing robustness.
+
+# 5. Conclusion
+
+This paper investigates the ADO problem within VPT. Based on our experimental insights into the nested relationship between ADO and VPT, we formalize an ADO-VPT co-design framework through the lens of nested optimization. Upon this formalization, we propose a novel PRO-VPT algorithm, which iteratively optimizes the distribution by integrating the PR (prompt relocation) strategy with VPT. Across 26 datasets, we show our method's significant performance improvement over VPT with superior robustness.
+
+# Acknowledgements
+
+This work was supported in part by the National Natural Science Foundation of China under Grants 62376233, 62306181, 62476063, and 62306313; in part by the NSFC / Research Grants Council (RGC) Joint Research Scheme under Grant N_HKBU214/21; in part by the General Research Fund of RGC under the Grant 12202622; in part by the RGC Senior Research Fellow Scheme under Grant SRFS2324-2S02; in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515010163; in part by the National Key Laboratory of Radar Signal Processing under Grant JKW202403; and in part by Xiaomi Young Talents Program.
+
+# References
+
+[1] Aochuan Chen, Yuguang Yao, Pin-Yu Chen, Yihua Zhang, and Sijia Liu. Understanding and improving visual prompting: A label-mapping perspective. In CVPR, pages 19133-19143, 2023. 1  
+[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, pages 16664-16678, 2022. 1  
+[3] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In CVPR, pages 9640-9649, 2021. 7  
+[4] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. vii  
+[5] Rajshekhar Das, Yonatan Dukler, Avinash Ravichandran, and Ashwin Swaminathan. Learning expressive prompting with residuals for vision transformers. In CVPR, pages 3366-3377, 2023. 2  
+[6] E Dataset. Novel datasets for fine-grained image categorization. In CVPR Workshop on Fine-grained Visual Classification, 2011. 5, v  
+[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 5, 7, i  
+[8] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 5  
+[9] Minghao Fu, Ke Zhu, and Jianxin Wu. Dtl: Disentangled transfer learning for visual recognition. In AAAI, pages 12082-12090, 2024. 5  
+[10] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In AAAI, 2017. 5, v  
+[11] Henry Gouk, Timothy Hospedales, and massimiliano pontil. Distance-based regularisation of deep networks for finetuning. In ICLR, 2021. 2
+
+[12] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In CVPR, pages 4805-4814, 2019. 2  
+[13] Cheng Han, Qifan Wang, Yiming Cui, Zhiwen Cao, Wenguan Wang, Siyuan Qi, and Dongfang Liu. E^2vpt: An effective and efficient approach for visual prompt tuning. In ICCV, pages 17445-17456, 2023. 2, 5, 6, iv  
+[14] Cheng Han, Qifan Wang, Yiming Cui, Wenguan Wang, Lifu Huang, Siyuan Qi, and Dongfang Liu. Facing the elephant in the room: Visual prompt tuning or full finetuning? In ICLR, 2024. 1  
+[15] Tianxiang Hao, Hui Chen, Yuchen Guo, and Guiguang Ding. Consolidator: Mergable adapter with group connections for visual adaptation. In ICLR, 2023. 5, 6, vi  
+[16] Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Sensitivity-aware visual parameter-efficient fin-tuning. In ICCV, pages 11825-11835, 2023. 5, 6, vi, vii  
+[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. v  
+[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 7  
+[19] Shihao Hou, Xinyi Shang, Shreyank N Gowda, Yang Lu, Chao Wu, Yan Yan, and Hanzi Wang. Capt: Class-aware prompt tuning for federated long-tailed learning with vision-language model. arXiv preprint arXiv:2503.06993, 2025. 2  
+[20] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 5, 6, vi  
+[21] Lingyun Huang, Jianxu Mao, Yaonan Wang, Junfei Yi, and Ziming Tao. Cvt: Cross-attention help visual prompt tuning adapt visual task. arXiv preprint arXiv:2408.14961, 2024. vi, vii  
+[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727, 2022. 1, 2, 5, 6, i, iv, v, vi, vii, viii  
+[23] Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. In AAAI, pages 1060-1068, 2023. 5, 6, vi  
+[24] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of Artificial Intelligence Research, 4:237-285, 1996. 3  
+[25] Gal Kaplun, Andrey Gurevich, Tal Swisa, Mazor David, Shai Shalev-Shwartz, and Eran Malach. Less is more: Selective layer finetuning with subtuning. arXiv preprint arXiv:2302.06354, 2023. 2  
+[26] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. CSUR, 54(10s):1-41, 2022. 1, i  
+[27] Youngeun Kim, Yuhang Li, Abhishek Moitra, Ruokai Yin, and Priyadarshini Panda. Do we really need a large number of visual prompts? Neural Networks, 177:106390, 2024. 5
+
+[28] Yunyong Ko, Dongwon Lee, and Sang-Wook Kim. Not all layers are equal: A layer-wise adaptive approach toward large-scale dnn training. In WWW, pages 1851-1859, 2022. 2  
+[29] Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical finetuning improves adaptation to distribution shifts. In ICLR, 2023. 2  
+[30] Mengke Li, Ye Liu, Yang Lu, Yiqun Zhang, Yiu-ming Cheung, and Hui Huang. Improving visual prompt tuning by gaussian neighborhood minimization for long-tailed visual recognition. In NeurIPS, pages 103985-104009, 2024. 1  
+[31] Xingjian Li, Haoyi Xiong, Hanchao Wang, Yuxuan Rao, Liping Liu, Zeyu Chen, and Jun Huan. Delta: Deep learning transfer using feature map with attention for convolutional networks. In ICLR, 2019. 2  
+[32] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In ACL, pages 4582-4597, 2021. 2  
+[33] Yuxi Li. Deep reinforcement learning: An overview. arXiv preprint arXiv:1701.07274, 2017. 3  
+[34] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, pages 109–123, 2022. 5, 6, vi, vii  
+[35] Siyu Lin and Peter A Beling. An end-to-end optimal trade execution framework based on proximal policy optimization. In IJCAI, pages 4548-4554, 2021. i  
+[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. v  
+[37] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. CSUR, 55(9):1-35, 2023. 2  
+[38] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In ACL, pages 61-68, 2022. 2  
+[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 7  
+[40] Fang Ma, Chen Zhang, Lei Ren, Jingang Wang, Qifan Wang, Wei Wu, Xiaojun Quan, and Dawei Song. XPrompt: Exploring the extreme of prompt tuning. In EMNLP, pages 11033-11047, 2022. 1  
+[41] Eran Malach, Gilad Yehudai, Shai Shalev-Schwartz, and Ohad Shamir. Proving the lottery ticket hypothesis: Pruning is all you need. In ICML, pages 6682-6691, 2020. 5  
+[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, pages 722-729, 2008. 5, v  
+[43] Junyoung Park, Jin Kim, Hyeongjun Kwon, Ilhoon Yoon, and Kwanghoon Sohn. Layer-wise auto-weighting for nonstationary test-time adaptation. In WACV, pages 1414-1423, 2024. 2
+
+[44] Alex Renda, Jonathan Frankle, and Michael Carbin. Comparing rewinding and fine-tuning in neural network pruning. arXiv preprint arXiv:2003.02389, 2020. 5  
+[45] Youngmin Ro and Jin Young Choi. Autolr: Layer-wise pruning and auto-tuning of learning rates in fine-tuning of deep networks. In AAAI, pages 2486-2494, 2021. 2  
+[46] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 5, 7  
+[47] Pedro Savarese, Hugo Silva, and Michael Maire. Winning the lottery with continuous sparsification. In NeurIPS, pages 11380-11390, 2020. 5  
+[48] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.5.i  
+[49] Zhiqiang Shen, Zechun Liu, Jie Qin, Marios Savvides, and Kwang-Ting Cheng. Partial is better than all: Revisiting fine-tuning strategy for few-shot learning. In AAAI, pages 9594–9602, 2021. 2  
+[50] Zhengxiang Shi and Aldo Lipani. Dept: Decomposed prompt tuning for parameter-efficient fine-tuning. In ICLR, 2024. 2  
+[51] Jan-Martin O Steitz and Stefan Roth. Adapters strike back. In CVPR, pages 23449-23459, 2024. 5, 6, iv, v, vi, vii  
+[52] Pengwei Tang, Xiaolin Hu, and Yong Liu. Adept: Adaptive decomposed prompt tuning for parameter-efficient fine-tuning. arXiv preprint arXiv:2501.03291, 2025. 2  
+[53] Junjiao Tian, Zecheng He, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, and Zsolt Kira. Trainable projected gradient method for robust fine-tuning. In CVPR, pages 7836-7845, 2023. 2  
+[54] Junjiao Tian, Yen-Cheng Liu, James S Smith, and Zsolt Kira. Fast trainable projection for robust fine-tuning. In NeurIPS, 2024. 2  
+[55] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In CVPR, pages 595-604, 2015. 5, v  
+[56] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. Technical report, 2011. 5, v  
+[57] Shaowen Wang, Linxi Yu, and Jian Li. Lora-ga: Low-rank adaptation with gradient approximation. In NeurIPS, pages 54905-54931, 2025. 1  
+[58] Yuzhu Wang, Lechao Cheng, Chaowei Fang, Dingwen Zhang, Manni Duan, and Meng Wang. Revisiting the power of prompt for visual tuning. In ICML, pages 50233-50247, 2024. 2, 5, 6, vi, vii  
+[59] Yuzhu Wang, Manni Duan, and Shu Kong. Attention to burstiness: Low-rank bilinear prompt tuning. arXiv preprint arXiv:2506.22908, 2025. iv  
+[60] Raphael Ngigi Wanjiku, Lawrence Nderu, and Michael Kimwele. Dynamic fine-tuning layer selection using kullback-Leibler divergence. Engineering Reports, 5(5):e12595, 2023. 2
+
+[61] Yi Xin, Siqi Luo, Haodi Zhou, Junlong Du, Xiaohong Liu, Yue Fan, Qing Li, and Yuntao Du. Parameter-efficient finetuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242, 2024. 1  
+[62] Yi Xin, Siqi Luo, Xuyang Liu, Haodi Zhou, Xinyu Cheng, Christina E Lee, Junlong Du, Haozhe Wang, MingCai Chen, Ting Liu, et al. V-petl bench: A unified visual parameter-efficient transfer learning benchmark. In NeurIPS, pages 80522-80535, 2025. 1  
+[63] Chen Xu, Yuhan Zhu, Haocheng Shen, Boheng Chen, Yixuan Liao, Xiaoxin Chen, and Limin Wang. Progressive visual prompt learning with contrastive feature re-formation. IJCV, pages 1–16, 2024. 2  
+[64] Jiacheng Yang, Wai Keung Wong, Lunke Fei, Shuping Zhao, Jie Wen, and Shaohua Teng. Decoupling visual and identity features for adversarial palm-vein image attack. Neural Networks, 180:106693, 2024. i  
+[65] Bruce XB Yu, Jianlong Chang, Haixin Wang, Lingbo Liu, Shijie Wang, Zhiyu Wang, Junfan Lin, Lingxi Xie, Haojie Li, Zhouchen Lin, et al. Visual tuning. CSUR, 56(12):1-38, 2024. 1  
+[66] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 5  
+[67] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. IEEE TPAMI, 2024. 5, 6, vi  
+[68] Henry Hengyuan Zhao, Pichao Wang, Yuyang Zhao, Hao Luo, Fan Wang, and Mike Zheng Shou. Sct: A simple baseline for parameter-efficient fine-tuning via salient channels. IJCV, 132(3):731-749, 2024. 5  
+[69] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. vi  
+[70] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633–641, 2017. vi  
+[71] Nan Zhou, Jiaxin Chen, and Di Huang. ivpt: Improving task-relevant information sharing in visual prompt tuning by cross-layer dynamic connection. arXiv preprint arXiv:2404.05207, 2024. 2, 5, 6, vi, vii

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5781fcfd44aa706a373da1e882339c6177e476ee9691d46d749b1482aeb9d4ee
+size 629345

ICCV/2025/PRO-VPT_ Distribution-Adaptive Visual Prompt Tuning via Prompt Relocation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:036ca93ac2f22358bc5c63d13f28753fa1ac48a3217383bb5e73ff68006950dd
+size 435325

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/91086143-b5ba-4d82-87ab-d2e978da69f3_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fe4c83d87268ac40a9d98ebef43d8bf11b8fc0c865caf5f18280dc18297eded
+size 75492

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/91086143-b5ba-4d82-87ab-d2e978da69f3_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81586b72b5f648863e73e4c2d9a9394404732f1ef22f9734b69a2ff48893bebf
+size 97909

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/91086143-b5ba-4d82-87ab-d2e978da69f3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2820b0c5decf248774ef37a9b373e37d9ae41799c148596e653e4c405b4b8d0b
+size 11321066

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/full.md

@@ -0,0 +1,339 @@
+# PROGRESSOR: A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement
+
+Tewodros W. Ayalew $^{1}$  Xiao Zhang $^{*,1}$  Kevin Yuanbo Wu $^{*,1}$  Tianchong Jiang $^{2}$  Michael Maire $^{1}$  Matthew R. Walter $^{2}$
+
+<sup>1</sup>University of Chicago <sup>2</sup>Toyota Technological Institute at Chicago
+
+https://ripl.github.io/progressor
+
+# Abstract
+
+We present PROGRESSOR, a novel framework that learns a task-agnostic reward function from videos, enabling policy training through goal-conditioned reinforcement learning (RL) without manual supervision. Underlying this reward is an estimate of the distribution over task progress as a function of the current, initial, and goal observations that is learned in a self-supervised fashion. Crucially, PROGRESSOR refines rewards adversarially during online RL training by pushing back predictions for out-of-distribution observations in order to mitigate distribution shift inherent in non-expert observations. Utilizing this progress prediction as a dense reward together with an adversarial pushback, we show that PROGRESSOR enables robots to learn complex behaviors without any external supervision. Pre-trained on large-scale egocentric human video from EPIC-KITCHENS, PROGRESSOR requires no fine-tuning on indomain task-specific data for generalization to real-robot offline RL under noisy demonstrations, outperforming contemporary methods that provide dense visual reward for robotic learning. Our findings highlight the potential of PROGRESSOR for scalable robotic applications where direct action labels and task-specific rewards are not readily available.
+
+# 1. Introduction
+
+Practical applications of reinforcement learning (RL) require that a domain expert design complex reward models that encourage desired behavior while simultaneously penalizing unwanted behavior [34, 37, 38]. However, manually constructing dense rewards that enable effective learning is difficult and can yield undesired behaviors [5, 14]. Sparse rewards require far less effort, but the sparsity of supervision typically results in a significant loss of sample efficiency. These issues are exacerbated for long-horizon
+
+![](images/360f7ae4000bdeffb294d7a27089711b4d4bf33fa79cae23c26f3916072bd43d.jpg)  
+Figure 1. Trained in a self-supervised manner on expert videos, PROGRESSOR predicts an agent's progress toward task completion, providing a reward signal for reinforcement learning. During online reinforcement learning, we employ an adversarial technique to refine this reward estimate, addressing the distribution shift between expert data and non-expert online rollouts.
+
+tasks, where credit assignment is particularly difficult.
+
+A promising alternative is to learn effective rewards from unlabeled (i.e., action-free) videos of task demonstrations [2, 8, 19, 25, 29, 41]. Large-scale video data is readily available on the internet, providing a effective alternative to costly, intensive task-specific data collection. An effective way to incorporate diverse video data into a single model without learning a new policy for each task is to condition the model on a goal image that visually specifies the desired environment changes upon task completion [9, 25].
+
+To that end, we propose PROGRESSOR, a self-supervised framework that learns a task-agnostic reward function from videos for goal-conditioned reinforcement learning that only requires image observations, without the need for corresponding action labels. Following Yang et al. [41], we assume that expert demonstrations make monotonic progress towards the goal and employ a self-supervised temporal learning approach that utilizes a proxy objective aimed at
+
+estimating task completion progress. PROGRESSOR predicts the distribution of progress of the current observation (i.e., the current image) relative to the initial and goal observations. This estimation provides dense supervision of an agent's progression toward reaching the goal—and thereby completing the task—guiding exploration in alignment with expert execution trajectories. However, during RL exploration, agents will often encounter states that were not visited as part of the expert trajectories. PROGRESSOR accounts for this distribution shift via an online adversarial refinement that pushes back predictions for out-of-distribution observations.
+
+We evaluate the effectiveness of PROGRESSOR on robot manipulation tasks in both simulation and the real world. These results demonstrate the benefits that PROGRESSOR's self-supervised reward has on policy learning, particularly in settings where existing methods otherwise require intricate reward design. Additionally, PROGRESSOR can be used in imitation learning with suboptimal demonstrations to guide the learning of optimal task execution. The key contributions of this paper are as follows:
+
+- We present PROGRESSOR, a self-supervised reward model trained on unlabeled videos that guides a reinforcement learning agent by providing dense rewards based on predicted progress toward a goal.  
+- PROGRESSOR achieves state-of-the-art performance on six diverse tasks from the Meta-World benchmark [44] without the need for environment reward.  
+- Pretrained on large-scale egocentric human videos from EPIC-KITCHENS [6], PROGRESSOR enables few-shot offline learning on real robots, even in scenarios where half of the demonstrations are unsuccessful.
+
+# 2. Related Work
+
+Learning from expert videos has gained traction due to the potential to scale the acquisition of complex skills using the vast amount of video content readily available online.
+
+Inverse RL Inverse reinforcement learning (IRL) is a framework for learning reward functions from expert demonstrations as an alternative to specifying them manually. In contrast to traditional RL, where the reward function is predefined and the goal is to learn an optimal policy, IRL focuses on deducing the underlying reward structure that results in the expert's behavior [1, 40, 47]. Recent IRL methods leverage adversarial techniques to learn reward functions [10, 11, 23], such as in Generative Adversarial Imitation Learning (GAIL) [18]. A key limitation of typical approaches to IRL is that they require action-labeled demonstrations, which are only available in small quantities. Another is the ambiguity that results from having multiple rewards that can explain the observed behavior [27], potentially leading to ill-shaped reward functions. In contrast, our notion of progress enables the learning of well
+
+shaped dense rewards.
+
+Imitation learning from video Several methods [4, 45] infer an inverse dynamics model from a small amount of action-annotated videos and then use this model to pseudolabel large-scale action-free videos for policy learning. Other methods employ hand pose estimation as an intermediate annotation followed by standard imitation learning [3, 32], but they are sensitive to the accuracy of the pose estimators and rely on camera calibration.
+
+Learning reward from video Recent methods learn visual representations useful for RL from video data using self-supervised temporal contrastive techniques [9, 25, 29, 36, 42]. Time-contrastive objectives treat observations close in time as positive pairs and those that are temporally distant as negative pairs. They then learn embeddings that encourage small distances between positive pairs and large distances between negative pairs. The learned embeddings can then be used to define reward functions, i.e., as the distance from the goal image in embedding space. Time-contrastive objectives can be sensitive to frame rate, and extracted rewards are assumed to be symmetric [25]. In comparison, our approach is agnostic to the sampling rate and can learn asymmetric dynamics depending on the specified initial and goal images.
+
+Additionally, generative approaches have been used for reward learning. VIPER [8] estimates the log-likelihood of an observation and uses it as a reward function by training on an expert dataset using a VideoGPT-based autoregressive model. Similarly, Huang et al. [19] employ a diffusion-based approach to learn rewards by leveraging conditional entropy. However, these methods necessitate costly generative processes to estimate rewards.
+
+Particularly relevant, Rank2Reward [41] demonstrates that learning the ranking of visual observations from a demonstration helps to infer well-shaped reward functions. By integrating ranking techniques with the classification of expert demonstration data vs. non-expert data obtained through on-policy data collection, Rank2Reward guides robot learning from demonstration. A key limitation of the method is that it necessitates training distinct models for each task due to the task-dependent nature of ranking frames. In contrast, our approach employs a single reward model for all tasks. Additionally, we utilize an adversarial training approach that tackles the domain shift observed during online RL, rather than incorporating a separate classifier as used in Rank2Reward.
+
+# 3. Preliminaries
+
+We formulate our method in the context of a finite-horizon Markov decision process (MDP)  $\mathcal{M}$  defined by the tuple  $(\mathcal{O},\mathcal{A},\mathcal{P},\mathcal{R},\gamma ,\rho_0)$  that represents the observation space, action space, transition probability function, reward function, discount factor, and initial state distribution, respec
+
+tively. The agent starts from an initial state with observation  $o_0 \in \mathcal{O}$  and at each step performs an action  $a \in \mathcal{A}$ .
+
+The fundamental objective in RL is to learn an optimal policy  $\pi^{*}$  that maximizes the expected discounted reward, i.e.,  $\pi^{*} = \arg \max_{\pi}\mathbb{E}_{\pi_{\theta}}[\sum_{t = 1}^{T}(\gamma^{t - 1}\mathcal{R}(\boldsymbol {o}_t))]$ .
+
+In this work, we assume a fully observable environment, where each observation  $\pmb{o}_k \in \mathbb{R}^{H \times W \times 3}$  corresponds to a high-dimensional RGB image. We assume access to a dataset  $\mathcal{D}_e = \{\tau_k\}_{k=1}^N$  of expert demonstrations, where each demonstration  $\tau_k$  is a trajectory (sequence) of observations  $\tau_k = \{o_0^{\tau_k}, o_1^{\tau_k}, \dots, o_n^{\tau_k}\}$ .
+
+However, we assume that neither the reward function  $\mathcal{R}$  nor the demonstrated actions  $a_{t} \sim \mathcal{A}$  are known. This setup mirrors the real-world robot task-learning scenario in which we have a plethora of videos of human-performed tasks, without knowledge of the human actions or the rewards being received. Indeed, the primary objective of this paper is to construct a model capable of predicting a well-structured dense reward from a dataset of expert demonstrations, without access to demonstrated actions or rewards. This model can subsequently serve as a reward estimating mechanism when training an agent using online RL in the absence of any environment rewards.
+
+# 4. Method
+
+We propose to learn a unified reward model via an encoder  $\pmb{E}_{\theta}(\pmb{o}_i, \pmb{o}_j, \pmb{o}_g)$  that estimates the relative progress of an observation  $\pmb{o}_j$  with respect to an initial observation  $\pmb{o}_i$  and a goal observation  $\pmb{o}_g$ , all of which are purely pixel-based. Figure 2 illustrates PROGRESSOR's framework for task progress estimation. Here, the objective is to learn the temporal ordering of observations that lead to the correct execution of a task. By learning this progress estimation, we can create a reward model that incentivizes progress toward the goal. Similar to Yang et al. [41], our method relies on the assumption that the values of states in optimal policies increase monotonically towards task completion.
+
+# 4.1. Learning the Self-Supervised Reward Model
+
+We measure an observation  $o_j$ 's progress towards a goal in terms of its relative position with respect to the initial observation  $o_i$  and the goal observation  $o_g$ . In order to learn this progress estimate, we first represent the relative position of an observation  $o_j$  within an expert trajectory  $\tau_k$  as
+
+$$
+\delta \left(\boldsymbol {o} _ {i} ^ {\tau_ {k}}, \boldsymbol {o} _ {j} ^ {\tau_ {k}}, \boldsymbol {o} _ {g} ^ {\tau_ {k}}\right) = \frac {| j - i |}{| g - i |}. \tag {1}
+$$
+
+Here,  $\delta (\pmb{o}_i^{\tau_k},\pmb{o}_j^{\tau_k},\pmb{o}_g^{\tau_k})$  calculates the ratio of the relative frame position differences between  $\pmb{o}_j$  and  $\pmb{o}_i$  and between  $\pmb{o}_g$  and  $\pmb{o}_i$  within a given expert trajectory  $\tau_{k}$ . This progress label is always such that  $\delta (\pmb{o}_i^{\tau_k},\pmb{o}_j^{\tau_k},\pmb{o}_g^{\tau_k})\in [0,1]$
+
+Predicting progress based on real-world video is challenging due to factors like camera motion and repetitive sub-trajectories. We address this by formulating the problem as one of estimating a Gaussian distribution over progress, allowing the model to represent the uncertainty in the estimate. With this approach, we approximate the ground-truth distribution for a triplet of frames  $(\pmb{o}_i^{\tau_k},\pmb{o}_j^{\tau_k},\pmb{o}_g^{\tau_k})$  within a video  $\tau_{k}$  as a normal distribution with mean  $\mu_{\tau_k} = \delta (\pmb{o}_i^{\tau_k},\pmb{o}_j^{\tau_k},\pmb{o}_g^{\tau_k})$  and standard deviation  $\sigma_{\tau_k} = \max \left(\frac{1}{g - i},\epsilon\right)$ :
+
+$$
+p _ {\text {t a r g e t}} \left(\boldsymbol {o} _ {i} ^ {\tau_ {k}}, \boldsymbol {o} _ {j} ^ {\tau_ {k}}, \boldsymbol {o} _ {g} ^ {\tau_ {k}}\right) = \mathcal {N} \left(\mu_ {\tau_ {k}}, \sigma_ {\tau_ {k}} ^ {2}\right) \tag {2}
+$$
+
+Our  $\epsilon$  upper-bound of  $\sigma_{\tau_k}$  downweights the penalty for triplets in which there are many frames between the initial and goal images, which we empirically find to improve robustness during training.
+
+We optimize our reward model  $\boldsymbol{r}_{\theta}$  to predict the distribution of progress on the expert trajectories. The architecture consists of a shared visual encoder to compute the per-frame representation, followed by several MLPs to produce the final estimate:
+
+$$
+\boldsymbol {E} _ {\theta} \left(\boldsymbol {o} _ {i}, \boldsymbol {o} _ {j}, \boldsymbol {o} _ {g}\right) = \mathcal {N} \left(\mu , \sigma^ {2}\right). \tag {3}
+$$
+
+Following Kingma and Welling [22], we estimate  $\log \sigma^2$  and then compute  $\sigma$  as  $\sigma = \exp \left(0.5\log \sigma^2\right)$ . We learn the parameters of the network by optimizing the Kullback-Leibler divergence loss between the ground-truth and the predicted distributions:
+
+$$
+\mathcal {L} _ {\text {e x p e r t}} = D _ {\mathrm {K L}} \left(p _ {\text {t a r g e t}} \left(\boldsymbol {o} _ {i} ^ {\tau_ {k}}, \boldsymbol {o} _ {j} ^ {\tau_ {k}}, \boldsymbol {o} _ {g} ^ {\tau_ {k}}\right) \| \boldsymbol {E} _ {\theta} \left(\boldsymbol {o} _ {i} ^ {\tau_ {k}}, \boldsymbol {o} _ {j} ^ {\tau_ {k}}, \boldsymbol {o} _ {g} ^ {\tau_ {k}}\right)\right) \tag {4}
+$$
+
+During training of our reward model  $\boldsymbol{r}_{\theta}$ , we randomly select and rank three frames from a video sequence of expert trajectories as a training triplet. Additionally, since these triplets can be randomly drawn from any trajectory within the dataset of different robotics tasks, a single reward model suffices to handle a variety of tasks.
+
+Using the Reward Model in Online RL: Due to the monotonic nature of task progress in our problem formulation, the estimates derived from the trained model can serve as dense rewards for training, replacing the unknown true reward. We create the reward model by defining a function in terms of the model's predicted outputs given a sample triplet of frames  $(\pmb{o}_i,\pmb{o}_j,\pmb{o}_g)$  from a trajectory as:
+
+$$
+\boldsymbol {r} _ {\theta} \left(\boldsymbol {o} _ {i}, \boldsymbol {o} _ {j}, \boldsymbol {o} _ {g}\right) = \mu - \alpha \mathcal {H} \left(\mathcal {N} \left(\mu , \sigma^ {2}\right)\right), \tag {5}
+$$
+
+where  $r_{\theta}$  denotes PROGRESSOR's reward estimate and  $\mathcal{H}$  is the entropy. The first term  $\mu$  in Eqn. 5 uses the mean prediction from the progress estimation model to determine the position of the current observation relative to the goal. The second term penalizes observations that yield high variance. Here,  $\alpha$  is hyperparameter, which we set to  $\alpha = 0.4$  in all our experiments.
+
+![](images/1a00e33c8480131565db3712afd01f92efd1b666a2c713fe070d85c5a27abb66.jpg)
+
+![](images/5d535dcc9dce3eb5bf22ef49f51cf1a1573a478a26c1c900fa3debea4b5977fd.jpg)
+
+![](images/176b73f331fa235a4416366a671273fb19a3956efae432e3fc9b5b59139f951c.jpg)  
+Figure 2. Top Left: Initial phase of reward model pretraining on expert data, where the model learns to predict the parameters of a Gaussian distribution centered on normalized progress, reflecting expected progress as demonstrated by experts. Top Right: In online reinforcement learning (RL) training, an adversarial online refinement (i.e., push-back) is applied to counteract non-expert predictions made by the reward model, effectively distinguishing expert from non-expert progress. Bottom: During online RL, the reward model is updated on expert and non-expert data.
+
+# 4.2. Adversarial Online Refinement via Push-Back
+
+Our reward model  $r_{\theta}$  can be used to train a policy that maximizes the expected sum of discounted rewards as:
+
+$$
+\pi^ {*} = \arg \max  _ {\pi} \mathbb {E} _ {\pi_ {\theta}} \left[ \sum_ {t = 1} ^ {T} \gamma^ {t - 1} \boldsymbol {r} _ {\theta} \left(\boldsymbol {o} _ {i} ^ {\tau^ {\prime}}, \boldsymbol {o} _ {t} ^ {\tau^ {\prime}}, \boldsymbol {o} _ {g} ^ {\tau^ {\prime}}\right) \right] \tag {6}
+$$
+
+The reward encourages the agent to take actions that align its observed trajectories  $\tau'$  with those of the expert demonstrations  $\tau$ , thereby promoting task completion. However, the all-but-random actions typical of learned policies early in training result in out-of-distribution observations relative to the expert trajectories. As a result, optimizing the policy using a reward  $r_{\theta}$  that is only trained on an expert trajectory  $\tau$  can result in unintended behaviors, since  $r_{\theta}$  does not reflect the actual progress of  $\tau'$ . To tackle this distribution shift, we propose an adversarial online refinement strategy [12, 13], which we refer to as "push-back", that enables the reward model  $r_{\theta}$  to differentiate between in- and out-of-distribution observations from  $\tau$  and  $\tau'$ , improving estimation accuracy. Specifically, for a frame triplet  $(o_i^{\tau_k}, o_j^{\tau_k}, o_g^{\tau_k})$  sampled from  $\tau_k'$  and the estimated progress  $\mu_{\tau_k'}$  from  $E_{\theta}(3)$ , we update  $E_{\theta}$  so that it learns to push-back
+
+# Algorithm 1 Training PROGRESSOR with Online RL
+
+Require: Expert demonstration data  $\mathcal{D}_e = \{\tau_k\}_{k=1}^N$  
+1: Initialize policy  $\pi$ , empty replay buffer  $\mathcal{D}_{RB}$  
+2: Initialize reward model  $\mathbf{r}_{\theta}$  
+3: // Pretraining PROGRESSOR  
+4: for step  $n$  in  $\{1, \dots, N_{\text{pretrain}}\}$  do  
+5: Sample triplets  $(\pmb{o}_i^{\tau_k},\pmb{o}_j^{\tau_k},\pmb{o}_g^{\tau_k})$  from  $\tau_{k}\in \mathcal{D}_{e}$  
+6: Learn  $\boldsymbol{r}_{\theta}$  with objective Eqn. 4  
+7: end for  
+8: // Policy optimization and reward model fine-tuning  
+9: for step  $n$  in  $\{1, \dots, N\}$  do  
+10: With  $\pi$ , collect transitions  $\{\tau_l\}_{l=1}^M$  and store in  $\mathcal{D}_{RB}$  
+11: if  $n \%$  reward_update频率  $= 0$  then  
+12: Update  $\boldsymbol{r}_{\theta}$  with Eqn. 4 using samples from  $\mathcal{D}_e$  
+13: Update  $\pmb{r}_{\theta}$  with Eqn. 7 using samples from  $\mathcal{D}_{RB}$  
+14: end if  
+15: Sample batch of transitions  $s_{\pi}$  from  $\mathcal{D}_{RB}$  
+16: Update  $\pi$  to maximize returns using Eqn. 5  
+17: end for
+
+the current estimate as  $\beta \mu_{\tau_k'}$ , where  $\beta \in [0,1]$  is the decay factor. In our experiments, we use  $\beta = 0.9$  (see Section 10
+
+![](images/8ef753083cc86bda59447c94938d1ffc3dda0c357eb617890769b1e7e6e8b1f5.jpg)  
+(a)Drawer-Open
+
+![](images/7d358f391b54bdb2b37a978deea2c4f778a50b2d11ec72a61a9374096bca782a.jpg)  
+(b)Drawer-Close
+
+![](images/9c1e78b9f15b94e093078fedddfbdfc5ba3d2c39d521d8431b8d70bc0e0d83c7.jpg)  
+(e) door-open
+
+![](images/dc9f919d63070feb97b69e2d333d27a283d5cc7ef3981be53225e0a3274c25a0.jpg)  
+(f) drawer-open
+
+![](images/d27dc575cb125502a04af3e1c3a93d8f3a10ce08fbecd6a29725dfd110e397e2.jpg)  
+(g) hammer
+
+![](images/8ece76161aad36138e3fea56afb574f4e4a1a4897b622239fa1bcddac1cc7df8.jpg)  
+(c)Push  
+Figure 3. Visualization of the robotic tasks: (a-d) Real world environments with a UR5 arm, (e-j) simulation environments for evaluation using the Meta-World [44] benchmark.
+
+![](images/77075d99d86291147d77d899d0ad5a2a1f3148d1c0f61ca5803ec952e2d30642.jpg)  
+(d) Pick-Place-Cup
+
+![](images/a0ade43f6cfcdd286f1c582a5b15a721cdcdc72286a97bff2c81df59c48aa7b4.jpg)  
+(h) peg-insert-side
+
+![](images/7d995282ac819cc92f1ce3319bfc63a2bdf4c7202472c28fb61cdbf7ca23237a.jpg)  
+(j) pick-place
+
+![](images/be06301878795182375c86c32dfa58dd6588975d537320e6361d6e309f692e00.jpg)  
+(i) reach
+
+for an ablation of  $\beta$ . We formulate our online refinement objective as:
+
+$$
+p _ {\text {p u s h - b a c k}} = \operatorname {s g} \left(\mathcal {N} \left(\beta \mu_ {\tau_ {k} ^ {\prime}}, \frac {1}{(g - i) ^ {2}}\right)\right)
+$$
+
+$$
+\mathcal {L} _ {\text {p u s h - b a c k}} = D _ {\mathrm {K L}} \left(p _ {\text {p u s h - b a c k}} \| \boldsymbol {E} _ {\theta} \left(\boldsymbol {o} _ {i} ^ {\tau_ {k} ^ {\prime}}, \boldsymbol {o} _ {j} ^ {\tau_ {k} ^ {\prime}}, \boldsymbol {o} _ {g} ^ {\tau_ {k} ^ {\prime}}\right)\right), \tag {7}
+$$
+
+where  $\mathrm{sg}(\cdot)$  denotes the stop gradient operator. With this approach, we differentiate between the two sources of trajectories by treating observations from the (inferior) online roll-out trajectories  $\tau'$  as corresponding to lower progress compared to expert trajectories  $\tau$ . In this case, optimizing for  $\pi^*$  would encourage the agent to consistently improve its actions until the policy trajectories are well aligned with the distribution of expert trajectories  $\tau_k$ .
+
+During online training, we fine-tune  $E_{\theta}$  (3) using hybrid objectives: we optimize  $\mathcal{L}_{\mathrm{push - back}}$  based on  $\tau_k^{\prime}$  sampled from the most recent buffers and at the same time, we optimize  $\mathcal{L}_{\mathrm{expert}}$  using samples from  $\tau_k$ , preventing  $E_{\theta}$  from being biased by  $\tau_k^{\prime}$ .
+
+# 5. Experimental Evaluation
+
+We evaluate the effectiveness with which PROGRESSOR learns reward functions from visual demonstrations that enable robots to perform various manipulation tasks in simulation as well as the real world. The results demonstrate noticeable improvements over the previous state-of-the-art.
+
+# 5.1. Simulated Experiments
+
+In our simulation experiments, we use benchmark tasks from the Meta-World environment [44], selecting six table-top manipulation tasks (Figure 3 e-j): door-open, drawer-open, hammer, peg-insert-side, pick-place, and reach. For all tasks, we train the model-free reinforcement learning algorithm DrQ-v2 [43] using the reward predicted by PROGRESSOR in place of the environment reward.
+
+We compare our method to five baselines. TCN [35] is a temporal contrastive method that uses optimized embed
+
+dings to measure task progress. GAIL [17] is an adversarial approach that aligns the state-action distribution of a policy with that of an expert. GAIL's discriminator is used as a reward function for our baseline following Yang et al. [41]. SOIL [33] jointly trains an inverse dynamics model and policy by predicting actions from state-only demonstrations. ROT [15] leverages optimal transport to match the expert and agent's trajectory to provide extra regularization for robot learning. Finally, Rank2Reward [41] is a state-of-the-art reward learning method that combines temporal ranking and classification of expert demonstrations to estimate rewards. Unlike our method, which uses a single model for all tasks, Rank2Reward requires a separate model for each task and needs an additional classifier for online adaptation.
+
+We collected 100 expert trajectories for each of the six tasks using the Oracle from Meta-World. All methods were trained for 1.5 million steps, relying solely on the reward estimates produced by their respective models. To ensure a fair comparison, we used the default hyperparameters reported for each method under the same experimental setup.
+
+We evaluate performance using episodic return, which measures the cumulative reward an agent receives during each episode. To ensure robustness, we average the episodic returns over five different seeds for both our approach and the baselines. This setup enables a comprehensive and controlled comparison across all tasks, allowing us to assess the impact of our proposed method in a simulated environment.
+
+We present the results in Figure 4. Our method significantly outperforms the baseline methods across most tasks, with a notable advantage in efficiency. In the drawer-open and hammer tasks, PROGRESSOR requires only  $10\%$  of the total training budget to achieve higher rewards than the baselines. In the door-open and peg-insert-side tasks, where almost all other methods fail entirely, our approach demonstrates strong performance and a higher success rate.
+
+Ablating Push-back We examine the impact of adversarial online refinement via push-back in our framework. Across various tasks (Figure 4), we find that online pushback significantly enhances PROGRESSOR's performance, particularly on challenging tasks such as door-open and hammer. These tasks demand more precise discrimination between expert behavior and online roll-outs to provide an informative reward signal. Notably, PROGRESSOR without push-back still outperforms every baseline on the drawer-open and reach tasks.
+
+In the case of the reach task, we observe that PROGRESSOR with push-back shows a decrease in return as training progresses, a trend that is less pronounced in the version without push-back. We hypothesize that this occurs because the model initially achieves a high success rate—reaching a peak with only  $10\%$  of the total steps—leading
+
+![](images/98ef3a101aa21e2fc028c4d1eb65d24d53b1312fe229f4ebd725f72bc68cceb3.jpg)  
+Figure 4. Visualization of policy learning in the Meta-World [44] simulation environment. We compare PROGRESSOR to five contemporary baselines on six diverse tasks of various levels of difficulty. We also run PROGRESSOR without online push-back as an ablation. We follow the standard evaluation protocol and plot the mean and standard deviation of (left) the environment reward during training and (right) the task success rate from ten rollouts averaged over five seeds. PROGRESSOR demonstrates clear advantages in both metrics, especially at early stages of training.
+
+the refinement process to overly penalize the agent even when its behavior closely resembles the expert.
+
+Online training with non-expert data is crucial for robust performance. TCN, which is only trained on expert demonstrations without online updates, fails at every task due to the domain shift between pretraining and online RL.
+
+# 5.2. Real-World Robotic Experiments
+
+Going beyond simulation, we use experiments with a real-world manipulator to demonstrate how PROGRESSOR, pretrained on the egocentric EPIC-KITCHENS dataset [6, 7], can efficiently learn robotic tasks from a limited number of demonstrations, even when some are unsuccessful. Our approach enhances sample efficiency and robustness to noisy data in offline RL, making it more effective than traditional behavior cloning (BC) methods.
+
+# 5.2.1. Pretraining on Kitchen Dataset
+
+Using ResNet34 [16] as a backbone, we first pretrain our encoder  $E_{\theta}$  to minimize  $\mathcal{L}_{\mathrm{expert}}$  (Eqn. 4) using the P01-P07 sequences from the EPIC-KITCHENS dataset that consist of approximately 1.29M frames. We randomly sample frame triplets  $(o_i, o_j, o_g)$  from the videos ensuring a maximal frame gap  $\| i - g \| \leq 2000$ . To improve the robustness
+
+<table><tr><td>Task</td><td>Behavior of Failed Demonstrations</td></tr><tr><td>Drawer-Open</td><td>The robot approaches the drawer but misses the handle by a few centimeters, failing to open it.</td></tr><tr><td>Drawer-Close</td><td>The robot misses the drawer by a few centimeters, failing to close it.</td></tr><tr><td>Push-Block</td><td>The robot moves the block halfway toward the target cup but stops before reaching it.</td></tr><tr><td>Pick-Place-Cup</td><td>The robot lowers to pick up the cup but misses it by a few centimeters and remains in that position.</td></tr></table>
+
+Table 1. Failure modes in the collected suboptimal trajectories for the real-world robotic experiments. For each task we collected 20 failed demonstrations to go with the 20 expert demonstrations.
+
+of  $E_{\theta}$ , we additionally train with distractor frames where we replace the current observation  $o_j$  with a frame  $o_{j'}$  randomly sampled from a different video sequence as a negative example. For negative triplets, we replace  $\mu$  in Eqn. 3 with -1 as the label. We train our model with batch size
+
+![](images/1a7bf6c63ad2fc5b1b59f75648fce36dd349ddbbb94ddb006e5dc1de631a173b.jpg)  
+Figure 5. Success rates for four real-world tasks, where RWR- ACT is trained on a combination of correct and failed demonstrations using PROGRESSOR, GVL, R3M, and VIP as reward models.
+
+128 for 30000 iterations using the Adam optimizer [21] with constant learning rate  $2e - 4$ .
+
+# 5.2.2. Baselines
+
+We compare PROGRESSOR with GVL [26], R3M [28], and VIP [24]. VIP and R3M are self-supervised visual representations shown to provide dense reward functions for robotic tasks. GVL frames value estimation as a temporal ordering task over shuffled video frames, leveraging pretrained vision-language models (VLMs). To compare with GVL, we use Gemini-1.5-Pro as the VLM following Ma et al. [26]. R3M is trained via time contrastive learning, where the distance between frames closer in time is smaller than for images farther apart in time. Additionally, they leverage L1 weight regularization and language embedding consistency losses. VIP uses a contrastive approach treating frames close in time as negative pairs and those farther away as positive pairs towards learning visual goal-conditioned value functions. For fair evaluation, we pretrained both R3M and VIP on the same EPIC-KITCHENS subsets that our method was trained on.
+
+# 5.2.3. Real-World Few-Shot Offline Reinforcement Learning with Noisy Demonstrations
+
+Following the offline reinforcement learning experiments of Ma et al. [24], we use the reward-weighted regression (RWR) method [30, 31]. Our aim in applying RWR is to show that a reward model trained on human videos can help robots learn to perform tasks even when the training data contains noisy, suboptimal trajectories. In scenarios with highly suboptimal trajectories, vanilla behavior loning (BC) approaches often struggle to learn performant policies, essentially regressing to the mean [20]. However, goal-conditioned reward weighting provides a signal that focuses learning from the correct sub-trajectories, enabling the visual imitation learning model to effectively learn accurate action executions.
+
+To this end, we start with a vanilla BC model and adapt the loss function to be weighted by a pretrained reward model. Specifically, we employ Action-Chunking
+
+Transformer (ACT) [46], a BC model designed for learning fine-grained manipulation tasks. We modified ACT's reconstruction loss—originally defined as the mean absolute error between the predicted action and the ground-truth action—to incorporate RWR as follows:
+
+$$
+\mathcal {L} _ {\text {r e c o n s t}} = \exp (\omega \cdot \hat {\mathbf {r}}) \cdot \| \pi_ {\mathrm {A C T}} (\boldsymbol {o} _ {j}) - a _ {j} \| _ {1}, \tag {8}
+$$
+
+where  $\hat{\mathbf{r}}$  is the reward prediction for the current observation  $o_j$ ,  $\omega$  is a temperature parameter,  $a_{j}$  is the ground-truth action, and  $\pi_{\mathrm{ACT}}(o_j)$  is the action predicted by ACT by taking the current observation image  $(o_j)$  as input. We refer to ACT trained with the reconstruction loss in Equation 8 as RWR-ACT, to distinguish it from the standard ACT.
+
+We compare PROGRESSOR with GVL, R3M and VIP by freezing the pretrained models and using them as reward prediction models to train RWR-ACT on downstream robot learning tasks. We use the reward predictions from these models in place of  $\mathbf{u}_{\theta}$  in Equation 8. In both VIP and R3M, the reward prediction is parameterized by the current observation, previous observation, and goal observation. Following Ma et al. [24], we define the reward for a given transition tuple  $(o_{t}, o_{t+1})$  as  $R(o_{t}, o_{t+1}; \phi, o_{g}) := S_{\phi}(o_{t+1}; o_{g}) - S_{\phi}(o_{t}; o_{g})$ , where  $S_{\phi}$  is the distance function in the  $\phi$ -representation space extracted from VIP and R3M, defined as  $S_{\phi}(o_{t}; o_{g}) := -\|\phi(o_{t}) - \phi(g)\|_{2}$ .
+
+In contrast, our method parameterizes reward using the initial observation, current observation, and goal observation. Additionally, we include vanilla ACT as a baseline, which applies uniform weighting (i.e.,  $\omega = 0$ ) across all observations in the training trajectories.
+
+We design four tabletop robotic manipulation tasks (Figure 3 a-d): **Drawer-Open**, **Drawer-Close**, **Pick-Place-Cup**, and **Push-Block**. For each task, we collect 40 demonstrations, half of which are suboptimal and fail to complete the task. Table 1 summarizes the behaviors in failed demonstrations. Including these failed demonstrations is crucial for evaluating whether the learned reward model can accurately signal progress toward a goal by assigning high reward to transitions that lead toward completion and low reward to those that do not. We categorize **Drawer-Close** as an easy task and the other three as hard tasks. This distinction in difficulty is based on the complexity of the tasks and the level of suboptimality in the failed demonstrations. Section 9 provides detailed task and data descriptions along with frame sequences from both successful and failed trajectories.
+
+We train all policies using the same hyperparameters employed by Zhao et al. [46] for training ACT in their real-world behavior cloning experiments (see Section 9.4 for more details). For all RWR experiments, we set  $\omega = 0.1$ . We evaluate each method over 20 rollouts and report the success rate considering the criteria in Table 4.
+
+The results presented in Figure 5 highlight the significant
+
+![](images/9efea4bd8c4b9271a753e78cc7451551a1208ff515d255864350a90a3b1370b3.jpg)  
+Figure 6. Estimated progress curves for correct and incorrect demonstrations in the Push-Block task from PROGRESSOR, GVL[R1], R3M, and VIP. PROGRESSOR exhibits the most distinct separation for incorrect demonstrations, contributing to the superior performance
+
+advantage of PROGRESSOR-RWR-ACT, which on average outperforms all baseline methods. This advantage is particularly pronounced on the more challenging tasks, such as Push-Block and Open-Drawer, for which other methods struggle. While GVL-RWR-ACT performs comparably to PROGRESSOR-RWR-ACT on the Pick-Place-Cup and Close-Drawer tasks, its performance drops significantly on the more challenging Open-Drawer and Push-Block tasks. Similarly, R3M-RWR-ACT and VIP-RWR-ACT show moderate success on simpler tasks, such as Driver-Close, yet their performance drops significantly for the Push-Block and Open-Drawer tasks. As shown in Figure 6, PROGRESSOR provides significantly more accurate progress estimates for both failed and successful trajectories compared to GVL, R3M, and VIP, particularly starting from around step 55 when the robot starts to deviate from the goal observation. This supports our hypothesis that PROGRESSOR provides more distinct reward weighting for both correct and failed trajectories compared to the baselines. Vanilla ACT performs poorly, especially on difficult tasks. Since ACT assigns equal weight to all sub-trajectories, it cannot prioritize better transitions or filter out noise effectively, leading to poor performance.
+
+Figure 7 shows reward predictions for a sample frame sequence from EPIC-KITCHENS and a trajectory from our Pick-Place-Cup robot task. The figure shows that, despite being pretrained solely on human videos, PROGRESSOR generates well-formed reward predictions for the robotic task in a zero-shot setting.
+
+# 6. Conclusion
+
+In this work, we presented PROGRESSOR, a self-supervised framework that learns task-agnostic reward functions from video via progress estimation. By learning to estimate progress between observations in expert trajectories, PROGRESSOR generates dense rewards that effectively guide policy learning. During online RL training, PROGRESSOR further refines this progress estimation through an adversar
+
+![](images/c42e99b3f94fb811498a0592ae6f024a9b1a80cbfd166384b1d9596c043d3edc.jpg)  
+(a) The evolution of the episodic reward on human video
+
+![](images/31073b920ea3009132a81590d5dd27c7aa2cae8f24e0835498648d9ef7cadd3c.jpg)  
+(b) The evolution of the episodic reward on robot video  
+Figure 7. Visualization of the predicted rewards from Epic Kitchen pretrained PROGRESSOR, evaluated on a sample clip from EPIC-KITCHENS (top) and on sample frames from an unseen Pick-Place-Cup demonstration on our robot (bottom).
+
+ial push-back strategy, helping the model handle non-expert observations and minimize distribution shift. Our method outperforms previous state-of-the-art approaches across a range of simulated experiments. Through real-world experiments, we show that by pretraining on in-the-wild human videos, PROGRESSOR enables robots to learn from limited and noisy task demonstrations, outperforming other visual reward models. Meanwhile, our framework opens several avenues for future improvement. Currently, it assumes that progress is monotonic, which may not be valid for all tasks. One means of relaxing this assumption may be to learn a multimodal or cycle-aware model of task progress. Another direction for future work is to employ a dynamic weighting factor  $\beta$  for refinement, which may further enhance performance.
+
+# References
+
+[1] Pieter Abbeel and Andrew Y. Ng. Apprenticeship learning via inverse reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2004. 2  
+[2] Minttu Alakuijala, Gabriel Dulac-Arnold, Julien Mairal, Jean Ponce, and Cordelia Schmid. Learning reward functions for robotic manipulation by observing humans. arXiv preprint arXiv:2211.09019, 2023. 1  
+[3] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. arXiv preprint arXiv:2207.09450, 2022. 2  
+[4] Bowen Baker, Ilge Akkaya, Peter Zhokov, Joost Huizinga, Jie Tang, Adrien Ecoffet, Brandon Houghton, Raul Sampedro, and Jeff Clune. Video pretraining (VPT): Learning to act by watching unlabeled online videos. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2  
+[5] Serena Booth, W Bradley Knox, Julie Shah, Scott Niekum, Peter Stone, and Alessandro Allievi. The perils of trial-and-error reward design: Misdesign through overfitting and invalid task specifications. In Proceedings of the National Conference on Artificial Intelligence (AAAI), 2023. 1  
+[6] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Scaling egocentric vision: The EPIC-KITCHENS dataset. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2, 6, 1  
+[7] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. International Journal on Computer Vision, pages 33--55, 2022. 6  
+[8] Alejandro Escontrela, Ademi Adeniji, Wilson Yan, Ajay Jain, Xue Bin Peng, Ken Goldberg, Youngwoon Lee, Danijar Hafner, and Pieter Abbeel. Video prediction models as rewards for reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2  
+[9] Benjamin Eysenbach, Tianjun Zhang, Ruslan Salakhutdinov, and Sergey Levine. Contrastive learning as goal-conditioned reinforcement learning. arXiv preprint arXiv:2206.07568, 2023. 1, 2  
+[10] Chelsea Finn, Sergey Levine, and Pieter Abbeel. Guided cost learning: Deep inverse optimal control via policy optimization. arXiv preprint arXiv:1603.00448, 2016. 2  
+[11] Justin Fu, Katie Luo, and Sergey Levine. Learning robust rewards with adversarial inverse reinforcement learning. arXiv preprint arXiv:1710.11248, 2018. 2  
+[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17(59):1-35, 2016. 4  
+[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and
+
+Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 4  
+[14] Dylan Hadfield-Menell, Smitha Milli, Pieter Abbeel, Stuart J Russell, and Anca Dragan. Inverse reward design. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 1  
+[15] Siddhant Haldar, Vaibhav Mathur, Denis Yarats, and Lerrel Pinto. Watch and match: Supercharging imitation with regularized optimal transport. In Proceedings of the Conference on Robot Learning (CoRL), 2023. 5  
+[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 1  
+[17] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2016. 5  
+[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. arXiv preprint arXiv:1606.03476, 2016. 2  
+[19] Tao Huang, Guangqi Jiang, Yanjie Ze, and Huazhe Xu. Diffusion reward: Learning rewards via conditional video diffusion. arXiv preprint arXiv:2312.14134, 2023. 1, 2  
+[20] Liyiming Ke, Sanjiban Choudhury, Matt Barnes, Wen Sun, Gilwoo Lee, and Siddhartha Srinivasa. Imitation learning as f-divergence minimization. In Proceedings of the Workshop on the Algorithmic Foundations of Robotics (WAFR), 2020. 7  
+[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7  
+[22] Diederik P Kingma and Max Welling. Auto-encoding variational Bayes. arXiv preprint arXiv:1312.6114, 2013. 3  
+[23] Yunzhu Li, Jiaming Song, and Stefano Ermon. InfoGAIL: Interpretable imitation learning from visual demonstrations. arXiv preprint arXiv:1703.08840, 2017. 2  
+[24] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Obert Bastani, Vikash Kumar, and Amy Zhang. VIP: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2022. 7  
+[25] Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. VIP: Towards universal visual reward and representation via value-implicit pre-training. arXiv preprint arXiv:2210.00030, 2023. 1, 2  
+[26] Yecheng Jason Ma, Joey Hejna, Chuyuan Fu, Dhruv Shah, Jacky Liang, Zhuo Xu, Sean Kirmani, Peng Xu, Danny Driess, Ted Xiao, et al. Vision language models are in-context value learners. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 7  
+[27] Alberto Maria Metelli, Filippo Lazzati, and Marcello Restelli. Towards theoretical understanding of inverse reinforcement learning. arXiv preprint arXiv:2304.12966, 2023. 2  
+[28] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3M: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 7
+
+[29] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3M: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 1, 2  
+[30] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-weighted regression: Simple and scalable off-policy reinforcement learning. arXiv preprint arXiv:1910.00177, 2019. 7  
+[31] Jan Peters and Stefan Schaal. Reinforcement learning by reward-weighted regression for operational space control. In Proceedings of the International Conference on Machine Learning (ICML), 2007. 7  
+[32] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. DexMV: Imitation learning for dexterous manipulation from human videos. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2  
+[33] Ilija Radosavovic, Xiaolong Wang, Lerrel Pinto, and Jitendra Malik. State-only imitation learning for dexterous manipulation. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021. 5  
+[34] Stuart J Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, Englewood Cliffs, NJ, USA, 1 edition, 1995. 1  
+[35] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 5  
+[36] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and Sergey Levine. Time-contrastive networks: Self-supervised learning from video. arXiv preprint arXiv:1704.06888, 2018. 2  
+[37] Satinder Singh, Richard L. Lewis, and Andrew G. Barto. Where do rewards come from? In Proceedings of the International Symposium on AI Inspired Biology, 2010. 1  
+[38] Richard S Sutton and Andrew G Barto. Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA, USA, 2 edition, 2018. 1
+
+[39] Michael C Welle, Nils Ingelhag, Martina Lippi, Maciej K. Wozniak, Andrea Gasparri, and Danica Kragic. Quest2ROS: An app to facilitate teleoperating robots. In Proceedings of the International Workshop on Virtual, Augmented, and Mixed-Reality for Human-Robot Interactions, 2024. 1  
+[40] Markus Wulfmeier, Peter Ondruska, and Ingmar Posner. Maximum entropy deep inverse reinforcement learning. arXiv preprint arXiv:1507.04888, 2016. 2  
+[41] Daniel Yang, Davin Tjia, Jacob Berg, Dima Damen, Pulkit Agrawal, and Abhishek Gupta. Rank2Reward: Learning shaped reward functions from passive video. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2024. 1, 2, 3, 5  
+[42] Mengjiao Yang and Ofir Nachum. Representation matters: Offline pretraining for sequential decision making. arXiv preprint arXiv:2102.05815, 2021. 2  
+[43] Denis Yarats, Ilya Kostrikov, and Rob Fergus. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 5  
+[44] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. MetaWorld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Proceedings of the Conference on Robot Learning (CoRL), 2020. 2, 5, 6, 1  
+[45] Qihang Zhang, Zhenghao Peng, and Bolei Zhou. Learning to drive by watching YouTube videos: Action-conditioned contrastive policy pretraining. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2  
+[46] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 7, 2  
+[47] Brian D. Ziebart, Andrew Maas, J. Andrew Bagnell, and Anind K. Dey. Maximum entropy inverse reinforcement learning. In Proceedings of the National Conference on Artificial Intelligence (AAAI), 2008. 2

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd1b1827e910ae581632d5db33609fb64ad94e8ba9c119d6bc62dea83c5e16fa
+size 563761

ICCV/2025/PROGRESSOR_ A Perceptually Guided Reward Estimator with Self-Supervised Online Refinement/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2a5539d4edbfd418103a5f17e5c780f6767853f1e856396ef0058e73b30ff4c
+size 444871

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/d5f28d0d-4814-43b7-b4fc-d4fe29db4270_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccdb8efdf9f6e413440b02919c55aeb4b25a4329c95178af5f5cd286001cd739
+size 84925

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/d5f28d0d-4814-43b7-b4fc-d4fe29db4270_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ec4e65d5df7a1b8a56906fece3b7f47d55084e1905ab7c4f089a9c138e77b27
+size 102459

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/d5f28d0d-4814-43b7-b4fc-d4fe29db4270_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e86880d6e3829db88add429032175d271d02341c357c65da17ed94ef4949091
+size 897434

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/full.md

@@ -0,0 +1,298 @@
+# PROL : Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning
+
+M. Anwar Ma'sum $^{1,*}$ , Mahardhika Pratama $^{1}$ , Savitha Ramasamy $^{2}$ , Lin Liu $^{1}$ , Habibullah Habibullah $^{1}$ , and Ryszard Kowalczyk $^{1,3}$
+
+$^{1}$ STEM University of South Australia,  $^{2}$ Institute for Infocomm Research, A*STAR & IPAL, CNRS@CREATE
+
+<sup>3</sup>Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
+
+# Abstract
+
+The data privacy constraint in online continual learning (OCL), where the data can be seen only once, complicates the catastrophic forgetting problem in streaming data. A common approach applied by the current SOTAs in OCL is the use of memory-saving exemplars or features from previous classes to be replayed in the current task. On the other hand, the prompt-based approach performs excellently in continual learning, but at the cost of a growing number of trainable parameters. The first approach may not be applicable in practice due to the data openness policy, while the second approach has the issue of throughput associated with the streaming data. In this study, we propose a novel prompt-based method for online continual learning that includes 4 main components: (1) single light-weight prompt generator as a general knowledge, (2) trainable scalar-and-shifter as specific knowledge, (3) pre-trained model (PTM) generalization preserving, and (4) hard-soft updates mechanism. Our proposed method achieves significantly higher performance than the current SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset. Our complexity analysis shows that our method requires a relatively smaller number of parameters and achieves moderate training time, inference time, and throughput. For further study, the source code of our method is available at https://github.com/anwarmaxsum/PROL.
+
+# 1. Introduction
+
+Continual learning (CL) has become a spotlight in the world of artificial intelligence (AI) involving dynamic environments, as it tackles catastrophic forgetting for a continually evolving sequence of tasks [5, 38]. CL methods are proven to be promising in various domains, e.g., computer vision, automation, NLP and graph analysis [1, 21, 31, 35]. Online continual learning (OCL) is an advanced subproblem of CL where CL is conducted in streaming data.
+
+The limitation of seeing data only once in OCL leads to insufficient learning of a deep learning model that results in difficulty in gaining knowledge of a new task while maintaining knowledge of previously learned tasks. In other words, it complicates how to resolve the catastrophic forgetting problem. To handle such a limitation, most of the state-of-the-art OCL methods (SOTAs) save previous task exemplars for a rehearsal process to maintain the knowledge from previously learned tasks [30, 39, 49]. Practically, this approach is not reliable since, due to privacy constraints or data openness policy, previously learned data can be unavailable anymore. In addition, the rehearsal process increases memory consumption and computational costs.
+
+Leveraging the recent advancement of the foundation model, recent SOTAs utilized a pre-trained model (PTM) that carries a generalization attribute as initial weight [22, 25, 48]. However, the SOTAs still store and replay previously learned exemplars or features in their memory. The insight tells us that rehearsal-free online continual learning remains an open problem. On the other hand, the prompt-based approach has been proven to be effective in rehearsal-free continual learning. Specifically, the growing prompt components methods [17, 28, 33] outperform the prompt-pool method [44] and task-wise prompt methods [7, 37, 42, 43] both in terms of accuracy and forgetting measure. However, from the perspective of OCL, the growing components prompt methods violate the throughput aspects, as the model parameters grow with the increasing number of classes. This growth incurs more operations and longer training time, leading to more skipped (unprocessed) data streams. Thus, the approach is not favorable for dealing with catastrophic forgetting in streaming data.
+
+These intriguing insights motivate us to develop a novel rehearsal-free method that can effectively address the OCL problem. We develop our proposed method based on two fundamental goals in OCL, i.e., (1) achieving stability-plasticity and (2) being efficient both in the training and inference phases. To achieve these goals, we design our method with the following components: (1) a single
+
+lightweight prompt generator for all tasks, (2) a set of scalers and shifters associated with the class-wise learnable keys, (3) preservation of PTM generalization via cross-correlation matrices, and (4) hard-soft learning updates for a better parameter tuning. The generator is trained only on the first task to avoid forgetting and achieve stability, while scalers and shifters are trained on every task to achieve plasticity. To achieve high throughput, our lightweight generator has fewer parameters than a single generator in [28] and fewer than a single task-specific prompt in [43]. The third component is adapted from SAFE[51] to the OCL problem.
+
+In this study, our contributions are: (1) We highlight rehearsal-free constraint in online continual learning and propose a novel prompt online learning (PROL) that addresses the two essential goals of OCL. (2). We design a new mechanism to generate prompts by a single generator with a few smaller-shifter parameters. We also propose a joint loss function for our method. (3). Our rigorous experiment and analysis prove the superiority of our method compared to the existing SOTAs in general and historical performance. Our analysis shows that our method has relatively low parameters, moderate training, inference time, and throughput compared to the SOTAs. (4). We provide a complexity analysis of our method and complete numerical results for further study; please see the supplementary.
+
+# 2. Literature Review
+
+Online Continual Learning (OCL): Recent OCL studies shed light on various approaches to handle catastrophic forgetting in streaming data. The backbone tuning approach tunes the whole model with unique mechanisms, e.g. cross-task class discrimination as in GSA [10], dual-view as in DVC [8], proximal point method as in LPR [49], real-synthetic similarity maximization as in ESRM [39], and adaptive layer freezing as in ALF [30]. The other approaches tune the class prototype along with the backbone as in ONPRO [45], augmenting and refining the prototype as in DSR [13], or generating multiple features in angular space as in EARL [29]. Another unique approach uses a student-teacher model and knowledge distillation between them, as in MKD [23], CCLDC [40], and MOSE [41]. The recent development of the foundation model gives rise to the PTM-based OCL, e.g., NSCE [48] that transfers PTM weight and fine-tunes it, RandDumb [25] that trains RBF kernel from PTM generated features, and RanPAC [22] that projects the feature with a randomized matrix. Tackling OCL can be achieved by a customized optimization technique, e.g., POCL [47] that utilizes Pareto optimization, and CAMA [14] that performs confidence-aware moving average. However, these aforementioned methods still save exemplar or features for rehearsal. It is not always applicable in real-world applications since the old data may not be available or accessible. Note that rehearsal-based CL meth
+
+ods [2, 3, 24, 26, 46] work poorly in OCL and even worse than the aforementioned methods.
+
+PTM-Based Continual Learning: The prompt-based approach is one of the most popular approaches in CL due to its simplicity and scalability. The prompt can be generated by learnable latent parameters or a vector generated by trainable generator networks and then prepended into visual transformer (ViT) multi head self attention (MSA) [18, 19]. The prompt structure can be organized by a pooling approach as in L2P [44], task-specific and task-shared structure as in DualPrompt, S-Prompt, HiDePrompt and CPrompt [7, 37, 42, 43], evolving prompt component as in CODA-P[33], evolving prompt generator as in ConvPrompt and EvoPrompt [17, 28]. The other approach utilizes Low-Rank Adaptation (LoRA) where it extends the frozen ViT weight  $W$  with learnable matrices  $A$  and  $B$  [6, 20, 32]. The adapter-based approach emerges as a new PTM approach for CL by extending the ViT with a sequence of (growing) adapters [34, 52]. Rather than extending the inner level, such as ViT weight, adapters work on the feature level by projecting the ViT feature into a more discriminative form. Last but not least, recent works focused on learning mechanisms, i.e., slow-learner rather than structures to elevate PTM model discrimination [50, 51]. Except for the adapters approach, all the mentioned methods were developed for rehearsal-free CL. Even thought they have been proven effective in rehearsal-free CL, the methods are not yet proven in rehearsal free OCL where the data is streaming and can be seen only once.
+
+# 3. Preliminary
+
+(a) Problem Formulation: Class incremental learning (CIL) problem is defined as the problem of learning a sequence of fully supervised tasks  $\{\mathcal{T}^t\}_{t = 1}^T$ , where after finishing learning a task  $\mathcal{T}^t$ , a model must recognize all learned tasks i.e.  $\mathcal{T}^1,\mathcal{T}^2,\dots \mathcal{T}^{t - 1},\mathcal{T}^t$ . Symbol  $T$  represents the number of consecutive tasks. Each task carries pairs of training samples, i.e.,  $\mathcal{T}^t = \{(x_i^t,y_i^t)\}_{i = 1}^{|\mathcal{T}^t|}$  where  $x_{i}\in \mathcal{X}^{t}$  and  $y_{i}\in \mathcal{V}^{t}$  denotes input image and its label, while  $|\cdot |$  denotes cardinality.  $\mathcal{C}^t$  denotes the unique class labels in  $\mathcal{V}^t$ , i.e.,  $\mathcal{C}^t = \text{unique} (\mathcal{V}^t)$ , and  $|\mathcal{C}^t|$  denotes the number of classes in  $\mathcal{T}^t$ . The task ID is visible only during the training phase of CIL, unlike in task incremental learning (TIL), which is visible in both the training and inference phases.
+
+Online Class incremental learning (OCIL) is similar to CIL but the data comes in streams  $s_1, s_2, s_3, \ldots$  where  $s_j = \{(x_i^t, y_i^t)\}_{i=1}^{|s_j|}$  rather than in a storable  $\mathcal{T}^t = \{(x_i^t, y_i^t)\}_{i=1}^{|\mathcal{T}^t|}$  manner.  $|s_j|$  is relatively small, e.g., 10. The data can be seen only once, as the data will not be available (stored) after streaming. Following [15], there are three main settings in OCL, i.e., disjoint task where each task  $t$  is disjoint from another task  $t'$ , i.e.,  $\forall t, t' \neq t$ ,  $(\mathcal{T}^t \cap \mathcal{T}^{t'} = \emptyset)$ , blurry task
+
+![](images/e895f685824edad59f3c173cf0a4cdaab6a74e241d16f5edccfda9cf408f3cba.jpg)  
+Figure 1. Topology of OCL and prompt-based CL methods (a) OCL Methods without pre-trained model (PTM), (b) OCL methods with PTM as weight initiation, (c) Prompt-based CL method. Majority of the OCL method requires memory from previously learned classes to be replayed along with the currently learned classes, while the prompt-based method is rehearsal-free.
+
+![](images/9ca3cb7d4539984bcb313f52275a37280458987bf8d98614e2569621b08485f2.jpg)
+
+![](images/bcf42e2a61ebdfcc23a7c527d27a4f603d104b7a3af6162148c346ecc30ea465.jpg)
+
+setting where there is no explicit boundary between tasks, and iBlurry setting where consist of  $N\%$  of the classes are put into disjoint tasks and the other  $(100 - N)\%$  classes appear in every task. In this study, we focus on the disjoint task setting that is the most common in OCL and CL.
+
+(b) The OCL Method and Its Gap: Figure 1 shows a topology that the OCL methods can be divided into 2 main categories, i.e., OCL method without PTM (Non-PTM) and with PTM. In the perspective of available foundation model, the main gap of this category is that it tunes the whole backbone  $f_{\theta}(.)$  and classifier  $g_{\phi}(.)$  from scratch in the case of streaming data where the data can be seen only once. In addition, the backbone training in every task increases the chance of the model forgetting. Even with the presence of sample memory to avoid forgetting, the methods can not perform well since its weight is not yet convergent. A customized backbone tuning technique such as the proximal point method, real-synthetic similarity maximization, dual view, or adaptive layer freezing [8, 10, 30, 39, 49], or customized optimization method as in [14, 47] can not compensate the insufficient training and avoid model forgetting. The other approach trains dual models, i.e., student  $(f_{\theta}^{student})$  and teacher  $(f_{\theta}^{teacher})$  backbones. Even though one of the model is frozen, this approach requires higher training complexity (longer training time) and reduces model throughput. The prototype approach replaces the MLP classifier  $(g_{\theta})$  with a prototype-based classifier  $(g_{Z})$ . Despite reducing model forgetting, the prototypes do not optimally discriminate the learned classes since the prototypes are calibrated with once-seen data.
+
+On the other side, the PTM-based method leverages PTM generalized weight as its initial values. NSCE[48] transfers the PTM weight and then fine-tunes its weight along with the classifier to adapt to the streaming data. Despite achieving better performance than non-PTM methods,
+
+it still requires memory samples that may not be applicable in real-world applications due to data privacy or openness policy. Feature projection network, e.g., RanPAC[22] and RanDumb[25] transform PTM features into a more discriminative form by a projection network  $Prj_{\psi}$ . Similarly, this approach requires memory (in feature space) from a previous task to be replayed in the currently learned class. The adapter methods [34, 52] that were originally for offline CL are considered as baselines in OCL. Aside from utilizing sample memory, this method has a growing number of adapters ( $A_1, A_2, \dots, A_t$ ) along with the increasing of learned classes. This mechanism is bounded to increase model complexity and reduce model throughput. F-OAL [53] forms a feature fusion from its PTM layers then classify it with analytical learning (AL). Despite working in rehearsal-free setting, the method requires more operation for AL since the least square solver requires higher computation than standard MLP or prototype-based classifier.
+
+(c). Prompt-based CL Methods and its Drawback: The pool-based prompting, e.g., L2P[44] selects top-k prompts from the pool  $[P^1, P^2, \ldots, P^M]$  where  $M$  is the pool size. Any prompt  $P^m$  in the pool can be trained for any task  $t \in [1..T]$ . The task-wise prompting, e.g., DualPrompt[43] trains only the corresponding prompt  $P^t$  that associated to task-wise key  $K^t$  during the training process on task  $t$ , while the prompts and keys associated with the previous tasks i.e  $[P^1, P^2, \ldots, P^{(t-1)}]$  and  $[K^1, K^2, \ldots, K^{(t-1)}]$  are frozen. The growing component approach dynamically increases the number of prompt components, e.g.,  $p^i, p^{i+1}, \ldots, p^M$  to learn new classes while maintaining knowledge on previously learned classes. The final prompt is generated by weighted summing the components, i.e.,  $P = \sum_{i=1}^{M} p^i \alpha^i$  where  $\alpha^i$  denotes similarity value. The prompt components  $p^t$  can be a learnable latent parameter as in [33] or generated by prompt generator
+
+network, i.e.,  $p^i = G^i(x)$  as in [28].
+
+The pool-based approach has a high risk of forgetting since all the prompts  $[P^1, P^2, \dots, P^M]$  are possible to be trained in all tasks  $t \in [1..T]$ . A prompt  $P^i$  is optimal for the  $t^{th}$  task but is tuned again in the  $t + 1^{th}$  task. Therefore, it is no longer optimal for  $t^{th}$  task and leads to forgetting. Task-specific prompting has the risk of high similarity keys between two different tasks, e.g.,  $K^t \approx K^{t'}$  for  $t \neq t'$ , which leads to inaccurate task prediction during inference. The growing component prompting adds new components, i.e.,  $p^i, p^{i+1}, \dots, p^M$  that could disrupt the previous components, i.e.,  $p^1, p^2, \dots, p^{i+1}$  that are already optimal for previous tasks. Similar to the adapters method, this approach could lead to higher model complexity and reduce its throughput due to the growing prompt components or prompt generators.
+
+# 4. Proposed Method
+
+In this study, we propose prompt online learning (PROL) to handle catastrophic forgetting in streaming data. The goal of the method is to achieve stability-and-plasticity and high throughput. Thus, we utilize only a single lightweight generator for all classes in all tasks. The generator is trained only in the base task, then frozen afterwards to maintain stability, and assisted by learnable scalers and shifters to achieve plasticity. Figure 2 shows the architecture and flow of PROL, i.e., part 2(a) shows the generator and learnable parameters, part 2(b) shows the prompt generation mechanism, and part 2(c) shows the model forward, loss, and parameter updates. The following subsections explain the details of the components and the learning mechanism.
+
+# 4.1. Learnable Parameters:
+
+(a). Single Lightweight Generator for All: We design a lightweight generator  $(G_{K}, G_{V})$  to achieve the highest possible throughput in online continual learning that is trained only in the first task yet can be utilized in all tasks. We design it in a pair following the prompt structure in prefix tuning [19] that divides prompt  $P$  into two parts, i.e.,  $P_{K}$  and  $P_{V}$  that prepend into  $K$  and  $V$  of the ViT layer's MSA, respectively. Following ConvPrompt [28], the generator is distributed into all MSA heads of the ViT layers. Thus, the generator can be noted as  $(G_{K}, G_{V}) = \{(G_{K,l,h}, G_{V,l,h})\}, l \in [1..L]h \in [1..H]$ , where  $L$  and  $H$  are the number of heads and MSA heads per layer respectively. Each  $G_{K,l,h}$  or  $G_{V,l,h}$  is a 1DCNN kernel with the size of 3, thus only saving 3 parameters. Therefore, the total parameters of our generator are only  $2 \times H \times L \times 3$ , which is generally less than 1K.
+
+(b). Learnable Scalers-and-Shifters and Class-wise Key: To assist the generator in achieving plasticity, we train a learnable scalers-and-shifters  $(a_{K}^{c},b_{K}^{c},a_{V}^{c},b_{V}^{c})$  associated with a learnable class-wise key  $K^c$  (figure 2(a)), where  $c$  is
+
+the class label that is represented by  $K^c$ . Following the generator pair, we can specify that  $(a_K^c, b_K^c)$  and  $(a_V^c, b_V^c)$  are the scalars-and-shifters for  $G_K$  and  $G_V$  respectively. However, the scaller and shifters are not distributed into all MSA heads and layers but are expanded based on prompt length  $l$ . Thus, each element of the tuple, e.g.,  $a_K^c$  is a vector with  $l - 1$  elements. Note that the first length  $(l = 1)$  is not scaled or shifted, but left original  $G(x)$ . Our method utilizes small prompt length as in DualPrompt [43], i.e.,  $l = 5$  to maximize the model throughput. Based on that setting, the total scalars and shifters for each class are  $4 \times 4 = 16$  parameters. To avoid uncontrollable values of  $(a_K^c, b_K^c, a_V^c, b_V^c)$  we deploy upper and lower bounds both for  $K$  and  $V$  scalars and shifters formulated by the expression below:
+
+$$
+\left(1 - \epsilon_ {a}\right) \leq a ^ {c} \leq \left(1 + \epsilon_ {a}\right), \text {a n d} - \epsilon_ {b} \leq b ^ {c} \leq \epsilon_ {b} \tag {1}
+$$
+
+where  $\epsilon_{a}$  and  $\epsilon_{b}$  are small positive real numbers. The boundaries are deployed to avoid too much perturbation of  $G_{K}(x)$  and  $G_{V}(x)$  that leads to degraded performance.  $K^{c}$  is a  $\mathbb{R}^D$  learnable vector that represents class  $c$ , where  $D$  is the embedding size. By representing class  $c$ ,  $K^{c}$  plays as a reference to select the appropriate scalars and shifters during the inference time based on its similarity to input  $x$ .
+
+# 4.2. Prompt Generation
+
+We propose a novel prompt generation mechanism i.e., by single lightweight generators and scalers-and-shifters. The mechanism is unique to existing prompt-generation methods [17, 28, 33]. Given prompt generator  $(G_K, G_V)$  and scalar-and-shifters  $(a_K^c, b_K^c, a_V^c, b_V^c)$  where each element as  $l$ -elements vector, the prompt is generated by executing equation 2.
+
+$$
+\begin{array}{l} P _ {K} = \left[ P _ {K, 1}; P _ {K, 2}; \dots ; P _ {K, l} \right], P _ {V} = \left[ P _ {V, 1}; P _ {V, 2}; \dots ; P _ {V, l} \right] \\ P _ {K, i} = a _ {K, i - 1} ^ {c} G _ {K} (x) + b _ {K, i - 1} ^ {c}, \tag {2} \\ P _ {V, i} = a _ {V, i - 1} ^ {c} G _ {V} (x) + b _ {V, i - 1} ^ {c}, i \in [ 2.. l ] \\ \end{array}
+$$
+
+where  $"$ ;" denotes the concatenation operation. We add similarity factor  $s$  to  $G(x)$ , i.e.,  $G(x) = s.G(x)$  where  $s$  is the cosine similarity between input and closest key, i.e.,  $s = \text{cosine}(x, K^c)$ . The addition of the similarity factor acts as a controller if the input chooses the wrong key. Following the previous study[28, 43], the matched input  $x$  can be pached embedding, class token, or class feature that we set as default in this study.
+
+# 4.3. Current-to-Previous Classes Orthogonality
+
+We apply current-to-previous classes orthogonality to increase the discriminative capability of our method. Let the method currently at task  $t$  that learns class  $j, j + 1, j + 2 \ldots m$ . The model must discriminate any currently learned class  $c \in [j, j + 1,..m]$  to any previously learned class  $c' \in [1, 2,..j - 1]$ . Due to rehearsal-free constraints, our model can not replay any samples from class  $c'$  with the samples of
+
+![](images/ea5f26d048486e916966f698a49908d159032b9a728eb13c4700fcc0ce319da6.jpg)  
+Figure 2. The proposed PROL diagram. PROL utilizes a single generator  $(G_K, G_V)$  for all classes in all tasks, learnable scalars  $(a_K^i, a_V^i)$ , and shifters  $(b_K^i, b_V^i)$  parameters associated to the class-wise key  $K^i$ . The generator is trained only in the first task and frozen afterward, while the scalers, shifters, and class-wise keys are trained in every task. The prompt is generated by scaling and shifting  $G(x)$  by the selected scalers and shifters associated with the highest similarity key. The model is updated by joining the loss function consisting of intra-task loss  $\mathcal{L}_{intra}$ , inter-task loss  $\mathcal{L}_{intra}$ , orthogonality loss  $\mathcal{L}_{ort}$ , input-to-key similarity loss  $\mathcal{L}_{sim}$  and generalization loss  $\mathcal{L}_{gen}$ .
+
+$c$ . Dealing with the constraint, we utilize  $K^c$  and  $K^{c'}$  that represent  $c$  and  $c'$  respectively, thus the model can always improve its discrimination capability. Following the vector principle, that states 2 orthogonal vectors have a zero dot product, our method optimizes  $K^c$  so that  $K^c.K^{c'} = 0$ .
+
+# 4.4. Preserving Generalization Property
+
+We utilize PTM that has generalization capability as the initial weight of our method. Dealing with the sequence of streaming data that may be infinite, the generalization capability must be preserved so that it can be taken advantage of by the upcoming data. Following the previous study[51], we preserve the generalization capability by cross correlating the feature of PTM (only), i.e.,  $f_{\theta}(x)$  and the feature of PTM and prompt, i.e.,  $f_{\theta;P}(x)$  into a matrix  $M$  by executing equation 3.
+
+$$
+M = 1 / B \left[ f _ {\theta} (x) \right] ^ {T} \left[ f _ {\theta ; P} (x) \right] \tag {3}
+$$
+
+$\mathbf{M}$  is  $D\times D$  matrix, while  $f_{\theta}(x)$  and  $f_{\theta ;P}(x)$  are  $B\times D$  where  $B$  and  $D$  is batch and embedding sizes respectively. The total quantity of the diagonal elements of  $\mathbf{M}$  shows the alignment between the PTM feature  $(f_{\theta}(x))$  and the PTM+prompt tuning feature  $(f_{\theta ;P}(x))$ . The higher the quantity, the higher the generalization preserving.
+
+# 4.5. Learning, Losses, and Final Objectives
+
+In this subsection, we explain the learning objective of our method. For a better understanding, we explain it in a per-component way as follows.
+
+(a) Tuning Generators and scalar-and-shifters: These 2 components work together to generate the prompt that is prepended into ViT layers and go through until the last module, i.e., MLP classifier. Along with the MLP classifier, these two components are tuned by using cross entropy loss that is derived into two types, i.e., intra-task loss  $(\mathcal{L}_{intra})$
+
+and inter-task loss  $\mathcal{L}_{inter}$ .  $\mathcal{L}_{intra}$  improves model discrimination within a task, specifically, currently learned classes, while  $\mathcal{L}_{inter}$  improves model discrimination for all learned classes, including current and old classes. The intra and inter-task loss are formulated as in equations 4 and 5.
+
+$$
+\mathcal {L} _ {i n t r a} = - \sum_ {c \in C ^ {t}} \log \frac {\exp \left(g _ {\phi} \left(f _ {\theta ; P} (x)\right) [ c ]\right)}{\sum_ {c ^ {\prime} \in C ^ {t}} \exp \left(g _ {\phi} \left(f _ {\theta ; P} (x)\right) [ c ^ {\prime} ]\right)} \tag {4}
+$$
+
+$$
+\mathcal {L} _ {\text {i n t e r}} = - \sum_ {t = 1} ^ {t = T} \sum_ {c \in \mathcal {C} ^ {t}} \log \frac {\exp \left(g _ {\phi} \left(f _ {\theta ; P} (x)\right) [ c ]\right)}{\sum_ {t = 1} ^ {t = T} \sum_ {c ^ {\prime} \in \mathcal {C} ^ {t}} \exp \left(g _ {\phi} \left(f _ {\theta ; P} (x)\right) [ c ^ {\prime} ]\right)} \tag {5}
+$$
+
+(b) Tuning Task-wise learnable key:  $K^c$  has two important roles in our method, i.e., as the class representation to select scalers-and-shifters and improving model discrimination through orthogonality. As the class representation, we tune  $K^c$  tuned closer to the input sample of class  $c$  with similarity loss  $\mathcal{L}_{sim}$  as expressed in the equation 6.
+
+$$
+\mathcal {L} _ {s i m} = - \left(x. K ^ {c}\right) / \left(\max  \left(\left| | x | | _ {2}. \right| \left| K ^ {c} \right| | _ {2}, \epsilon\right)\right) \tag {6}
+$$
+
+To improve class discrimination, we tune any  $K^c$  from currently learned classes to be orthogonal to any  $K^{c'}$  from previously learned classes. Let  $y = [c_1, c_2, \dots, c_B]$  is the class labels of batch-wise input  $x$ ,  $K_{new} = [K^{c_1}, K^{c_2}, \dots, K^{c_B}]$  is the class-wise keys matched with  $y$ , and  $K_{old} = [K^{c_1'}, K^{c_2'}, \dots, K^{c_B'}]$ , then the model will be improved by orthogonality loss  $\mathcal{L}_{\text{ort}}$  as formulated in equation 7.
+
+$$
+\mathcal {L} _ {o r t} = \frac {1}{B} \sum_ {i = 1} ^ {B} K ^ {c i}. K ^ {c ^ {\prime} i} \tag {7}
+$$
+
+(c) Tuning Generalization preserving Matrix: The diagonal elements of matrix  $M$  represent the agreement between PTM only prototype  $(f_{\theta}(x))$  and PTM+Prompt prototype  $(f_{\theta;P}(x))$ , while the non-diagonal represent the disagreement between them. Thus we tune the generalization matrix
+
+$M$  with generalization loss  $\mathcal{L}_{gen}$  as formulated in the equation 8. The first term of  $\mathcal{L}_{gen}$  push the model towards the agreement, while the second term of the  $\mathcal{L}_{gen}$  avoids the disagreement of the prototypes.
+
+$$
+\mathcal {L} _ {\text {g e n}} = \frac {1}{D} \sum_ {i = 1} ^ {D} \left(1 - M _ {i, i}\right) ^ {2} + \frac {1}{D (D - 1)} \sum_ {i = 1} ^ {D} \sum_ {j \neq i} M _ {i, j} ^ {2} \tag {8}
+$$
+
+(d) Joint Loss and Final Objective: Finally, accommodating the losses for all components of our proposed method, i.e., single lightweight generators, scalers-and-shifter, class-wise keys, and generalization matrix, we define a new joint loss function as in equation 9. To enhance the flexibility of our model we add losses coefficient  $\lambda_{1},\lambda_{2},\lambda_{3},\lambda_{4},\lambda_{5}$ .
+
+$$
+\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {i n t r a}} + \lambda_ {2} \mathcal {L} _ {\text {i n t e r}} + \lambda_ {3} \mathcal {L} _ {\text {s i m}} + \tag {9} \\ \lambda_ {4} \mathcal {L} _ {o r t} + \lambda_ {5} \mathcal {L} _ {g e n} \\ \end{array}
+$$
+
+(e) Hard-Soft Update (HSU) Strategy: Dealing with streaming data that the learning cannot be repeated into several epochs, we devise a learning strategy for updating the model. We utilize Hard-Soft Update (HSU) where the model adaptively switches from-and-into hard update, i.e., update with constant and high learning rate, and soft update, i.e., update with decayed learning rate. The hard update is started when the model see new class labels and switch into soft update if the cross entropy loss, i.e.,  $\mathcal{L}_{ce} = \lambda_1\mathcal{L}_{intra} + \lambda_2\mathcal{L}_{inter}$  less than a loss threshold  $\mathcal{L}_{thres}$ . The decay is executed by a cosine annealing scheduler.
+
+# 5. Result and Analysis
+
+# 5.1. Experiment Setting
+
+We evaluate our method in 4 of the most popular datasets in CL i.e. CIFAR100[11], ImageNet-R [16], ImageNet-A [12], and CUB[36]. CIFAR100 contains 100 classes images, while the other has 200 classes. We uniformly split the dataset into 10 tasks. We compare our proposed method with the current SOTAs of PTM-based OCL, i.e., RanDumb [25], RanPAC [22], and MOS [34], the extension of EASE [52]. We adapt the official code of RanDumb and RanPAC from [25] into our rehearsal free and disjoint setting. We also adapt the rehearsal version (denoted with -R) with 5 samples per class as memory, and the joint version (denoted with -J) for both of them. We adapted MOS from its official code into rehearsal free setting. We add L2P [44], DualPrompt [43], and ConvPrompt [28] as representation of pool-based, task-specific, and growing components prompting, respectively. To ensure fairness, all the consolidated methods are run with the same dataset split, backbone, i.e., ViT-B/16 pre-trained on Imagenet-21K, and a 10 chunk size. We also compare our method to non-PTM methods [3, 4, 9, 10, 27, 39, 45] as reported in [39]. The learning rates for all methods are set by grid search from the range
+
+of [0.001,0.005,0.01,0.05,0.1]. All the methods are run under the same environment, i.e., a single NVIDIA A100 GPU with 40 GB RAM, with 3 different random seeds. We utilize Adam optimizer for PROL, while the rest follows their official implementation. We evaluate final average accuracy (FAA), cumulative average accuracy (CAA), and final forgetting measure (FM), see supplementary.
+
+# 5.2. Result and Analysis
+
+a) Overall Performance: Table 1 shows the summarized performance of the consolidated method in 4 benchmark datasets. In term of accuracy, our method outperforms the existing SOTAs with a significant gap, i.e.,  $2 - 76\%$  FAA and  $2 - 64\%$  CAA, respectively, except in the CUB dataset. Surprisingly, our method outperforms the joint version of RanDumb and RunPAC with a  $2 - 20\%$  gap both for FAA and CAA, even though those methods join all the samples' features together. Our method outperforms their replay versions (RanPAC-R and RanDumb-R) with a high margin, i.e.,  $20 - 35\%$  FAA and CAA margins. The rehearsal free of those methods perform extremely poor indicated by the very low FAA and CAA. Our method outperforms the prompt-based methods with a significant margin, i.e.,  $2 - 20\%$  margin both for FAA and CAA. ConvPrompt is the closest competitor to our method with  $2 - 3\%$  difference in all cases. Our method outperforms the adapter method (MOS) with a significant margin, i.e.,  $2 - 21\%$  FAA and  $2 - 10\%$  CAA, respectively. In CUB dataset, our method has lower performance than RanPAC-J and RanDumb-J, comparable performance with RanPAC-R and RanDumb-J, and outperforms the other methods with a convincing gap, i.e.,  $2 - 63\%$  FAA and  $2 - 52\%$  CAA, respectively. In terms of forgetting, our method achieves a fairly low forgetting rate, i.e., slightly higher than Convprompt methods in many cases but significantly lower than RanPAC, RanDumb, and MOS. Our method even achieves lower forgetting in 2 of 4 datasets than RanPAC-J and RanDumb-J. Table 2 shows that despite having a smaller backbone size, our method outperform the non-PTM methods with a significant gap, i.e.,  $\geq 16\%$ . It shows that PTM and parameter-efficient tuning are preferable to a bigger size backbone and memory.
+
+b) Historical Performance: Figure 3 plots the detailed performance of the consolidated methods after finishing each task. The figure shows that in CIFAR100, ImageNet-R, and ImageNet-A datasets, our method consistently achieves the highest average accuracy from the first task until the last task. In addition, the figure shows that our method has the smallest slope, indicating that it suffers from the smaller performance degradation in the learning process. In CUB dataset, our method consistently achieves higher performance than all rehearsal-free methods and competes with RanPAC-R and RanDumb-R, even though it has lower performance than RanPAC-J and RanDUmb-J. Note that both the -R and -J methods replay the previous data, and those
+
+Table 1. Summarized performance of consolidated methods in 4 benchmark dataset, i.e., CIFAR100, Magnet-R, Magnet-A, and CUB200. The official code of ConvPrompt is not runnable on Imagenet-A due to its descriptor unavailability to determined new generators for continual tasks. PET and Rhsl denotes parameter efficient tuning and rehearsal respectively.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">PET Type</td><td rowspan="2">Rhsl</td><td colspan="3">CIFAR100</td><td colspan="3">ImageNet-R</td><td colspan="3">ImageNet-A</td><td colspan="3">CUB200</td></tr><tr><td>FAA</td><td>CAA</td><td>FM</td><td>FAA</td><td>CAA</td><td>FM</td><td>FAA</td><td>CAA</td><td>FM</td><td>FAA</td><td>CAA</td><td>FM</td></tr><tr><td>RanPAC-J [22]</td><td>Rand. matrix</td><td>yes</td><td>79.94</td><td>85.91</td><td>6.43</td><td>52.09</td><td>56.21</td><td>6.63</td><td>45.03</td><td>53.30</td><td>6.97</td><td>85.20</td><td>86.31</td><td>2.63</td></tr><tr><td>RanDumb-J [25]</td><td>RBF kernel</td><td>yes</td><td>80.39</td><td>86.15</td><td>6.25</td><td>52.63</td><td>56.13</td><td>6.04</td><td>44.07</td><td>52.30</td><td>4.98</td><td>85.82</td><td>86.84</td><td>2.52</td></tr><tr><td>RanPAC-R [22]</td><td>Rand. matrix</td><td>yes</td><td>59.66</td><td>69.35</td><td>34.30</td><td>28.16</td><td>35.37</td><td>39.81</td><td>28.54</td><td>35.16</td><td>32.05</td><td>75.63</td><td>79.37</td><td>14.65</td></tr><tr><td>RanDumb-R [25]</td><td>RBF kernel</td><td>yes</td><td>50.93</td><td>62.14</td><td>47.35</td><td>21.91</td><td>29.90</td><td>48.81</td><td>24.86</td><td>30.82</td><td>34.79</td><td>73.81</td><td>78.70</td><td>17.45</td></tr><tr><td>RanPAC [22]</td><td>Rand. matrix</td><td>no</td><td>9.52</td><td>27.92</td><td>94.52</td><td>7.24</td><td>18.81</td><td>67.27</td><td>3.77</td><td>14.86</td><td>57.76</td><td>9.52</td><td>27.06</td><td>88.98</td></tr><tr><td>RanDumb [25]</td><td>RBF kernel</td><td>no</td><td>9.56</td><td>27.94</td><td>94.51</td><td>7.32</td><td>18.80</td><td>67.67</td><td>3.69</td><td>14.56</td><td>55.15</td><td>9.53</td><td>27.21</td><td>89.69</td></tr><tr><td>L2P [44]</td><td>Prompt (Latent)</td><td>no</td><td>80.99</td><td>86.66</td><td>6.71</td><td>52.08</td><td>55.90</td><td>5.67</td><td>14.64</td><td>21.88</td><td>5.35</td><td>61.98</td><td>71.46</td><td>9.81</td></tr><tr><td>DualPrompt [43]</td><td>Prompt (Latent)</td><td>no</td><td>82.53</td><td>88.06</td><td>4.53</td><td>61.17</td><td>62.73</td><td>3.75</td><td>20.05</td><td>21.57</td><td>1.91</td><td>58.06</td><td>49.95</td><td>6.23</td></tr><tr><td>ConvPrompt [28]</td><td>Prompt Gen.</td><td>no</td><td>84.79</td><td>88.19</td><td>2.38</td><td>70.92</td><td>74.66</td><td>3.39</td><td>-</td><td>-</td><td>-</td><td>70.12</td><td>77.48</td><td>7.21</td></tr><tr><td>MOS [34]</td><td>Adapters</td><td>no</td><td>83.99</td><td>89.53</td><td>11.34</td><td>52.13</td><td>67.97</td><td>35.72</td><td>44.79</td><td>55.14</td><td>15.50</td><td>61.74</td><td>77.03</td><td>25.25</td></tr><tr><td>PROL (Ours)</td><td>Prompt. Gen+SS</td><td>no</td><td>86.32</td><td>91.35</td><td>6.34</td><td>73.50</td><td>78.13</td><td>4.82</td><td>47.72</td><td>58.96</td><td>3.29</td><td>72.51</td><td>79.89</td><td>8.92</td></tr></table>
+
+![](images/f141733a5de79fbdfef2d833a4df7a2a599dd79329d2434e7023479391dae467.jpg)
+
+![](images/2865be8c9f8dc4a16c73ff27f865cdbb0ea927fa025e37810ed72a950ef6a17d.jpg)
+
+![](images/5cb85924c451de2a2961e3b9f2bc0cabf5256d5d4a769059ff20ba2a0d7bda6c.jpg)
+
+![](images/a56457c0a51107d2238ff4575ffab94a139a831aba019f86824f1a2fc3690e28.jpg)
+
+Figure 3. Historical performance plot of the consolidated methods in CIFAR100, Magnet-R, Magnet-A, and CUB200 dataset  
+![](images/eaa1d16473819804eb94ab55b7c2f978adbbf35a5bea91913eeabf50d019555b.jpg)  
+RanPAC-J RanDumb-J RanPAC-R RanDumb-R RanPAC RanDumb L2P DualPrompt ConvPrompt MOS PROL
+
+![](images/faa8e50c8978fcc0ffa2fd984eab09f741e47cba0fc72e1d4f322cb0c5efb1ca.jpg)
+
+![](images/1b4484e226b373c9a6eabf038c7c2a29a764b845fdaafc3bc6d166fce2b73b17.jpg)
+
+![](images/6a3b5b9b5f8dd722dd7fbaaa9fb06234fe37ac0963cf656ca46402c0b677340a.jpg)
+
+Table 2. Comparison of proposed method to non-PTM method on CIFAR100 dataset, the result is copied from ERSM[39]  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Mem</td><td colspan="3">Backbone</td><td rowspan="2">FAA on CIFAR100</td></tr><tr><td>Type</td><td>Size</td><td>PTM</td></tr><tr><td>ER [27]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>38.70</td></tr><tr><td>DER++ [3]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>37.62</td></tr><tr><td>ERACE [4]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>39.82</td></tr><tr><td>OCM [9]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>42.01</td></tr><tr><td>GSA [10]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>42.27</td></tr><tr><td>OnPro [45]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>41.47</td></tr><tr><td>ERSM [39]</td><td>5000</td><td>ResNet-18</td><td>11.2M</td><td>no</td><td>47.72</td></tr><tr><td>PROL (Ours)</td><td>-</td><td>ViT-Tiny</td><td>5.6M</td><td>yes</td><td>64.59</td></tr></table>
+
+methods are significantly outperformed by PROL in the other 3 datasets. In terms of the forgetting measure, PROL consistently achieves a fairly low forgetting rate along with the prompt-based methods. The plot shows that our method is consistently in the bottom group while RanPAC, RanDumb, and MOS are in the middle and upper groups.
+
+c) Parameter, Throughput and Execution time: Table 3 presents the parameter and summarized throughput and execution time while the detailed version is presented by fig.4. The throughput is computed in the training phase, where the
+
+learning process is conducted. The table shows that PROL requires one of the smallest number of trainable parameters, i.e., only slightly bigger than RanPAC and L2P and far smaller than the rest of the methods. RanDumb saves the biggest (> 7M) RBF kernel parameters, while ConvPrompt and MOS require 1.4M and 3.2M parameters, respectively. The table show that PROL has a moderate throughput, i.e., comparable with L2P and DualPrompt, lower than RanDumb and RanPAC. In line with this, PROL achieves moderate training and inference time. RanDumb and RanPAC have higher throughput and lower execution time since the methods update their feature projection and classifier only once, not iteratively per data stream. In the streaming process, it only extracts the input feature, stores it in memory, and then updates the projection and classifier only once after the current task is finished. The weakness of this mechanism is that it can not handle any time inference where a request for inference comes in the middle of the streaming. Figure 4 confirms that PROL has moderate throughput, training time, and testing time from the first task until
+
+Table 3. Trainable Parameter (including classifier), summarized throughput and execution time of consolidated methods in, ImageNet-R dataset.  
+
+<table><tr><td>Method</td><td>#Trainable 
+Parameter 
+In million (M)</td><td>Average 
+Throughput 
+(sample/s)</td><td>Average 
+Training 
+Time (s)</td><td>Average 
+Inference 
+Time (s)</td></tr><tr><td>RanPAC-J [22]</td><td>0.154</td><td>82.92</td><td>29.60</td><td>55.24</td></tr><tr><td>RanDumb-J [25]</td><td>7.834</td><td>77.03</td><td>32.00</td><td>53.79</td></tr><tr><td>RanPAC-R [22]</td><td>0.154</td><td>89.46</td><td>28.30</td><td>84.20</td></tr><tr><td>RanDumb-R [25]</td><td>7.834</td><td>116.89</td><td>21.20</td><td>43.36</td></tr><tr><td>RanPAC [22]</td><td>0.154</td><td>86.83</td><td>27.70</td><td>45.21</td></tr><tr><td>RanDumb [25]</td><td>7.834</td><td>85.90</td><td>29.90</td><td>46.81</td></tr><tr><td>L2P [44]</td><td>0.200</td><td>40.94</td><td>59.80</td><td>54.50</td></tr><tr><td>DualPrompt [43]</td><td>0.701</td><td>46.50</td><td>53.20</td><td>52.60</td></tr><tr><td>ConvPrompt [28]</td><td>1.434</td><td>11.77</td><td>235.70</td><td>103.92</td></tr><tr><td>MOS [34]</td><td>3.197</td><td>16.06</td><td>165.70</td><td>322.20</td></tr><tr><td>PROL (Ours)</td><td>0.213</td><td>35.42</td><td>68.70</td><td>55.39</td></tr></table>
+
+the last task. In terms of throughput, it competes with DualPrompt and L2P and is far better than ConvPrompt and MOS. ConvPrompt and MOS low throughput and high execution time confirm our preliminary analysis on growing prompt components/adapters.
+
+![](images/89dce0af13be97e66954199ece8a4fba0ee6854853f3641499c508249bd0830e.jpg)  
+Figure 4. Detailed throughput and execution time of consolidated methods in, Magnet-R dataset.
+
+d) Robustness Analysis: Table 4 shows the robustness of our method in comparison to the closest competitor, i.e., ConvPrompt, on different prompt lengths and the number of ViT layers pre-pended. The table shows that PROL successfully maintains its performance in different prompt lengths and numbers of layers, even in the smaller prompt length or fewer layers. In addition, it maintains the performance gap to the ConvPrompt as the closest competitor with the convincing margin, i.e.  $\geq 2\%$ . The table shows that PROL achieves a higher CAA in a higher number of layers. The table also shows that PROL has a better performance stability than ConvPrompt with  $2\%$  performance swing.  
+e) Ablation Study: We conducted an ablation study to evaluate the contribution of each component of our method as presented in table 5. The table shows that fine-tuning (FT) alone can not handle OCL problem since the performance is too low. However, combining both inter-task loss  $\mathcal{L}_{inter}$  and intra-task loss  $\mathcal{L}_{intra}$  has  $1 - 2\%$  better accuracy than utilizing only one. The generator (G) plays the most important role as it tremendously boosting the accuracy and reduce the forgetting i.e. up to  $57\%$ . The learnable scalar-and-shifetrs (SS) and class-wise keys (K) and input-to-key similarity
+
+Table 4. Robustness analysis of the proposed method compared to the competitor methods in ImageNet-R dataset.  
+
+<table><tr><td colspan="10">Robustness on Different Prompt Length</td></tr><tr><td rowspan="2">Method</td><td>3</td><td>5</td><td>7</td><td>10</td><td>15</td><td></td><td></td><td></td><td></td></tr><tr><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>CAA</td></tr><tr><td>ConvPrompt</td><td>69.88</td><td>72.89</td><td>70.92</td><td>74.66</td><td>70.68</td><td>74.11</td><td>70.02</td><td>74.18</td><td>69.45</td></tr><tr><td>PROL</td><td>73.42</td><td>77.71</td><td>73.50</td><td>78.13</td><td>73.43</td><td>77.92</td><td>73.45</td><td>78.03</td><td>73.30</td></tr><tr><td colspan="10">Robustness on Different Number of ViT Layers</td></tr><tr><td rowspan="2">Method</td><td>3</td><td>5</td><td>7</td><td>9</td><td>-</td><td></td><td></td><td></td><td></td></tr><tr><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>FAA</td><td>CAA</td><td>-</td></tr><tr><td>ConvPrompt</td><td>71.43</td><td>75.95</td><td>70.92</td><td>74.66</td><td>69.38</td><td>74.21</td><td>72.12</td><td>77.47</td><td>-</td></tr><tr><td>PROL</td><td>73.72</td><td>77.80</td><td>73.50</td><td>78.13</td><td>73.55</td><td>78.14</td><td>73.75</td><td>78.69</td><td>-</td></tr></table>
+
+Table 5. Ablation study of our method in ImageNet-R dataset.  
+
+<table><tr><td>Component</td><td>Loss</td><td>FAA</td><td>CAA</td><td>FM</td></tr><tr><td>FT</td><td>Linter</td><td>11.98</td><td>26.25</td><td>75.72</td></tr><tr><td>FT</td><td>Linter+Lintra</td><td>12.68</td><td>28.13</td><td>81.15</td></tr><tr><td>FT+G</td><td>Linter+Lintra</td><td>69.77</td><td>74.43</td><td>3.25</td></tr><tr><td>FT+G+SS+K</td><td>Linter+Lintra+Lsim</td><td>70.82</td><td>75.83</td><td>3.96</td></tr><tr><td>FT+G+SS+K</td><td>Linter+Lintra+Lsim+Lort</td><td>71.13</td><td>76.09</td><td>4.31</td></tr><tr><td>FT+G+SS+K+HSU</td><td>Linter+Lintra+Lsim+Lort</td><td>72.53</td><td>77.24</td><td>4.95</td></tr><tr><td>FT+G+SS+K+HSU+M</td><td>Linter+Lintra+Lsim+Lort+Lgen</td><td>73.50</td><td>78.13</td><td>4.82</td></tr></table>
+
+loss  $\mathcal{L}_{sim}$  improve the performance with  $1\%$  margin. At this point, our method is comparable to ConvPrompt as the closest competitor and even achieves higher CAA. Adding orthogonality loss  $\mathcal{L}_{ort}$  on K improves both FAA and FAA even though with a small margin. Applying the hard-soft-update strategy improves the model accuracy with  $1.5\%$  FAA and  $1.2\%$  CAA improvement, respectively. However, it increases the forgetting along with the accuracy. Finally, preserving the PTM generalization property by matrix M and generalization loss  $\mathcal{L}_{gen}$  improves the accuracy with  $1\%$  margin, and reduces a slight magnitude of forgetting.
+
+# 6. Concluding Remark
+
+We highlight the rehearsal-free aspect of continual learning in streaming data and propose a novel rehearsal-free method that incorporates four key concepts (1) Single lightweight prompt generator for all instances in all classes (2) learnable scalers-and-shifters parameters associated with learnable class-wise keys (3) cross-correlation matrix for preserving PTM generalization (4) hard-soft update for adaptive learning rate assignment. Our experiment and analysis show the superiority of our method to existing SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset with significant margins, i.e.,  $2 - 76\%$  and  $2 - 64\%$  margin for FAA and CAA, respectively. Our method significantly outperforms the joint version of existing SOTAs in 3 datasets while running in a rehearsal-free manner. Our historical analysis shows the consistency of our method in maintaining stability-plasticity in every task. Our method achieves moderate training time, and throughput. Our method achieves significantly higher throughput and lower training and testing time than the growing prompt generator and adapters methods. Our robustness analysis shows that our method consistently achieves the highest performance in various prompt lengths and numbers of ViT layers.
+
+# Acknowledgement
+
+M. Anwar Ma'sum acknowledges the support of Tokopedia-UI Centre of Excellence for GPU access to run the experiments.
+
+# References
+
+[1] Magdalena Biesialska, Katarzyna Biesialska, and Marta R Costa-Jussa. Continual lifelong learning in natural language processing: A survey. arXiv preprint arXiv:2012.09823, 2020. 1  
+[2] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5497-5512, 2022. 2  
+[3] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 6, 7  
+[4] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2104.05025, 2021. 6, 7  
+[5] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Ales Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern analysis and machine intelligence, 44(7):3366-3385, 2021. 1  
+[6] Qiankun Gao, Chen Zhao, Yifan Sun, Teng Xi, Gang Zhang, Bernard Ghanem, and Jian Zhang. A unified continual learning framework with general parameter-efficient tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11483-11493, 2023. 2  
+[7] Zhanxin Gao, Jun Cen, and Xiaobin Chang. Consistent prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28463-28473, 2024. 1, 2  
+[8] Yanan Gu, Xu Yang, Kun Wei, and Cheng Deng. Not just selection, but exploration: Online class-incremental continual learning via dual view consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7442-7451, 2022. 2, 3  
+[9] Yiduo Guo, Bing Liu, and Dongyan Zhao. Online continual learning through mutual information maximization. In International conference on machine learning, pages 8109-8126. PMLR, 2022. 6, 7  
+[10] Yiduo Guo, Bing Liu, and Dongyan Zhao. Dealing with cross-task class discrimination in online continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11878-11887, 2023. 2, 3, 6, 7  
+[11] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. ICCV, 2021. 6
+
+[12] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 6  
+[13] Fushuo Huo, Wenchao Xu, Jingcai Guo, Haozhao Wang, and Yunfeng Fan. Non-exemplar online class-incremental continual learning via dual-prototype self-augment and refinement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12698-12707, 2024. 2  
+[14] Byeonghwi Kim, Minhyuk Seo, and Jonghyun Choi. Online continual learning for interactive instruction following agents. arXiv preprint arXiv:2403.07548, 2024. 2, 3  
+[15] Hyunseo Koh, Dahiyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. arXiv preprint arXiv:2110.10031, 2021. 2  
+[16] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6  
+[17] Muhammad Rifki Kurniawan, Xiang Song, Zhiheng Ma, Yuhang He, Yihong Gong, Yang Qi, and Xing Wei. Evolving parameterized prompt memory for continual learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13301-13309, 2024. 1, 2, 4  
+[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2  
+[19] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 4  
+[20] Yan-Shuo Liang and Wu-Jun Li. Inflora: Interference-free low-rank adaptation for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23638-23647, 2024. 2  
+[21] Hao Liu, Yong Zhou, Bing Liu, Jiaqi Zhao, Rui Yao, and Zhiwen Shao. Incremental learning with neural networks for computer vision: a survey. Artificial intelligence review, 56 (5):4557-4589, 2023. 1  
+[22] Mark D McDonnell, Dong Gong, Amin Parvaneh, Ehsan Abbasnejad, and Anton Van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. Advances in Neural Information Processing Systems, 36:12022-12053, 2023. 1, 2, 3, 6, 7, 8  
+[23] Nicolas Michel, Maorong Wang, Ling Xiao, and Toshihiko Yamasaki. Rethinking momentum knowledge distillation in online continual learning. In Proceedings of the 41st International Conference on Machine Learning. JMLR.org, 2024. 2  
+[24] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 524-540. Springer, 2020. 2  
+[25] Ameya Prabhu, Shiven Sinha, Ponnurangam Kumaraguru, Philip Torr, Ozan Sener, and Puneet Dokania. *Randumb: Random representations outperform online continually learned representations.* Advances in Neural Information Processing Systems, 37:37988-38006, 2024. 1, 2, 3, 6, 7, 8
+
+[26] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 2  
+[27] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 6, 7  
+[28] Anurag Roy, Riddhiman Moulick, Vinay K Verma, Saptarshi Ghosh, and Abir Das. Convolutional prompting meets language models for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23616-23626, 2024. 1, 2, 4, 6, 7, 8  
+[29] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 2  
+[30] Minhyuk Seo, Hyunseo Koh, and Jonghyun Choi. Budgeted online continual learning by adaptive layer freezing and frequency-based sampling. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 3  
+[31] Khadija Shaheen, Muhammad Abdullah Hanif, Osman Hasan, and Muhammad Shafique. Continual learning for real-world autonomous systems: Algorithms, challenges and frameworks. Journal of Intelligent & Robotic Systems, 105 (1):9, 2022. 1  
+[32] James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora. arXiv preprint arXiv:2304.06027, 2023. 2  
+[33] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 4  
+[34] Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, De-Chuan Zhan, and Han-Jia Ye. Mos: Model surgery for pre-trained model-based class-incremental learning. arXiv preprint arXiv:2412.09441, 2024. 2, 3, 6, 7, 8  
+[35] Zonggui Tian, Du Zhang, and Hong-Ning Dai. Continual learning on graphs: A survey. arXiv preprint arXiv:2402.06330, 2024.1  
+[36] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6  
+[37] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 1, 2  
+[38] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method
+
+and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1  
+[39] Maorong Wang, Nicolas Michel, Jiafeng Mao, and Toshihiko Yamasaki. Dealing with synthetic data contamination in online continual learning. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 3, 6, 7  
+[40] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23460-23469, 2024. 2  
+[41] Maorong Wang, Nicolas Michel, Ling Xiao, and Toshihiko Yamasaki. Improving plasticity in online continual learning via collaborative learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23460-23469, 2024. 2  
+[42] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. Advances in Neural Information Processing Systems, 35:5682-5695, 2022. 1, 2  
+[43] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 3, 4, 6, 7, 8  
+[44] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 1, 2, 3, 6, 7, 8  
+[45] Yujie Wei, Jiaxin Ye, Zhizhong Huang, Junping Zhang, and Hongming Shan. Online prototype learning for online continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18764-18774, 2023. 2, 6, 7  
+[46] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 374-382, 2019. 2  
+[47] Yichen Wu, Hong Wang, Peilin Zhao, Yefeng Zheng, Ying Wei, and Long-Kai Huang. Mitigating catastrophic forgetting in online continual learning by modeling previous task interrelations via pareto optimization. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3  
+[48] Wang Xinrui, Chuanxing Geng, Wenhai Wan, Shao-Yuan Li, and Songcan Chen. Forgetting, ignorance or myopia: Revisiting key challenges in online continual learning. Advances in Neural Information Processing Systems, 37:58341-58375, 2024. 1, 2, 3  
+[49] Jason Yoo, Yunpeng Liu, Frank Wood, and Geoff Pleiss. Layerwise proximal replay: a proximal point method for online continual learning. In Proceedings of the 41st International Conference on Machine Learning. JMLR.org, 2024. 1, 2, 3
+
+[50] Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca: Slow learner with classifier alignment for continual learning on a pre-trained model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19148-19158, 2023. 2  
+[51] Linglan Zhao, Xuerui Zhang, Ke Yan, Shouhong Ding, and Weiran Huang. Safe: Slow and fast parameter-efficient tuning for continual learning with pre-trained models. Advances in Neural Information Processing Systems, 37: 113772-113796, 2025. 2, 5  
+[52] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan
+
+Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554-23564, 2024. 2, 3, 6  
+[53] Huiping Zhuang, Yuchen Liu, Run He, Kai Tong, Ziqian Zeng, Cen Chen, Yi Wang, and Lap-Pui Chau. F-oial: Forward-only online analytic learning with fast training and low memory footprint in class incremental learning. Advances in Neural Information Processing Systems, 37: 41517–41538, 2024. 3

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffa9484b2ed35c8275e557d42ee98feb3d321f9bf6c50ba14168b4f6611468d7
+size 695885

ICCV/2025/PROL _ Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a08a9eff694c145287a0b29ec43fa3ce004d164349d1a65e82f0616aaeb4b00a
+size 475688

ICCV/2025/PRVQL_ Progressive Knowledge-guided Refinement for Robust Egocentric Visual Query Localization/eb4759bd-cd46-42ae-a79f-87abd8db733e_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9113e87836edc13e8be4ffa10d2b988e22de5b593c551adbae5e3b3355629e89
+size 85544

ICCV/2025/PRVQL_ Progressive Knowledge-guided Refinement for Robust Egocentric Visual Query Localization/eb4759bd-cd46-42ae-a79f-87abd8db733e_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dbebf50f18e733551be861cb4734f63c1c71e5475d743783e45afd26fbd37cd
+size 101951