Add Batch 0812bccf-3436-4c24-bb5a-f58edb5496be

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/2318884c-d71a-4947-ab84-f1c482a132fd_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6aa9b1ff49ee11ca471706aeedcd27e02758e89e5b0f22ad6b19959ac8ceab8c
+size 73111

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/2318884c-d71a-4947-ab84-f1c482a132fd_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25f1f912140c1117714b16327c24c116f22f00db3be0c3973225c3bf705edb45
+size 90536

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/2318884c-d71a-4947-ab84-f1c482a132fd_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43fdebc53fd8476f7775d707e3ce74df79421334ad5509c20bc05628e0d7a3e8
+size 720261

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/full.md

@@ -0,0 +1,289 @@
+# Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception
+
+Baixuan Lv $^{1*}$  Yaohua Zha $^{1,2*}$  Tao Dai $^{3\dagger}$  Xue Yuerong $^{1}$  Ke Chen $^{2}$  Shu-Tao Xia $^{1,2}$  Tsinghua University Pengcheng Laboratory Shenzhen University
+
+# Abstract
+
+Point cloud video understanding is becoming increasingly important in fields such as robotics, autonomous driving, and augmented reality, as they can accurately represent object motion and environmental changes. Despite the progress made in self-supervised learning methods for point cloud video understanding, the limited availability of 4D data and the high computational cost of training 4D-specific models remain significant obstacles. In this paper, we investigate the potential of transferring pre-trained static 3D point cloud models to the 4D domain, pointing out the limitations of static models that capture only spatial information while neglecting temporal dynamics. To address this, we propose a novel Cross-frame Spatio-temporal Adaptation (CSA) strategy by introducing the Point Tube Adapter as the embedding layer and the Geometric Constraint Temporal Adapter (GCTA) to enforce temporal consistency across frames. This strategy extracts both short-term and long-term temporal dynamics, effectively integrating them with spatial features and enriching the model's understanding of temporal changes in point cloud videos. Extensive experiments on 3D action and gesture recognition tasks demonstrate that our method achieves state-of-the-art performance, establishing its effectiveness for point cloud video understanding. Code is available at: https://github.com/LvBaixuan/Point-CSA.
+
+# 1. Introduction
+
+Four-dimensional (4D) point cloud videos offer an accurate representation of object motion, making them increasingly valuable in fields such as robotics, autonomous driving, and augmented reality. A comprehensive understanding of point cloud videos is crucial for effectively perceiving environmental changes and facilitating interactions with the real world. Consequently, the learning-based point cloud video understanding [9-11, 20, 33] has attracted significant re
+
+![](images/020e3c8aa6140894301c64f722b0043a4156f722663e8aacb25a4b6d62b7178b.jpg)  
+Figure 1. Comparison of direct transfer of static point cloud pretraining model and point cloud video pre-training model on MSR-Action dataset. Blue bars represent the performance of 4D-specific models, yellow bars indicate the results from directly transferring static point cloud pre-trained models, and orange bars depict the performance after applying our Adaptation Method.
+
+search interest in recent years.
+
+Self-supervised representation learning has become the most popular approach for current point cloud video understanding, as it learns general representations using unlabeled 4D data. Although various self-supervised methods [25-28, 46] have achieved some success in point cloud video modeling, the difficulty in collecting 4D data results in relatively small pre-training datasets, which limits the ability to learn general representations. Furthermore, the introduction of the temporal dimension in point cloud videos increases the time and computational resources required during the pre-training phase, thereby raising the costs of training 4D-specific models.
+
+Pre-trained models [4, 8, 23, 37-39, 42, 44] for 3D static point clouds have seen significant success within the community and have gained widespread adoption. Static and dynamic point clouds exhibit spatial consistency, raising the question of whether existing pre-trained models for static point clouds can be effectively leveraged for point cloud videos, thereby reducing the need for additional training of 4D-specific models. To explore this, we initially di
+
+rectly transfer existing popular pre-trained models for static point clouds to the 4D domain. However, as shown in Figure 1, the direct transfer of static point cloud pre-trained models results in significantly inferior performance compared to existing pre-training methods [25, 26, 28] specifically designed for point cloud videos. This performance gap is primarily due to the static models capturing only spatial information while neglecting critical temporal dynamics. These observations have motivated us to design an adaptation strategy that incorporates temporal awareness.
+
+The straightforward idea is to incorporate temporal information into static point clouds. Inspired by previous work [9, 11], we introduce the Point Tube Adapter as our embedding layer, replacing the embedding layer in static point cloud pre-trained models. As shown in Table 3, although this modification significantly enhances the model's ability to understand point cloud videos, there remains a gap compared to dedicated 4D models. We believe this is due to the self-attention mechanism in the frozen backbone network being unable to dynamically capture the long-range spatio-temporal correlations between patches. To this end, we propose the Geometric Constraint Temporal Adapter (GCTA), which introduces geometric constraints on neighboring points to ensure the consistency of point cloud structures across different time frames during interaction, capturing cross-temporal geometric information. This strategy enables the model to perceive structural changes in spatial regions throughout the entire video and integrates global dynamic information into the pre-trained encoder, enriching the model's understanding of temporal dynamics.
+
+The Cross-frame Spatio-temporal Adaptation (CSA) strategy not only allows the model to capture short-term structure changes but also emphasizes long-term dynamics, significantly enriching its understanding of temporal information. We evaluated our model on 3D action recognition and gesture recognition tasks. Experiments on the MSR-Action3D [18] and SHREC'17 [7] datasets demonstrate the effectiveness of our approach. The contributions of this paper include:
+
+- We explore methods for transferring spatial semantic information from static point clouds to point cloud videos. To our knowledge, this is the first exploration of applying static point cloud pre-trained models in point cloud video tasks.  
+- We propose a novel Cross-frame Spatio-temporal Adaptation (CSA) strategy by introducing the Point Tube Adapter as the embedding layer and the Geometric Constraint Temporal Adapter (GCTA) to enforce temporal consistency across frames. This strategy extracts both short-term and long-term temporal dynamics, effectively integrating them with spatial features and enriching the model's understanding of temporal changes in point cloud videos.
+
+- Extensive experiments on 3D action and gesture recognition tasks demonstrate that our method achieves state-of-the-art performance, highlighting its efficacy in point cloud video understanding.
+
+# 2. Related Work
+
+# 2.1. Static Point Cloud Pre-trained Model
+
+Recent research on pre-training methods for 3D point cloud analysis has made significant progress. These methods typically involve learning latent semantic information from point clouds through pretext tasks, followed by fine-tuning the model for specific applications to enhance performance. Existing pretext tasks can be categorized into two types: discriminative tasks [2, 5, 12, 34, 45], and generative tasks [13, 19, 30, 36, 40, 41]. Discriminative methods enable the model to distinguish similar point clouds by contrasting various views or instances. For example, techniques such as PointContrast [35] and CrossPoint [1] leverage both in-domain and cross-domain features. Generative methods learn the latent features of the data relying on autoencoders to reconstruct the original input. For example, PointBERT [37] generates discrete point tokens that contain meaningful local information derived from masked tokens. Point-MAE [23] employs an autoencoder to learn high-level latent features from unmasked patches, with the aim of reconstructing the masked point patches. ACT [8] significantly improves the performance of 3D self-supervised networks on real scanned point clouds by combining pretrained language and image models. PointGPT [4] introduces an autoregressive generation task for point clouds to pre-train Transformer models, offering high scalability.
+
+However, these pre-trained models have primarily been fine-tuned for downstream tasks associated with static point clouds. Our research aims to extend their applicability by transferring these models to the domain of point cloud video understanding.
+
+# 2.2. Point Cloud Video Understanding
+
+Point cloud videos exhibit a dual nature, being unordered in the spatial dimension while ordered in the temporal dimension. This characteristic necessitates the design of specialized methods to aggregate spatio-temporal information to perceiving both geometry and dynamics. Current approaches in this domain can be broadly classified into voxel-based [15, 43] and point-based [9-11, 20, 21, 33] techniques. Voxel-based methods, such as MinkowskiNet [6], involve voxelizing the original point clouds and then employing 4D convolution to extract features from 4D voxels. However, this voxelization process inherently results in quantization loss. In contrast, point-based methods operate directly on the raw point data. MeteorNet [20] introduces the temporal dimension and tracks motion at the point level,
+
+![](images/df570255063dd93aac7158cbde2603dc98fd2241d04947a77f32406c81c605c6.jpg)  
+Figure 2. Illustration of the proposed Cross-frame Spatio-temporal Adaptation strategy. We divide point cloud videos into point tubes. Aggregate points within the point tube to obtain local spatio-temporal embedding, and construct a cross-frame KNN graph based on the coordinates of the anchor point of the point tubes. Our proposed Geometric Constraint Temporal Adapter (GCTA) aggregates long-term changes based on the cross-frame KNN graph. During fine-tuning, the parameters of Attention and Feed Forward Network(FFN) of the pre-trained model are frozen.
+
+while PSTNet [11] constructs point tubes along the time axis to facilitate 4D point convolution. Recent advances, including P4Transformer [9], PPTr [33], and LeaF [21], have integrated transformer architectures to expand the receptive field. PST-Transformer [10] builds on P4Transformer, enhancing its ability to model spatio-temporal structures through spatial displacement encoding. To effectively learn the inherent information in point cloud videos, rather than focusing solely on downstream tasks, several methods employ self-supervised learning to enhance generalization capabilities. One approach [31] pretrains an encoder by predicting the temporal order of shuffled segments. C2P [46] introduces a complete-to-part 4D distillation method to predict representations of point cloud frames over short time windows. Additionally, several methods [26-28] utilize spatial and inter-frame contrastive learning to extract inherent geometric and dynamic features. However, these techniques often rely on clip-level or frame-level pretext tasks to learn motion information, which may be limited by inconsistencies in points across frames. MaST-Pre [25] proposes a masked prediction approach that segments continuous point clouds into point tubes for reconstruction. While this method effectively encodes relative spatio-temporal information within the point tubes, it falls short in capturing global semantics.
+
+Our approach preserves the model's strong encoding
+
+capability for spatial structures learned from static point clouds while also emphasizing the capture of short-term morphological changes and long-term dynamics.
+
+# 3. Method
+
+# 3.1. Overview
+
+The general framework of our method is shown in Figure 2. To enable the static point cloud pre-trained model to perceive dynamic information within point cloud videos, we introduce the Cross-frame Spatio-temporal Adaptation (CSA) strategy, which allows the model to comprehend point cloud videos through both short-term and long-term spatial structural changes. Initially, we incorporate point tubes during the feature embedding phase. This grouping strategy combines local points along both spatial and temporal dimensions. By integrating relative spatio-temporal information, the model effectively captures temporal dynamics along with spatial structures. Subsequently, to enhance attention to the spatial structural changes of point clouds across the entire video, we propose a Geometric Constraint Temporal Adapter(GCTA). This module facilitates feature aggregation over a global temporal range, allowing each point tube to traverse time and identify relevant features from other point tubes throughout the video. By aggregating spatial information from various frames, we in
+
+fuse global dynamic knowledge into the pre-trained model.
+
+Through this approach, we successfully adapt the static point cloud pre-trained model for point cloud video understanding tasks while keeping its parameters frozen.
+
+# 3.2. Point Tube Adapter
+
+Some widely used static point cloud pretraining methods process input point clouds using farthest point sampling (FPS) and K-Nearest Neighbors (KNN) algorithms to obtain center points and point patches. This encoding technique effectively aggregates information from local spatial neighborhoods. However, point cloud video data exhibit dynamic characteristics, necessitating a neighborhood partitioning method that can extract spatial structures while concurrently encoding temporal information.
+
+Building on previous work [9, 11], we introduce the concept of point tubes. Point cloud videos can be represented as a series of 3D point coordinates  $P \in \mathbb{R}^{L \times N \times 3}$ , where  $L$  is the length of the video and  $N$  is the number of points per frame. First, we uniformly select anchor frames within the point cloud video, obtaining anchor points using FPS within these frames and extending them to adjacent frames. Next, we search for neighbors within a spatial radius  $r_s$  around the anchor points in both the anchor frame and its adjacent frames, forming a point tube, denoted as  $T \in \mathbb{R}^{l \times k \times 3}$ , where  $l$  is the temporal length of the point tube and  $k$  is the number of points per frame in the point tube.
+
+Similarly to the embedding methods for static point clouds, we explicitly calculate the relative distances between the points in the tube and the anchor points, which aids the model in capturing local spatial structures. In addition, we incorporate relative temporal information to encode short-term local structural dynamics. The obtained spatio-temporal displacements  $(\delta_x,\delta_y,\delta_z,\delta_t)$  serve as point-level features  $F_{p}$ . The embedding of point tubes is accomplished by an extended lightweight 4D PointNet [24], which mainly consists of feature mapping functions and pooling layers. This embedding can be represented as:
+
+$$
+T _ {e} = \operatorname {m a x p o o l} [ \psi (F _ {p}) ], \tag {1}
+$$
+
+where  $\psi : \mathbb{R}^{lk \times 4} \longrightarrow \mathbb{R}^{lk \times C}$  aims to increase the dimensionality of point features to enhance feature representation capabilities. We directly input the embeddings  $T_{e}$  into the pre-trained encoder.
+
+# 3.3. Geometric Constraint Temporal Adapter
+
+The classical Transformer [29] relies on the self-attention mechanism to perceive long-range correlations among all patches globally and has achieved great success in static 3D domains. However, simply replacing point patch embeddings with point tube embeddings still makes it challenging to adapt static point cloud pre-trained models for point cloud video. The temporal scope of point tubes remains
+
+![](images/7fdd5372e234196414af714833976bcd8555f37cfc1f07611b43f1ac5326c4e0.jpg)  
+(a) Cross-frame KNN.
+
+![](images/98c7ba51ebd397422148e5d5f32eb3d8976c353c76698f62f0d15a6c341cdc68.jpg)  
+(b) Aggregation of point tube features through cross-frame KNN Graph.  
+Figure 3. Illustration of Geometric Constraint Temporal Adapter.
+
+limited by their length, capturing only short-term changes. Moreover, static point clouds only contain spatial structural information, which results in the frozen parameter static point cloud pre-trained models having a limited ability to perceive temporal relationships between point tubes.
+
+To address this issue, we propose Geometric Constraint Temporal Adapter (GCTA) to capture long-term dynamic information from point cloud videos. The structure of GCTA is shown in Figure 3. Inspired by EdgeConv [32], We construct a directed cross-frame KNN graph  $G = (T, E)$  based on the spatial distance between all point tubes, allowing each point tube to explore more relevant point tubes across the entire video without being constrained by time. Consider a point cloud video divided into  $N'$  point tubes with features denoted as  $F = \{f_1, f_2, \dots, f_{N'}\}$ , we define the edge feature between the  $i$ -th and the  $j$ -th point tube as:
+
+$$
+e _ {i j} = \phi (\operatorname {c o n c a t} \left(f _ {j} - f _ {i}, f _ {i}\right)), \tag {2}
+$$
+
+where  $\phi : \mathbb{R}^{2C} \longrightarrow \mathbb{R}^{C'}$  is a function parameterized by a set of learnable parameters. We use a symmetric operation, max pooling, to aggregate the edge features and project them into feature dimension  $C$  using  $\xi : \mathbb{R}^{C'} \longrightarrow \mathbb{R}^C$ :
+
+$$
+f _ {i} ^ {\prime} = \xi (m a x p o o l (e _ {i j})). \tag {3}
+$$
+
+GCTA explicitly combines global variation, captured by  $f_{j} - f_{i}$ , with local structural features, captured by the point tube embedding  $f_{i}$ . By aggregating point tubes distributed across different timestamps, GCTA can learn the global dynamics of spatial structures, thereby enhancing long-term
+
+<table><tr><td>Methods</td><td>Reference</td><td>8 Frames</td><td>12 Frames</td><td>16 Frames</td><td>24 Frames</td></tr><tr><td colspan="6">Supervised Learning</td></tr><tr><td>MeteorNet[20]</td><td>ICCV2019</td><td>81.14</td><td>86.53</td><td>88.21</td><td>88.50</td></tr><tr><td>PSTNet[11]</td><td>ICLR2021</td><td>83.50</td><td>87.88</td><td>89.90</td><td>91.20</td></tr><tr><td>P4Transformer[9]</td><td>CVPR2021</td><td>83.17</td><td>87.54</td><td>89.56</td><td>90.94</td></tr><tr><td>Kinet[47]</td><td>CVPR2022</td><td>83.84</td><td>88.53</td><td>91.92</td><td>93.27</td></tr><tr><td>PPTr[33]</td><td>ECCV2022</td><td>84.02</td><td>89.89</td><td>90.31</td><td>92.33</td></tr><tr><td>LeaF[21]</td><td>ICCV2023</td><td>84.50</td><td>-</td><td>91.50</td><td>93.84</td></tr><tr><td>PST-Transformer[10]</td><td>TPAMI2023</td><td>83.97</td><td>88.15</td><td>91.98</td><td>93.73</td></tr><tr><td>X4D-SceneFormer[17]</td><td>AAAI 2024</td><td>86.47</td><td>-</td><td>92.56</td><td>93.90</td></tr><tr><td colspan="6">Self-Supervised Learning (End-to-End Fine-Tuning)</td></tr><tr><td>CPR[27]</td><td>AAAI2023</td><td>86.53</td><td>91.00</td><td>92.15</td><td>93.03</td></tr><tr><td>C2P[46]</td><td>CVPR2023</td><td>87.16</td><td>-</td><td>91.89</td><td>94.76</td></tr><tr><td>PointCMP[26]</td><td>CVPR2023</td><td>89.56</td><td>91.58</td><td>92.26</td><td>93.27</td></tr><tr><td>PointCPSC[28]</td><td>ICCV2023</td><td>88.89</td><td>90.24</td><td>92.26</td><td>92.68</td></tr><tr><td>MaST-Pre[25]</td><td>ICCV2023</td><td>-</td><td>-</td><td>-</td><td>94.08</td></tr><tr><td colspan="6">Adaptation for Pre-Trained Models</td></tr><tr><td>Point-BERT+CSA</td><td>-</td><td>90.72</td><td>92.39</td><td>93.38</td><td>94.77</td></tr><tr><td>Point-MAE+CSA</td><td>-</td><td>91.41</td><td>92.04</td><td>93.73</td><td>95.12</td></tr><tr><td>PointGPT-S+CSA</td><td>-</td><td>91.41</td><td>93.08</td><td>94.08</td><td>94.42</td></tr></table>
+
+Table 1. Action recognition accuracy (%) on the MSR-Action3D.
+
+temporal perception.
+
+# 3.4. Feature Interaction
+
+In this section, we fuse local and global spatio-temporal information. The Point Tube Adapter provides the model with local spatio-temporal information, while the GCTA learns global dynamic knowledge. The global information related to spatial positions is provided by a customized position encoder (PE) implemented as a simple multilayer perceptron (MLP). The position encoders in different layers maintain the same architecture but have different parameters. The specific architecture is illustrated in the figure. The feature interaction process at the  $i$ -th layer can be represented as:
+
+$$
+\begin{array}{l} \mathcal {F} _ {i + 1} = F F N \left(\text {A t t e n t i o n} \left(\mathcal {F} _ {i}\right)\right) \tag {4} \\ + G C T A (\mathcal {F} _ {i}) + P E (\mathcal {F} _ {i}), \\ \end{array}
+$$
+
+where FFN is a feedforward neural network whose parameters, along with the attention parameters, are pre-trained on static point cloud data and frozen during the adaptation process.
+
+# 4. Experiments
+
+We evaluated the performance of our proposed Cross-frame Spatio-temporal Adapting (CSA) strategy on 3D action recognition and gesture recognition tasks. We also discuss the benefits of our adaptation strategy and architecture in
+
+Sec 4.4. We used three classic pre-trained models, Point-Bert [37], Point-MAE [23], and PointGPT-S [4], as our baseline.
+
+# 4.1. Experiment Setting
+
+We apply our CSA strategy on three baseline models: Point-BERT [37], Point-MAE [23], and PointGPT-S [4]. All three are pre-trained on ShapeNet [3] dataset for 300 epochs.
+
+To ensure a fair comparison, we use the same experimental settings for each baseline. We freeze the weights of the pre-trained static point cloud model and only update the parameters of the Point Tube Embedding module, the GCTA, and the head during downstream task training. We employ the SGD optimizer. The learning rate warms up to 0.01 in the first 10 epochs and decays at a rate of 0.1 in the 20th and 30th epochs. The model is trained for a total of 50 epochs.
+
+# 4.2.3D Action Recognition
+
+We only use the coordinates of the points, without incorporating color information. The point cloud videos are divided into multiple clips with a fixed number of frames as input. During evaluation, we average the prediction probabilities at the clip level to obtain the video-level prediction result.
+
+The experiment is conducted on the MSR-Action3D [18] dataset. MSR-Action3D consists of 567 depth videos, with 20 action categories. Each video in the MSR-Action3D dataset contains an average of 40 frames. Following pre
+
+vious work [9], we partition the dataset into 270 training videos and 297 testing videos. To evaluate the performance of our adaptation strategy under different temporal resolutions, we conduct experiments by extracting 8, 12, 16, and 24 frames, with each frame randomly sampled to 2048 points.
+
+We select an anchor frame every two frames and sample 64 anchor points from it. We set the temporal length of the point tube  $l = 3$ , the neighborhood space radius  $r_s = 0.3$  and randomly sample 32 points in the neighborhood space. The results summarized in Table 1 demonstrate that our adaptation method outperforms both supervised and self-supervised learning approaches across all temporal configurations. These results underscore the potential of transferring knowledge from static point cloud pre-trained models to point cloud video understanding. The robust spatial feature extraction capabilities inherent in these models provide valuable insights that enhance action recognition in dynamic settings. The CSA strategy effectively bridges the gap between static and dynamic representations, ensuring that the model can capitalize on both short-term and long-term temporal dynamics. The consistent improvements across varying input configurations suggest that our approach adapts flexibly to different temporal resolutions. In the experiments with 8, 12, and 16 frames, the application of our CSA to the three baseline models outperformed existing methods, demonstrating the effectiveness of this approach at low temporal resolutions.
+
+Overall, our results not only validate the effectiveness of the proposed adaptation method but also highlight the importance of leveraging existing pre-trained models to improve performance in downstream tasks.
+
+# 4.3. Gesture Recognition
+
+We conduct gesture recognition experiments on the SHREC'17 [7] dataset. SHREC'17 consists of 2800 videos in 28 classes for gesture recognition, of which  $70\%$  (2960 videos) are training data and the other  $30\%$  (840 videos) are the test set. To adapt our models, we freeze the parameters of the pre-trained models and retain the same architecture used in the action recognition tasks from the MSRAction3D dataset.
+
+As shown in Table 2, our method achieved higher recognition accuracy compared to both supervised and self-supervised methods. Notably, pre-trained models using the Cross-Frame Spatio-temporal Adaptation (CSA) strategy achieved impressive accuracies: PointGPT-S+CSA reached  $96.5\%$ , Point-BERT+CSA attained  $96.2\%$ , and Point-MAE+CSA achieved  $95.2\%$ . The high performance across various tasks reinforces the adaptability of our approach. By effectively leveraging the consistency in spatial structures between static point clouds and point cloud videos, our method successfully adapts static pre-trained
+
+models to a variety of downstream tasks. This indicates that the structural insights gained from static point clouds can be effectively transferred to dynamic contexts.
+
+<table><tr><td>Methods</td><td>Reference</td><td>Acc.(%)</td></tr><tr><td colspan="3">Supervised Learning</td></tr><tr><td>PLSTM-base [22]</td><td>CVPR2020</td><td>87.6</td></tr><tr><td>PLSTM-early [22]</td><td>CVPR2020</td><td>93.5</td></tr><tr><td>PLSTM-PSS [22]</td><td>CVPR2020</td><td>93.1</td></tr><tr><td>PLSTM-middle [22]</td><td>CVPR2020</td><td>94.7</td></tr><tr><td>PLSTM-late [22]</td><td>CVPR2020</td><td>93.5</td></tr><tr><td>Kinet [47]</td><td>CVPR2022</td><td>95.2</td></tr><tr><td colspan="3">Self-Supervised Learning (End-to-End Fine-Tuning)</td></tr><tr><td>PointCMP [26]</td><td>CVPR2023</td><td>93.3</td></tr><tr><td>MaST-Pre [25]</td><td>ICCV2023</td><td>92.4</td></tr><tr><td colspan="3">Adaptation for Pre-Trained Models</td></tr><tr><td>Point-BERT+CSA</td><td>-</td><td>96.2</td></tr><tr><td>Point-MAE+CSA</td><td>-</td><td>95.2</td></tr><tr><td>PointGPT-S+CSA</td><td>-</td><td>96.5</td></tr></table>
+
+Table 2. Gesture recognition accuracy  $(\%)$  on the SHREC'17.
+
+# 4.4. Ablation Studies
+
+To analyze the impact of each module in our proposed method on performance, we utilize Point-MAE as the baseline model and conduct ablation experiments on the MSRAction3D dataset.
+
+# 4.4.1. Architecture Design
+
+Our proposed Cross-frame Spatio-temporal Adaptation strategy enhances static point cloud pre-trained models with spatio-temporal information in three key ways: (1) Point Tube Adapter: This component aggregates local spatiotemporal neighborhood information. (2) Position Encoders: These encoders are independently learned for each layer, providing adaptive global spatial context. (3) Geometric Constraint Temporal Adapter(GCTA): This module extends the temporal field of view, enabling the model to perceive global dynamics effectively. we conduct extensive ablation studies, as shown in Table 3.
+
+We first present the performance of model A0 as the baseline (freeze the pre-trained model and only tuned the task head), achieving an accuracy of  $77.35\%$ . Model A1, which integrates Point Tube Adapter, increases the accuracy by  $6.62\%$ , highlighting the benefit of local spatio-temporal aggregation. Adding Position Encodings in model A2 further enhances accuracy to  $87.11\%$ , emphasizing the importance of adaptive global spatial context. Model A3, which combines Point Tube Adapter with the GCTA, achieves an accuracy of  $94.08\%$ , demonstrating the value of extend
+
+<table><tr><td></td><td>PTA</td><td>PE</td><td>GCTA</td><td>Acc.(%)</td></tr><tr><td>A0</td><td>×</td><td>×</td><td>×</td><td>77.35</td></tr><tr><td>A1</td><td>✓</td><td>×</td><td>×</td><td>83.97</td></tr><tr><td>A2</td><td>✓</td><td>✓</td><td>×</td><td>87.11</td></tr><tr><td>A3</td><td>✓</td><td>×</td><td>✓</td><td>94.08</td></tr><tr><td>A4(ours)</td><td>✓</td><td>✓</td><td>✓</td><td>95.12</td></tr></table>
+
+Table 3. Ablation studies on architecture designs. PTA denotes Point Tube Adapter, and PE denotes additional Position Encoders.
+
+ing the temporal field of view. Finally, our full model A4, incorporating all components, reaches an accuracy of  $95.12\%$ , confirming the effectiveness of our Cross-frame Spatio-temporal Adaptation strategy in leveraging both local and global information to improve recognition performance in dynamic point cloud scenarios.
+
+We visualize the attention distribution of the model with and without the GCTA. As shown in the Figure 4, for the same input video, the model adding the GCTA focuses more on the key frames where the action occurs, and allocates less attention to the frames that are weakly related to the motion (the first two frames). This demonstrates that the GCTA helps the model understand long-term motion.
+
+![](images/9e99cca72102df5355d0d027c0d65855ab8a1bcbc4a20c6ea8e9e0049b837e36.jpg)  
+(a) Without Geometric Constraint Temporal Adapter
+
+![](images/c1c19478334db726035c77289e3a0e6c15b9d66ff16b363492dad233f6aa938a.jpg)  
+(b) With Geometric Constraint Temporal Adapter  
+Figure 4. Visualization of the attention of model with/without the Geometric Constraint Temporal Adapter. Warmer color indicates higher attention. The model with the Geometric Constraint Temporal Adapter can correctly identify important frames and moving areas.
+
+# 4.4.2. Advantages of CSA compared to Other Tuning strategies
+
+To demonstrate the effectiveness of our Cross-frame Spatiotemporal Adaptation strategy, we conduct comprehensive comparative experiments against various tuning strategies while keeping the parameters of the pre-trained model frozen. As shown in the Figure 5, we compare our method with the vanilla adapter [14], VPT-Deep [16], and IDPT [39]. The baseline is consistent with that described in Sec 4.4.1. Our CSA outperforms other tuning strategies in terms of performance. We fully consider the spatially disordered and temporally ordered characteristics of point cloud
+
+![](images/dae291fa3b63c04987fb1f7e71cdbcb8eed7e92c3ad7c6cbc9776afe7df7f29b.jpg)  
+Figure 5. Comparison of different tuning strategies.
+
+videos, and enhance the ability of models to capture dynamics by combining spatio-temporal information. This focus on dynamics is crucial for point cloud video tasks such as action recognition and gesture detection, where temporal changes play a significant role.
+
+# 4.4.3. Receptive Field of Geometric Constraint Temporal Adapter
+
+The size of the convolutional kernel significantly influences the ability to model local structures. In the Geometric Constraint Temporal Adapter, the kernel size of EdgeConv [32] is determined by the number of neighbors  $k$  selected for each anchor point during graph construction. We investigate the impact of varying  $k$  values on spatio-temporal modeling. The kernel radius is defined as the distance between the center point of a point tube and the farthest point tube among its k-nearest neighbors. Figure 6a shows the average kernel radius corresponding to different  $k$  values. Figure 6b shows the video-level recognition accuracy under different  $k$  values.
+
+![](images/24d2e69d841479dd647312bf91ed0597795f6579e157b2cafbdecd396c948c5e.jpg)  
+(a) Kernel radius.  
+Figure 6. The effect of number of edge per point tube in the Geometric Constraint Temporal Adapter.
+
+![](images/bb763ba388302e4fc1090f984ac759b094b44d4fac9f496aaa39caaa9082bdd4.jpg)  
+(b) Accuracy.
+
+As  $k$  increases, the kernel radius also increases, indicating that a larger  $k$  allows the model to consider a broader spatial context when processing local structures. The kernel radius  $(< 0.1)$  is much smaller than the point tube spatial radius (0.3), which means that the GCTA can focus on changes in spatial structure at similar locations in different frames.
+
+In the case of  $k = 1$ , point tubes do not aggregate infor
+
+mation from others. Increasing  $k$  to 5 results in a slight decrease in accuracy to  $91.99\%$ , suggesting that too few neighbors may limit the ability to capture relevant features or even confuse its own information. The peak accuracy of  $95.12\%$  is observed at  $k = 20$ , demonstrating optimal performance when the model has access to a sufficient number of neighbors. However, as  $k$  increases to 25 and 30, the accuracy decreases by  $1.04\%$  and  $0.69\%$ , respectively, suggesting that excessive neighborhood information may introduce redundancy that negatively impacts performance.
+
+We further cut the point cloud video into slices, which limits the process of searching for spatial K-nearest neighbors within those slices. We evaluate the impact of the temporal range on the GCTA. The results shown in Table 4 indicate a clear trend: as the temporal range increases, model performance tends to improve, peaking when the full video is utilized. The model achieves the highest accuracy of  $95.12\%$  when using the full video as the slice length. This result underscores the advantage of utilizing the entire temporal span of the video, enabling the model to integrate global context and dynamics effectively.
+
+<table><tr><td>Slice Lenth</td><td>Acc.(%)</td></tr><tr><td>4</td><td>93.38</td></tr><tr><td>6</td><td>93.03</td></tr><tr><td>8</td><td>94.08</td></tr><tr><td>12</td><td>94.77</td></tr><tr><td>24(full clip)</td><td>95.12</td></tr></table>
+
+Table 4. The effect of temporal range.
+
+# 4.4.4. Location of Geometric Constraint Temporal Adapter
+
+We add the GCTA at different locations to explore how spatio-temporal information can be effectively combined with static point cloud pre-trained models. As shown in the Figure 7, we conducted experiments at 4 locations: (a) Bypass (b) Prepend (c) Post (d) Middle.
+
+The model achieves the highest accuracy of  $95.12\%$  when the GCTA is positioned at the Bypass location. This suggests that integrating spatio-temporal information in a way that allows the model to process it alongside static features leads to optimal performance. The Bypass configuration allows the model to maintain the integrity of the pretrained static features while benefiting from additional temporal dynamics. Compared to the result of placing at the Middle location, those at the Post location indicate that applying the GCTA after adding the global spatial position information still maintains a strong impact on the model's performance, suggesting that the temporal dynamics introduced at this stage enhance the overall understanding of the data. The lowest accuracy is observed at the Prepend loca
+
+![](images/88e359d88179f7419b466c1c8dc5683509051adb906b964f320c09784e636a68.jpg)
+
+![](images/d5e3dcc64c1995e2b8da52f8bd14165b50fdabce0f84cfdf80c0094fc704f54f.jpg)
+
+![](images/faffdac3e9e6a527144e69233fd425d6c0ec77ab7db34f61777b24e2aa13b4b4.jpg)  
+Figure 7. Different location of Geometric Constraint Temporal Adapter.
+
+tion. This placement disrupts the initial processing of static features, indicating that introducing temporal dynamics too early in the pipeline can hinder the model's ability to effectively leverage pre-trained static information.
+
+# 5. Conclusion
+
+In this paper, we explored transferring pre-trained static point cloud models to point cloud video tasks and identified the limitations of static models in capturing temporal dynamics. We proposed a novel Cross-frame Spatiotemporal Adaptation (CSA) strategy, integrating the Point Tube Adapter and Geometric Constraint Temporal Adapter (GCTA) to capture both short-term and long-term temporal dynamics. Our method outperforms existing approaches in 3D action and gesture recognition tasks, demonstrating its effectiveness in bridging the gap between static and dynamic point cloud representations. This work advances point cloud video understanding and offers valuable insights for future research.
+
+# Acknowledgment
+
+This work is supported in part by the National Natural Science Foundation of China, under Grant (62302309,62171248), Shenzhen Science and Technology Program (JCYJ20220818101014030, JCYJ20220818101012025), and the PCNL KEY project (PCL2023AS6-1).
+
+# References
+
+[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 2  
+[2] Suzanne Becker and Geoffrey E Hinton. Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356), 1992. 2  
+[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5  
+[4] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 5  
+[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2  
+[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2  
+[7] Quentin De Smedt, Hazem Wannous, Jean-Philippe Vandeborre, Joris Guerry, Bertrand Le Saux, and David Filliat. Shrec'17 track: 3d hand gesture recognition using a depth and skeletal dataset. In 3DOR-10th Eurographics Workshop on 3D Object Retrieval, pages 1–6, 2017. 2, 6  
+[8] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaiheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2  
+[9] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14204-14213, 2021. 1, 2, 3, 4, 5, 6  
+[10] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2181-2192, 2022. 3, 5  
+[11] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and Mohan Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. arXiv preprint arXiv:2205.13713, 2022. 1, 2, 3, 4, 5  
+[12] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 2
+
+[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2  
+[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 7  
+[15] Jian-Fang Hu, Wei-Shi Zheng, Jiahui Pan, Jianhuang Lai, and Jianguo Zhang. Deep bilinear learning for rgb-d action recognition. In Proceedings of the European conference on computer vision (ECCV), pages 335-351, 2018. 2  
+[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 7  
+[17] Linglin Jing, Ying Xue, Xu Yan, Chaoda Zheng, Dong Wang, Ruimao Zhang, Zhigang Wang, Hui Fang, Bin Zhao, and Zhen Li. X4d-sceneformer: Enhanced scene understanding on 4d point cloud videos through cross-modal knowledge transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2670-2678, 2024. 5  
+[18] Wanqing Li, Zhengyou Zhang, and Zicheng Liu. Action recognition based on a bag of 3d points. In 2010 IEEE computer society conference on computer vision and pattern recognition-workshops, pages 9-14. IEEE, 2010. 2, 5  
+[19] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV, 2022. 2  
+[20] Xingyu Liu, Mengyuan Yan, and Jeannette Bohg. Meteornet: Deep learning on dynamic 3d point cloud sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9246-9255, 2019. 1, 2, 5  
+[21] Yunze Liu, Junyu Chen, Zekai Zhang, Jingwei Huang, and Li Yi. Leaf: Learning frames for 4d point cloud sequence understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 604-613, 2023. 2, 3, 5  
+[22] Yuecong Min, Yanxiao Zhang, Xiujuan Chai, and Xilin Chen. An efficient pointlstm for point clouds based gesture recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5761-5770, 2020. 6  
+[23] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 5  
+[24] 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. 4  
+[25] Zhiqiang Shen, Xiaoxiao Sheng, Hehe Fan, Longguang Wang, Yulan Guo, Qiong Liu, Hao Wen, and Xi Zhou. Masked spatio-temporal structure prediction for self-supervised learning on point cloud videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16580-16589, 2023. 1, 2, 3, 5, 6
+
+[26] Zhiqiang Shen, Xiaoxiao Sheng, Longguang Wang, Yulan Guo, Qiong Liu, and Xi Zhou. Pointcmp: Contrastive mask prediction for self-supervised learning on point cloud videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1212-1222, 2023. 2, 3, 5, 6  
+[27] Xiaoxiao Sheng, Zhiqiang Shen, and Gang Xiao. Contrastive predictive autoencoders for dynamic point cloud self-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9802-9810, 2023. 5  
+[28] Xiaoxiao Sheng, Zhiqiang Shen, Gang Xiao, Longguang Wang, Yulan Guo, and Hehe Fan. Point contrastive prediction with semantic clustering for self-supervised learning on point cloud videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16515-16524, 2023. 1, 2, 3, 5  
+[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4  
+[30] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 2  
+[31] Haiyan Wang, Liang Yang, Xuejian Rong, Jinglun Feng, and Yingli Tian. Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3762-3771, 2021. 3  
+[32] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 4, 7  
+[33] Hao Wen, Yunze Liu, Jingwei Huang, Bo Duan, and Li Yi. Point primitive transformer for long-term 4d point cloud video understanding. In European Conference on Computer Vision, pages 19-35. Springer, 2022. 1, 2, 3, 5  
+[34] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2  
+[35] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 2  
+[36] Siming Yan, Zhenpei Yang, Haoxiang Li, Chen Song, Li Guan, Hao Kang, Gang Hua, and Qixing Huang. Implicit autoencoder for point-cloud self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14530-14542, 2023. 2  
+[37] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313-19322, 2022. 1, 2, 5  
+[38] Yaohua Zha, Rongsheng Li, Tao Dai, Jianyu Xiong, Xin Wang, and Shu-Tao Xia. Sfr: Semantic-aware feature ren
+
+dering of point cloud. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023.  
+[39] Yaohua Zha, Jinpeng Wang, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Instance-aware dynamic prompt tuning for pre-trained point cloud models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14161-14170, 2023. 1, 7  
+[40] Yaohua Zha, Huizhen Ji, Jinmin Li, Rongsheng Li, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Towards compact 3d representations via point feature enhancement masked autoencoders. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6962-6970, 2024. 2  
+[41] Yaohua Zha, Naiqi Li, Yanzi Wang, Tao Dai, Hang Guo, Bin Chen, Zhi Wang, Zhihao Ouyang, and Shu-Tao Xia. Lcm: locally constrained compact point cloud model for masked point modeling. arXiv preprint arXiv:2405.17149, 2024. 2  
+[42] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1  
+[43] Shiwen Zhang, Sheng Guo, Weilin Huang, Matthew R Scott, and Limin Wang. V4d: 4d convolutional neural networks for video-level representation learning. arXiv preprint arXiv:2002.07442, 2020. 2  
+[44] Yabin Zhang, Jiehong Lin, Ruihuang Li, Kui Jia, and Lei Zhang. Point-dae: Denoising autoencoders for self-supervised point cloud learning. arXiv e-prints, pages arXiv-2211, 2022. 1  
+[45] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252-10263, 2021. 2  
+[46] Zhuoyang Zhang, Yuhao Dong, Yunze Liu, and Li Yi. Complete-to-partial 4d distillation for self-supervised point cloud sequence representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17661-17670, 2023. 1, 3, 5  
+[47] Jia-Xing Zhong, Kaichen Zhou, Qingyong Hu, Bing Wang, Niki Trigoni, and Andrew Markham. No pain, big gain: classify dynamic point cloud sequences with static models by fitting feature-level space-time surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8510-8520, 2022. 5, 6

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fb1eb41d647c0f3b94e16d31e48b8a757b4f47ad18e7a98e65b7f601e90e7ce
+size 477031

CVPR/2025/Adapting Pre-trained 3D Models for Point Cloud Video Understanding via Cross-frame Spatio-temporal Perception/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:825602eb2cf8b9dcf509a7cec946aa7a2602aa707dade8541d3455ae6b984b1d
+size 337790

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/0abe3b3e-a1ff-4d75-954a-4898bc338177_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2b73b31084dcb110f83523f9a1e15873a5e748cccab3c7517b5151cfde60a2a
+size 95030

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/0abe3b3e-a1ff-4d75-954a-4898bc338177_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9e5b3bb5a541bfdfd39b58f7732749e99d8591cac81d843352e5f0e578ce36a
+size 127142

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/0abe3b3e-a1ff-4d75-954a-4898bc338177_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3aca0f315f615fa4f532d97272a857c813464842ebfb5722d9f30e9e556d89bd
+size 9993750

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/full.md

@@ -0,0 +1,415 @@
+# Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration
+
+Chao Wang $^{1,2*}$ , Hehe Fan $^{4}$ , Huichen Yang $^{2}$ , Sarvnaz Karimi $^{2}$ , Lina Yao $^{2,3}$ , Yi Yang $^{4\dagger}$ $^{1}$ ReLER, AAII, University of Technology Sydney  
+ $^{2}$ Data61, CSIRO  
+ $^{3}$ University of New South Wales  
+ $^{4}$ CCAI, Zhejiang University
+
+![](images/4b69f5f47df889f965d49123685c2a410717e383b68c662fe2defb26065cf788.jpg)  
+Deraining
+
+![](images/726f4fcafb771b7ea2c2f0f499f6808b13cc907f8c5896c234c16c12af67d05d.jpg)  
+Dehazing
+
+![](images/02cd9c718762293adad24f7f687733c1f83a0fe0e0baa49c206e305d9d164be7.jpg)  
+Desnowing  
+Figure 1. Our image restoration results. The proposed DiffRes introduces cross-modal understanding by leveraging multimodal vision-language models through an efficient adapter, without modifying the pre-trained diffusion model. This enables DiffRes to effectively produce clean and sharp results over 10 image restoration tasks. For simplicity, we only show several representative tasks here.
+
+![](images/91af981aa4f4fd61a0d558250aa4ed9ac0255e71ce70269961248ccfa8de3ba6.jpg)  
+Low-light Enhancement
+
+![](images/beff1d6e23afa4e796bb89d6423764d6e7cd245d91513a6af1e800ce291ba310.jpg)  
+Denoising
+
+![](images/8a6531bef0bd89815ee21ed40f5c4c54935200641f621b1de3b07ba510d490b3.jpg)  
+Motion Deblurring
+
+![](images/3ffcf2e48af0790780544ced5a906d5c5f10d469764aa2668249104eac52a0f5.jpg)  
+Mixed Degradations
+
+# Abstract
+
+Diffusion-based Text-to-Image (T2I) models have demonstrated significant potential in image restoration. However, existing models continue to grapple with challenges such as complex training and prompt design. We introduce a new perspective for improving image restoration by injecting knowledge from pretrained vision-language models into current T2I models. We empirically show that the degradation and content representations in BLIP-2 can be linearly separated, providing promising degradation guidance for image restoration. Specifically, the Feature Difference Instruction (FDI) is first extracted by Q-Formers through a simple subtraction operation based on reference image pairs. Then, we propose a multi-scale FDI adapter to decouple the degradation style and corrupted artifacts, and inject the styleflow exclusively into specific blocks through adapter-tuning, thereby preventing noise interference and eschewing the need for cumbersome weight retraining. In this way, we can train various task-specific adapters according to different degradations, achieving rich detail enhancement in the restoration results. Furthermore, the proposed FDI adapters have attractive properties of practical value, such as composability and generalization abl
+
+ity for all-in-one and mixed-degradation restoration. Extensive experiments under various settings demonstrate that our method has promising repairing quality over 10 image restoration tasks and a wide range of other applications.
+
+# 1. Introduction
+
+As a classical yet challenging Computer Vision (CV) task for many real-world applications, image restoration aims to recover high-quality images from degraded inputs by removing rain [19, 27, 63], haze [22, 64, 84], snow [10, 11, 42], low-light [36, 70, 90], etc. These traditional methods typically concentrate on specific, pre-defined weather conditions. To improve model scalability, researchers have proposed all-in-one networks [12, 31, 38, 45, 77] capable of handling different types of degradations through combined model ensembles. However, such approaches often necessitate separate training for each distinct condition or struggle with multiple degradations within a single image. Aiming at this problem, several works [21, 79, 80] have been proposed to restore images with mixed corruptions in a blind setting, i.e., both the mixing degradation types as well as the mixing mechanism are unknown. Nevertheless, these methods still
+
+![](images/c77406c80d7dd61bc744b4399d094ce3a3faa3f8e7d4573d2b1edda62b9b4ada.jpg)  
+Figure 2. Our framework employs BLIP-2 to extract features from both clean and degraded image sets. The Feature Difference Instruction (FDI) is obtained by subtracting the Q-former output of clean features from that of degraded features. FDI is not only interpretable by the LLM to provide textual descriptions of degradation, it can also be leveraged in our DiffRes as the repairing instruction in a difference-based "DA-BLIP" way.
+
+heavily rely on the complex multi-head structure design or tedious pre-training on large datasets.
+
+Meanwhile, recent advancements in large Multimodal Vision-Language Models (MVLMs) have dramatically improved the capability to process and interpret visual and textual information in tandem. These models, which leverage both vision and language modalities, are integral to a range of applications, from understanding models for image captioning [34, 39, 61, 82, 98], visual question answering [4, 13, 33, 35, 74], to generating models for text-guided image creation [20, 52, 62, 66, 69] and editing [6, 8, 49, 50, 97]. Taking into account both effectiveness and efficiency, how to stably and reliably combine the strengths of both understanding and generating models for downstream CV tasks becomes intractable.
+
+Several studies leverage pretrained MVLMs for downstream computer vision tasks. CLIPStyler [30] and InstanStyle [81] propose to use CLIP-based guidance for style transfer, aligning input image content with target styles described by text prompts. Similar to CLIPStyler, DA-CLIP [45] and MPerceiver [3] further explore acquiring Degradation-Aware (DA) hints from the pretrained CLIP model and then inject the guidance into Text-to-Image (T2I) models for image restoration. However, two problems remain: (1) Both studies heavily rely on the design of prompt generation based on input-degraded images, which leads to mutual interference between content and degradation within the visual embedding [56]; and, (2) For complex illumination-related degradations such as haze, their visual degradative effects cannot be accurately described solely through text descriptions [46].
+
+In this paper, we aim to address these challenges for a more efficient and robust restoration framework, named DiffRes. We observe that BLIP-2 can effectively capture features encompassing both semantic content and degradation attributes, which are linearly separable to some extent. Specifically, we find the feature difference between
+
+clean and degraded images from BLIP-2 can be interpreted by Large Language Models (LLMs) [17], and effectively used as degradation instructions/conditions for T2I models, guiding image restoration during the denoising process (see Figure 2). Compared to past studied that utilize CLIP-based prompts derived from the input degraded image, our reference-image-pair-based Feature Difference Instruction (FDI) is easily accessible while being able to circumvent the impact of degradation factors during the CLIP image encoding, and encapsulate more directed DA-aware information through feature-level subtraction. Combined with the input image features, FDI provides more effective and efficient guidance for high-quality restoration.
+
+Building on this, we further design an adapter tuning-based framework that leverages FDI as a core component to enhance image restoration. Our framework, DiffRes, integrates FDI within a pretrained T2I model, enabling it to precisely reconstruct images using the feature differences as restoration cues. DiffRes consists of an inversion-based T2I model for high-fidelity reconstruction, a multi-scale FDI adapter for precise restoring guidance, and a degradation classifier for mixed degradation removal. During training, the pretrained Stable Diffusion (SD) remains frozen, in which DDIM inversion projects the degraded image into the latent space of LDM to provide the content information. The adapters extract defacement-related guidance instructions from FDI, such as color, depth, illumination, and stylization, and then inject them into the reverse diffusion process in a tuning way. A multi-label classifier enables composite degradation restoration and more customized image processing. By injecting the FDI within the denoising process, DiffRes addresses both semantic content and degradation factors simultaneously. This design not only enhances restoration quality but also improves the model adaptability across various types of degradation, paving the way for a more versatile and robust restoration solution.
+
+Our contributions are threefold:
+
+- We show that the BLIP-2 feature differences between clean and degraded images have the potential to instruct image restoration, with more targeted and clear Degradation-Aware (DA) guidance.  
+- We propose Feature Difference Instruction (FDI) adapter, a more effective tuning-based method to inject the implicit DA instructions into pretrained SD in a low-cost feature addition manner.  
+- Without loss of generalizability, we conduct extensive experiments under various settings to verify the superiority of DiffRes. Results show that our method achieves state-of-the-art performance over 10 image restoration tasks.
+
+# 2. Related Work
+
+Multimodal Vision-Language Models Broadly, multimodal vision-language models can be categorized into two
+
+![](images/90e6862fb0c94e09dbf214ff6fc8bc5b06e7fcf10f42f2f7715f1c6810e6a6bd.jpg)  
+Figure 3. The overview of the proposed DiffRes. DiffRes consists of pretrained stable diffusion, an FDI adapter, and a degradation classifier. Given a degraded image input, DiffRes first utilizes inversion to obtain an accurate latent space representation  $Z_{T}$ . During the denoising process (take step  $S$  as an example), FDI is further optimized through cross-attention with image features  $A_{S}^{i}$ , which are derived from four feature extraction blocks and three downsample blocks. These blocks adjust the feature resolution to match the latent space resolution of stable diffusion  $E_{S}^{i}$ . The FDI adapter then injects the final restoration-relevant guidance  $\hat{A}_{S}^{i}$  into the encoding layers of the Denoising U-Net through direct addition. Meanwhile, the visual modulator in the adapter integrates the optimized FDI  $\hat{F}_{S}^{i}$ , promoting feature regularization  $\theta_{S}^{i}$  in the decoding layers to encourage alignment with the distribution of clean images. Note that all trainable parameters are in the FDI adapter, which can be removed at any time without affecting the original T2I model. Additionally, an optional degradation classifier is employed to enable more challenging all-in-one, mixed-degradation, and customized image restoration.
+
+distinct types: Understanding and Generating. Understanding models such as CLIP (Contrastive Language-Image Pretraining) [61], BLIP (Bootstrapped Language-Image Pretraining) [34], and BLIP-2 [35] advanced cross-modal understanding by aligning visual and linguistic representations. Leveraging these foundations, a recent study [30, 32, 45, 81, 89] explored their applications in computer vision tasks. For instance, InstantStyle [81] leverages CLIP by capturing style features through the difference in encoded features between images and text. Similarly, DEADiff [56] employs BLIP-2 to manually decouple style and content using Q-formers during training. Meanwhile, inspired by the great success of diffusion-based vision-language generation models [52, 62, 66, 69], several recent text-to-image (T2I) methods [3, 40, 53, 68] have been proposed for both supervised and zero-shot computer vision tasks such as image stylization and super-resolution.
+
+In this paper, we propose DiffRes to efficiently combine two pre-trained MVLMs, i.e., vision-language understanding model BLIP-2 [35] and vision-language generating model Stable Diffusion [66], for universal image restoration. Considering the restoration as a process of decoupling degradation style and generating corrupted content, we discover that the feature difference in BLIP-2 can be effectively applied to image restoration tasks as repairing instruction for pre-trained T2I models.
+
+Image Restoration While single degradation image restoration methods [19, 22, 27, 42, 64, 90] have received significant progress, recent studies [9, 12, 38, 45, 54] seek to use an all-in-one framework to handle different degradations simultaneously. For instance, DA-CLIP [45] utilizes an additional image controller based on the pretrained CLIP model, which learns the degradation embeddings matched to the input image through contrastive learning. However, these methods can only repair one single type of degradation every time during inference, making them inapplicable to real-world scenarios where multiple weather degradations may coexist in a single image. To solve this challenge, several methods [3, 21, 79, 80] are proposed for mixed degradations removal. For example, Han et al. propose the Blind Image Decomposition (BID) [21] setting, regarding rain and other real-world weather corruptions as superimposing and separable to a clean image. MPerceiver [3] designs a multimodal prompt-based network with a learnable degradation predictor for restoration. Luo et al. [46] further improved DA-CLIP by incorporating mixed degradation labels into the dataset and optimizing the discrepancy between degraded and clean image embeddings through a fidelity loss. However, these methods still require tedious network training or prompt optimization against input noise interference, thus impeding the real-world implementation.
+
+In contrast, through efficient reference-based adapter tuning on pre-trained T2I models, our DiffRes enables both
+
+degradation-specific or -agnostic, and single- or mixed-degraded image restoration with high flexibility. Meanwhile, benefiting from the multimodal understanding capabilities of BLIP-2 and the generative priors of pre-trained diffusion models, our method can also be extended to other image-to-image translation applications.
+
+# 3. Method
+
+# 3.1. Preliminary
+
+Our method is based on the Latent Diffusion Model (LDM), which operates in a compressed latent space to enable efficient generation. The input image is first encoded into a latent representation using a pretrained Variational Autoencoder (VAE) [28]. The model then combines a forward diffusion process that incrementally adds noise to the latent representation and a reverse denoising process that reconstructs the original data. This approach is formulated as:
+
+$$
+\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {z _ {0}, c, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}, c, t) \| ^ {2} \right], \tag {1}
+$$
+
+where  $z_{t}$  is the noisy latent representation at time step  $t$ ,  $\epsilon$  represents the true Gaussian noise, and  $c$  is the conditional information provided. The network  $\epsilon_{\theta}$  is trained to estimate this noise given the condition  $c$ , guiding the reverse diffusion process to reconstruct the original latent embedding  $z_{0}$ .
+
+After the reverse process, the final denoised latent representation is passed through the VAE decoder to generate the final image in pixel space.
+
+# 3.2. Feature Difference Instruction (FDI)
+
+Given a pretrained BLIP-2 model with a pair of corrupted  $S_{noise}$  and clean image  $S_{clean}$  sets, the Feature Difference Instruction (FDI), denoted as  $F$ , can be obtained as follows:
+
+$$
+\begin{array}{l} F = \mathrm {Q} - \text {f o r m e r (E n c o d e r} (S _ {\text {n o i s y}})) \\ - \text {Q - f o r m e r} (\operatorname {E n c o d e r} \left(S _ {\text {c l e a n}}\right)). \tag {2} \\ \end{array}
+$$
+
+The Q-former, pretrained on large-scale datasets generates features with a certain degree of linear separability, and the differences between paired images can be further interpreted by LLM (see Figure 2). This leads to the potential of aligning these differences with a pretrained T2I model to address downstream CV tasks. Based on this observation, we provide customized alignment between the vision-language understanding models (e.g., BLIP-2) and generating models (e.g., stable diffusion). Instead of retraining T2I models from scratch, we use adapter-tuning methods by integrating external plugins. An overview of our method is presented in Figure 3, consisting of a pretrained inversion-based T2I model for high-fidelity reconstruction, a multiscale FDI adapter for fine-grained instruction, and an optional degradation classifier for mixed degradation removal.
+
+# 3.3. Inversion-based Reconstruction
+
+In our image-to-image restoration task, standard diffusion sampling cannot directly map an existing image to a corresponding noise state. First, we use the pretrained VAE encoder to encode the clean image into the latent space as  $z_0$ . Then, starting from  $z_0$ , we gradually add noise to a specific state  $z_t$  following DDIM inversion [72], thus enabling the restoration conditioned on our FDI. This inversion process can be expressed as:
+
+$$
+\begin{array}{l} z _ {t} = \frac {\sqrt {\alpha_ {t}}}{\sqrt {\alpha_ {t - 1}}} z _ {t - 1} + \sqrt {\alpha_ {t}} \left(\sqrt {\frac {1}{\alpha_ {t}} - 1} \right. \\ \left. - \sqrt {\frac {1}{\alpha_ {t - 1}} - 1}\right) \epsilon_ {\theta} \left(z _ {t - 1}, t - 1\right). \tag {3} \\ \end{array}
+$$
+
+Unlike previous methods [3, 65] that inject content features through extra encoding on degraded images, our inversion-based method injects the content information in a direct and accurate way. It also encourages the model to generate content in regions obscured by degradation during the sampling process. Combined with the FDI adapter, our method facilitates higher-quality restoration from perspectives of noise removal and content inpainting.
+
+# 3.4. FDI Adapter
+
+Noise Decoupling. Since FDI already contains the degradation style and spatial information, we employ an inversion-based diffusion process to inpaint the corrupted areas. Consequently, the FDI, serving as the degradation or de facto restoration guidance, should ideally retain only degradation-related information. However, in practice, when dealing with complex degraded images, such as those severely obscured by rain or haze, the FDI still suffers from certain noise interference from the corrupted unpaired regions. To mitigate this problem, we propose emphasizing the degradation-relevant features in FDI combined with a more comprehensive pixel-level image feature space. We employ cross-attention [78] to refine the input FDI  $F$  (Query) based on the multi-stage image features  $A_{S}^{i}$  (Key, Value) extracted from the adapter. In this way, the refined output FDI  $\hat{F}_S^i$  become more focused on the degradation-related parts, enhancing the accuracy of subsequent image restoration or enhancement tasks.
+
+Multi-level Guidance. After the FDI refinement, we first inject the repairing guidance into the encoding layers of the denoising U-Net. Considering for low-level CV tasks, different encoding channels focus on varying types of information [5, 93]. For example, in the style transfer field [16, 25], features at different levels typically represent different aspects of style: lower-level features capture textures and details, while higher-level features are more inclined towards
+
+overall color and abstract style. Thus, we utilize the optimized FDI  $\hat{F}_S^i$  to further generate channel-wise weights based on the classical SEBlock [24], guiding the focus of information in each encoding layer through channel attention. The channel attention process can be formulated as:
+
+$$
+D _ {S} ^ {i} ^ {'} = D _ {S} ^ {i} \cdot \sigma \left(W _ {2} \cdot \operatorname {R e L U} \left(W _ {1} \cdot \operatorname {G A P} \left(\hat {F} _ {S} ^ {i}\right)\right)\right), \tag {4}
+$$
+
+where  $\cdot$  is the dot product, GAP refers to global average pooling,  $W_{1}$  and  $W_{2}$  are the weight matrices for the fully connected layers, and  $\sigma (\cdot)$  denotes the Sigmoid function.
+
+After channel attention, the final instruction  $\hat{A}_S^i$  has the same dimension as the intermediate feature  $E_S^i$  in the denoising U-Net. Similar to the T2I-Adapter [50], we directly add  $\hat{A}_S^i$  with  $E_S^i$  at each scale  $i$  for guidance injection.
+
+Visual Modulator. For the decoding layers in U-Net, we also utilize SEBlock, which is based on the refined FDI  $\hat{F}_S^i$  and the decoding features  $D_S^i$  following the skip-connection correspondence. Furthermore, we incorporate hierarchical stylized degradation information into the generation process through AdaIN [25], aligning with the encoding process. As shown in Figure 3 (a), our adaptive instance normalization can be expressed as:
+
+$$
+\hat {D} _ {S} ^ {i} = \gamma \left(\hat {F} _ {S} ^ {i}\right) \odot \operatorname {N o r m} \left(D _ {S} ^ {i} ^ {\prime}\right) + \beta \left(\hat {F} _ {S} ^ {i}\right), \tag {5}
+$$
+
+where  $\odot$  represents the element-wise multiplication,  $\gamma (\cdot)$  and  $\beta (\cdot)$  are implemented through two convolutional layers. Note that this visual modulator is also a plug-and-play module as a part of our FDI Adapter.
+
+# 3.5. Degradation Classifier
+
+In real-world scenarios, multiple types of degradation frequently co-occur, leading to the coexistence of various distortions within a single image. For instance, an image may depict weather conditions involving both rain and fog simultaneously. To address this complexity, we additionally deploy an MLP-based multi-label classifier to identify the degradation types present in the input image, as illustrated in Figure 3 (b). This multi-label classifier enables our DiffRes model to perform both all-in-one and mixed-degradation removal with high flexibility from two perspectives. (1) Composite: Simultaneously employ multi-branch FDI adapters trained on mixed datasets with different degradation types. Each branch can be combined through direct addition with adjustable weights obtained from the classifier. In this setting, the input  $\hat{F}_S^i$  of our visual modulator is the mean value of each adapter branch. (2) Progressive: Sequentially apply different task-specific adapters. This approach also allows for flexible customization to selectively remove specific weather effects, as shown in Figure 4.
+
+![](images/4837ab8c13c49c73b6594aeca40acc43b8af1b8b006dcf717a21b6decc734880.jpg)  
+Figure 4. Illustration of composite restoration (red) and progressive restoration (blue). Both approaches can produce satisfactory results for mixed-degradation removal. Zoom in to see the details.
+
+# 4. Experimental Setup and Results
+
+Datasets. For the task-specific and task-agnostic settings, we consider 10 restoration tasks: deraining, raindrop removal, dehazing, desnowing, low-light enhancement, denoising, motion deblurring, defocus deblurring, shadow removal, and JPEG decompression. For each degradation type, we collect data from the corresponding datasets, Rain: Rain100 (L&H) [88] and Rain1400 [19]; Haze: RESIDE-6K [58]; Snow: Snow100K [42]; Raindrop: RainDrop [57]; Low-light: LOL-v1 [83] and MEF [47]; Noise: DIV2K [2], Flickr2K [75] and CBSD68 [48]; Motion blur: GoPro [51]; Defocus blur: DPDD [1]; Shadow: SRD [59]; JPEG compression: DIV2K [2], Flickr2K [75] and LIVE1 [71]. For the mixed-degradation setting, we use RainDS [60] for raindrop + rainstreak, NYU-Rain [37] and OutDoor-Rain [77] for rainstreak + haze + low-light, and BID [21] for more complex combinations. Note that during the testing stage, to ensure a fair comparison, we use the whole training set corresponding to the test set to obtain the average FDI. Details of the dataset selection under different task settings are provided in the Supp.
+
+Evaluation Metrics. Apart from the commonly used pixel-level metrics PSNR and SSIM, we also adopt a variety of widely used metrics to provide a comprehensive evaluation of the different methods. In terms of feature-level perceptual quality, we use FID [23] and LPIPS [99] as reference-based metrics to measure the distance and similarity between ground truth images and the restored images.
+
+Implementation Details. We choose the pre-trained SD V2.1-512 model with 50 sampling steps as our base model. During training, we employ the AdamW optimizer [43]  $(\beta_{1} = 0.9, \beta_{2} = 0.99)$  over 50 epochs with a learning rate of  $1 \times 10^{-5}$ . All experiments are conducted on 4 NVIDIA A100 GPUs with a batch size of 32. The total training time of one FDI adapter is approximately 8 hours
+
+Table 1. Quantitative comparison between our method with other state-of-the-art approaches on nine different degradation-specific datasets.  $\downarrow$  represents the smaller the better, and  $\uparrow$  represents the bigger the better. Best and second best performance are in red and blue, respectively.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Deraining (Rain100H)</td><td rowspan="2">Method</td><td colspan="2">Dehazing (RESIDE)</td><td rowspan="2">Method</td><td colspan="2">Desnowing (Snow100K)</td></tr><tr><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td></tr><tr><td>MPRNet [91]</td><td>30.42 / 0.891</td><td>61.59 / 0.158</td><td>DehazeFormer [73]</td><td>30.29 / 0.964</td><td>7.58 / 0.045</td><td>DesnowNet [42]</td><td>27.53 / 0.899</td><td>9.84 / 0.147</td></tr><tr><td>MAXIM [76]</td><td>30.81 / 0.903</td><td>58.72 / 0.133</td><td>MAXIM [76]</td><td>29.12 / 0.932</td><td>8.12 / 0.043</td><td>DDMSNet [96]</td><td>28.85 / 0.877</td><td>3.24 / 0.096</td></tr><tr><td>IR-SDE [44]</td><td>31.65 / 0.904</td><td>18.64 / 0.047</td><td>IR-SDE [44]</td><td>25.25 / 0.906</td><td>8.33 / 0.060</td><td>PromptIR [55]</td><td>30.91 / 0.913</td><td>3.79 / 0.100</td></tr><tr><td>DA-CLIP [45]</td><td>33.91 / 0.926</td><td>11.79 / 0.031</td><td>DA-CLIP [45]</td><td>30.16 / 0.936</td><td>5.52 / 0.030</td><td>DA-CLIP [45]</td><td>30.83 / 0.907</td><td>3.01 / 0.082</td></tr><tr><td>Ours</td><td>34.55 / 0.935</td><td>9.97 / 0.023</td><td>Ours</td><td>30.87 / 0.970</td><td>5.08 / 0.023</td><td>Ours</td><td>31.32 / 0.925</td><td>2.10 / 0.077</td></tr><tr><td rowspan="2">Method</td><td colspan="2">Randrop Remocal (RainDrop)</td><td rowspan="2">Method</td><td colspan="2">Low-light Enhancement (LOL-v1)</td><td rowspan="2">Method</td><td colspan="2">Denoising (CBSD68, sigma=50)</td></tr><tr><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td></tr><tr><td>AttGAN [57]</td><td>31.59 / 0.917</td><td>22.38 / 0.058</td><td>MIRNet [90]</td><td>24.14 / 0.830</td><td>69.18 / 0.250</td><td>Restormer [92]</td><td>26.85 / 0.753</td><td>75.12 / 0.221</td></tr><tr><td>IDT [87]</td><td>31.87 / 0.931</td><td>25.54 / 0.059</td><td>URetinex-Net [85]</td><td>19.84 / 0.824</td><td>52.38 / 0.237</td><td>Xformer [95]</td><td>29.45 / 0.848</td><td>59.33 / 0.152</td></tr><tr><td>PromptIR [55]</td><td>32.05 / 0.933</td><td>25.13 / 0.054</td><td>Retinexformer [7]</td><td>24.08 / 0.825</td><td>58.31 / 0.244</td><td>PromptIR [55]</td><td>28.82 / 0.816</td><td>65.33 / 0.177</td></tr><tr><td>DA-CLIP [45]</td><td>30.75 / 0.892</td><td>24.31 / 0.061</td><td>DA-CLIP [45]</td><td>23.75 / 0.833</td><td>34.03 / 0.083</td><td>DA-CLIP [45]</td><td>28.55 / 0.807</td><td>66.35 / 0.201</td></tr><tr><td>Ours</td><td>33.70 / 0.939</td><td>19.65 / 0.047</td><td>Ours</td><td>24.55 / 0.839</td><td>29.17 / 0.075</td><td>Ours</td><td>29.17 / 0.823</td><td>61.29 / 0.161</td></tr><tr><td rowspan="2">Method</td><td colspan="2">Motion Deblur (GoPro)</td><td rowspan="2">Method</td><td colspan="2">Defocus Deblur (DPDD)</td><td rowspan="2">Method</td><td colspan="2">JPEG Compression (LIVE1, Q=10)</td></tr><tr><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td><td>PSNR ↑ / SSIM ↑</td><td>FID ↓ / LPIPS ↓</td></tr><tr><td>DeblurGAN [29]</td><td>28.70 / 0.858</td><td>27.02 / 0.178</td><td>DRBNet [67]</td><td>25.73 / 0.791</td><td>49.04 / 0.183</td><td>FBCNN [26]</td><td>27.77 / 0.803</td><td>30.16 / 0.997</td></tr><tr><td>Restormer [92]</td><td>30.05 / 0.937</td><td>13.54 / 0.138</td><td>FocalNet [15]</td><td>26.25 / 0.810</td><td>47.83 / 0.210</td><td>Restormer [92]</td><td>27.65 / 0.805</td><td>30.52 / 0.103</td></tr><tr><td>DiffIR [86]</td><td>32.55 / 0.952</td><td>10.21 / 0.087</td><td>PromptIR [55]</td><td>25.66 / 0.791</td><td>53.64 / 0.197</td><td>PromptIR [55]</td><td>27.47 / 0.793</td><td>32.34 / 0.078</td></tr><tr><td>DA-CLIP [45]</td><td>30.88 / 0.903</td><td>6.15 / 0.058</td><td>DA-CLIP [45]</td><td>26.15 / 0.802</td><td>48.11 / 0.193</td><td>DA-CLIP [45]</td><td>27.35 / 0.790</td><td>32.97 / 0.121</td></tr><tr><td>Ours</td><td>32.58 / 0.952</td><td>6.15 / 0.055</td><td>Ours</td><td>26.22 / 0.810</td><td>44.39 / 0.188</td><td>Ours</td><td>27.77 / 0.802</td><td>30.05 / 0.081</td></tr></table>
+
+Table 2. Quantitative results on five restoration tasks with state-of-the-art all-in-one methods.  $\downarrow$  represents the smaller the better, and  $\uparrow$  represents the bigger the better. Best and second best performance are in red and blue, respectively.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Deraining (Rain100L)</td><td colspan="2">Dehazing (SOTS)</td><td colspan="2">Denoising (BSD68, average)</td><td colspan="2">Deblurring (GoPro)</td><td colspan="2">Low-light Enhance. (LOL-v1)</td><td colspan="2">Average</td></tr><tr><td>PSNR ↑</td><td>SSIM ↑</td><td>PSNR ↑</td><td>SSIM ↑</td><td>PSNR ↑</td><td>SSIM ↑</td><td>PSNR ↑</td><td>SSIM ↑</td><td>PSNR ↑</td><td>SSIM ↑</td><td>PSNR ↑</td><td>SSIM ↑</td></tr><tr><td>DL [18]</td><td>21.96</td><td>0.762</td><td>20.54</td><td>0.826</td><td>23.09</td><td>0.745</td><td>19.86</td><td>0.672</td><td>19.83</td><td>0.712</td><td>21.05</td><td>0.743</td></tr><tr><td>TransWeather [77]</td><td>29.43</td><td>0.905</td><td>21.32</td><td>0.885</td><td>29.00</td><td>0.841</td><td>25.12</td><td>0.757</td><td>21.21</td><td>0.792</td><td>25.22</td><td>0.836</td></tr><tr><td>TAPE [41]</td><td>29.67</td><td>0.904</td><td>22.16</td><td>0.861</td><td>30.18</td><td>0.855</td><td>24.47</td><td>0.763</td><td>18.97</td><td>0.621</td><td>25.09</td><td>0.801</td></tr><tr><td>AirNet [31]</td><td>32.98</td><td>0.951</td><td>21.04</td><td>0.884</td><td>30.91</td><td>0.882</td><td>24.35</td><td>0.781</td><td>18.18</td><td>0.735</td><td>25.49</td><td>0.846</td></tr><tr><td>IDR [94]</td><td>35.63</td><td>0.965</td><td>25.24</td><td>0.943</td><td>31.60</td><td>0.887</td><td>27.87</td><td>0.846</td><td>21.34</td><td>0.826</td><td>28.34</td><td>0.893</td></tr><tr><td>PromptIR [55]</td><td>36.75</td><td>0.963</td><td>26.63</td><td>0.945</td><td>31.58</td><td>0.886</td><td>28.44</td><td>0.863</td><td>22.34</td><td>0.826</td><td>28.79</td><td>0.894</td></tr><tr><td>DA-CLIP [45]</td><td>35.91</td><td>0.972</td><td>26.28</td><td>0.939</td><td>31.58</td><td>0.887</td><td>28.81</td><td>0.882</td><td>22.57</td><td>0.832</td><td>29.23</td><td>0.898</td></tr><tr><td>InstructIR-5D [14]</td><td>36.84</td><td>0.973</td><td>27.10</td><td>0.956</td><td>31.40</td><td>0.883</td><td>29.40</td><td>0.886</td><td>23.00</td><td>0.836</td><td>29.55</td><td>0.907</td></tr><tr><td>Ours</td><td>37.25</td><td>0.979</td><td>27.23</td><td>0.958</td><td>32.07</td><td>0.890</td><td>29.33</td><td>0.883</td><td>23.13</td><td>0.843</td><td>29.78</td><td>0.911</td></tr></table>
+
+with 45M parameter count. We divide the sampling process into three stages and inject the restoration instructions into the denoising U-Net during the beginning stage, as adding guidance in the middle and late stages has minimal effect on the results [50]. In addition to  $\mathcal{L}_{\mathrm{LDM}}$ , we incorporate an  $\mathcal{L}_1$  reconstruction loss between the generated and GT images in the latent space. Note that during testing, since degradation types are known, the degradation classifier is not needed. During inference, our classifier gives users more customized options. Given a random inference image, DiffRes first identifies potential degradation types (choosing one or more as needed) using the classifier and then computes the FDI from any dataset of that type. They
+
+can either manually set thresholds based on predicted labels to select the most similar degradation types for composite restoration, or progressively eliminate each specified type of degradation. More details are provided in the Supp.
+
+# 4.1. Results
+
+Single-task setting. We independently train the FDI-Adapter without the degradation classifier and evaluate it on task-specific restoration. The quantitative results across different datasets are summarized in Table 1. Compared to other task-specific methods, our approach achieves the best overall performance on most tasks. Figure 5 provides a visual comparison. Compared to other CLIP-based methods,
+
+![](images/06593735ef561bb78d19c015fb864c11faa5d23cbc4cd5cf76b24768df8bb6e2.jpg)
+
+![](images/c374f9bc27442d391d03b6ff4ff567b11ef01294ad20eed80f6331bdca82fc5f.jpg)
+
+![](images/794859a674cc75e6bb2968e7beade6022e0ff0224304d9d9ec9db6712f5471b1.jpg)  
+Figure 5. Qualitative comparison of different methods on four restoration tasks. Please zoom in to see the details.
+
+![](images/d83037939281501000ed4da2fd9a92487473c88e4cf149b1a87e40107659cfca.jpg)
+
+![](images/c05dd098806eed1ce74fcc2d65adcbf3a07aea82ed5dfac8a616c7bad6a3d70d.jpg)
+
+![](images/ebd0fdd18aa0d1bb538d819813abdb7cdc00362033f21abd6b69165aaee112ba.jpg)
+
+![](images/3a4ffe5e2685510f48e7efb28dc66f2713428d7e1e5acf3cd44ecf5243a9c845.jpg)
+
+![](images/a0bd2afe937729d0f57415d6d53a54da0f56d061c04f7c3097a180448cffd439.jpg)
+
+![](images/653641fa019eba5b8feecf1955739236f4ed210903af5e14b0f7aafc398a02bb.jpg)  
+Mixed-degraded Input
+
+![](images/24162be6815fd0308e867b1cba7f0c65b022a95734cfce3f4050fd309f8bb7ad.jpg)  
+Composite Restoration Output
+
+![](images/625f1a715db8cd2012593dfd5db823ddce93833ffe13d3826c7a334284e5d209.jpg)  
+Mixed-degraded Input
+
+![](images/62a1245baf086fad18d5a8d4a8fd098b49761e28a0ed75d865da1085c0121706.jpg)  
+Composite Restoration Output
+
+![](images/937ffa44b5fa0cee301175d81870ac8ff73b1ec4f58ee12a8d1fc8d7e4f83587.jpg)  
+Mixed-degraded Input
+
+![](images/ac183078d9ce26197e622123544f696630f0286e2bb06bb34d0b1b2bc9f6f353.jpg)  
+Composite Restoration Output
+
+![](images/5dab6cce59781ec106c6d45d61975d725356f06dc941f649d8f087fe77ecc137.jpg)  
+Single LLIE Output
+
+![](images/8578d8134fc338a0eab6e5558537d01b1a6537dfbdaeddcadb7890ac89259fb7.jpg)  
+Single Deraining Output
+
+![](images/1509c5f9ece23a4eb07e4f2071b15e827e46c05008122912a7297ffcc30c4a1c.jpg)  
+Single Dehazing Output
+
+![](images/78370798e7620f317b2bd9f06aef1c2e956cc51198d0ea2acf03c8f75f4ab74c.jpg)  
+Single Deraining Output  
+Figure 6. Visual examples on hybrid-degraded images. Our method remains superior performance when input images are simultaneously corrupted by mixed degradations. Our DiffRes can also achieve customized specific degradation removal with composable adapters.
+
+![](images/68bfcb818fe0640b8fde85c3aed77d72104c8bb4a20bce4d18cc9d7d94465b3b.jpg)  
+Single Dehazing Output
+
+![](images/4a4198ac0eff746709aefe809ef6f226db0aee5a2e30690d4db868e90944d68c.jpg)  
+Single Deraining Output
+
+such as DA-CLIP [45], our approach generates sharp and visually pleasing results that closely resemble high-quality images. This demonstrates that FDI from BLIP-2 leads to better performance for downstream restoration models. Note that DiffRes is particularly suitable for noise types that the degraded image differs noticeably from the original, such as adverse weather conditions, where the noise affects the original image semantics, allowing BLIP-2 to capture it and extract degradation clues through subtraction.
+
+Multi-task setting. We evaluate our method for unified image restoration, which is trained on a mixed dataset with multiple adapters. Table 2 presents a comprehensive comparison with SOTA all-in-one methods across 5 tasks. Our method consistently outperforms these compared methods on all datasets. Compared to other all-in-one methods, our DiffRes achieves the best average performance, primarily benefiting from the complementary interactions between multiple adapters in handling various degradations. This advantage is especially obvious in challenging test inputs, which often contain elements of other degradation types (e.g., noise in low-light images).
+
+Mixed-degradation setting. During the multi-branch adapter training, each adapter independently captures the restoration specific to different degradations. Therefore, our
+
+Table 3. Quantitative comparisons on mixed-degradation images. InstructIR is manually given text instructions for progressive restoration, each time restoring one type of degradation.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Rainstreak + Raindrop (RainDS, Real-world)</td><td colspan="3">Rainstreak + Haze + Low-light (Outdoor-Rain, Test 1)</td></tr><tr><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>AirNet [31]</td><td>23.41</td><td>0.633</td><td>0.390</td><td>19.29</td><td>0.615</td><td>0.438</td></tr><tr><td>TransWeather [77]</td><td>24.13</td><td>0.655</td><td>0.378</td><td>20.77</td><td>0.631</td><td>0.386</td></tr><tr><td>PromptIR [55]</td><td>24.05</td><td>0.661</td><td>0.380</td><td>21.47</td><td>0.704</td><td>0.314</td></tr><tr><td>Restormer [92]</td><td>23.39</td><td>0.628</td><td>0.396</td><td>20.80</td><td>0.687</td><td>0.335</td></tr><tr><td>InstructIR-5D [14]</td><td>24.26</td><td>0.647</td><td>0.330</td><td>21.45</td><td>0.707</td><td>0.285</td></tr><tr><td>Ours (composite)</td><td>25.11</td><td>0.702</td><td>0.230</td><td>22.30</td><td>0.791</td><td>0.221</td></tr></table>
+
+Table 4. Cross-dataset generalization test on different training sets.  
+
+<table><tr><td>Test Set</td><td colspan="3">Deraining (GT-Rain)</td><td colspan="3">Dehazing (O-Haze)</td></tr><tr><td>FDI source</td><td>Rain800</td><td>NYU-Rain</td><td>GT-Rain</td><td>ITS</td><td>OTS</td><td>O-Haze</td></tr><tr><td>PSNR ↑</td><td>19.52</td><td>22.33</td><td>22.49</td><td>21.33</td><td>25.09</td><td>26.88</td></tr><tr><td>LPIPS ↓</td><td>0.401</td><td>0.221</td><td>0.207</td><td>0.167</td><td>0.146</td><td>0.131</td></tr></table>
+
+DiffRes also achieves single-image mixed-degradation removal with improved results. Table 3 presents quantitative results for two common real-world degradation combinations. Additionally, Figure 8 shows that, beyond composite removal, DiffRes can also perform progressive restoration through consecutive single-task FDI adapters. Further mixed-degradation results with different combinations are provided in the Supp 3.
+
+![](images/61913da65e6f94a43634edaab235d920ed0f6b7466561a8d90c5deab567e2849.jpg)  
+OutDoor-Rain Test Input
+
+![](images/d21489fc52e0e9e73cd046511c49423d966e2f1b0fb3b1d40d67bbb3644881f6.jpg)  
+Ground Truth
+
+![](images/a7901730d63c04dddb6d0b2ea5be85a3feca6d7be21efa276f7d294f05e25044.jpg)
+
+![](images/4a88bf335b555901b5414996eb3612e97da894f61b9879300c20557be19f7783.jpg)  
+FDI from Rain100H PSNR: 14.28dB
+
+![](images/1c6469de7d881e42aaa8e7b91f57d43428054a56ae859c8c971da73bf9a2e95c.jpg)  
+FDI from Rain1400 PSNR: 17.46dB
+
+![](images/3bf7bf0037fb67c6574eca0d5af5e759684e7757bc958b90687a9eed0b548f1b.jpg)  
+FDI from OutDoor-Rain PSNR: 24.30dB  
+FDI from NYU-Rain PSNR:22.20dB  
+Figure 7. Examples of cross-dataset experiments: given a test input, we apply FDI from different datasets within the same task to perform restoration. Examples are evaluated using PSNR.
+
+Cross-dataset. We observed empirically that for tasks with abundant datasets (e.g., deraining), the FDI from these datasets within the same task often demonstrates direct cross-dataset compatibility. As illustrated in Figure 7, datasets such as NYU-Rain [37] and OutDoor-Rain [77], which both emphasize depth-aware rain accumulation effects like illumination and accompanying haze, exhibit superior cross-dataset FDI performance compared to datasets that merely depict synthetic rain streaks. Table 4 further presents a generalization test of different training sets on real-world test sets. Note that this is only a preliminary, performance could improve with more advanced datasets.
+
+Cross-task. Apart from being composable, FDI from certain tasks also demonstrates cross-task compatibility (e.g., leveraging FDI from denoising datasets for image desnowing). This suggests that we can address multiple restoration tasks with fewer adapter combinations. Results for different task combinations are provided in the Supp.
+
+Progressive vs. Composite. Figure 8 compares progressive and composite restoration using three task-specific adapters. While no clear pattern emerged in the progressive restoration order, it generally outperformed composite restoration by addressing specific degradations and minimizing interference between them [46].
+
+Furthermore, our FDI mechanism can be extended to multiple applications. For example, by customizing reference image pairs, DiffRes can be applied to image-to-image translation tasks such as style transfer. This flexibility enables our method to adapt to various image transformation objectives, using a unified framework to achieve diverse translation results in image processing. Other applications are provided in Supp 3.
+
+# 4.2. Ablation Study
+
+To understand the contribution of each of the components in DiffRes, we conduct further ablation experiments with different model variants. The quantitative results are pre
+
+![](images/1b90c71520fae4ee00349de1df5812f18b859508a94e114528b3106924ad95a4.jpg)  
+Multi-degradation Input
+
+![](images/cf52a377cddc9c76686caac5f96eeeedbf5cd58d4b320a19ffa471a16f68e799.jpg)
+
+![](images/a135bf462b7ed8908da210fa27b76fc13e66506d7423fa79ae973b52d003ac22.jpg)
+
+![](images/275cc4fa92dcba15f2ac7b645d781648b3c6385fa008281ea1d9c5eb58f4be01.jpg)
+
+![](images/c1b1a460acf65765b364b69a1323443f48eef8bb4ae1e85ed588b0ff2043fce5.jpg)  
+Multi-degradation Input  
+Figure 8. Comparisons on progressive and composite restorations.
+
+![](images/49ebea9b3c201ab2c2be3efe4ca88f3728b32415b1bec015e11c6a6a3ffb3ca1.jpg)  
+Dehazing  $\rightharpoonup$  Derainaging  $\rightarrow$  LLIIE PSNR:24.25dB  
+Deraining  $\rightarrow$  Dehazing  $\rightarrow$  LLIIE PSNR:23.15dB
+
+![](images/73c19d490d17eb2a226d59253781d77344ac6dd2c2860d5c65d5614f65a9e679.jpg)  
+PSNR:24.61dB  
+$\rightharpoonup$  Dehazing  $\rightarrow$  Deraining PSNR:29.22dB
+
+![](images/023e4900d0c154077a908f050bac90015edff65bc43e7ff2b94622fc52658feb.jpg)  
+Composite (1:1:1) PSNR: 21.18dB  
+Composite (1:1:1) PSNR: 22.47dB
+
+Table 5. Ablation experiments of DiffRes. The metrics are reported on the average of the restoration tasks with snow, haze, rainstreak, and low-light degradations.  
+
+<table><tr><td>Method</td><td>PSNR ↑</td><td>SSIM ↑</td><td>LPIPS ↓</td></tr><tr><td>Baseline (Stable Diffusion)</td><td>15.33</td><td>0.423</td><td>0.655</td></tr><tr><td>+ Visual Modulator (replace \(\hat{F}_S^i\) with \(F\))</td><td>15.90</td><td>0.435</td><td>0.651</td></tr><tr><td>+ Visual Branch w/o Channel-Attn</td><td>15.28</td><td>0.419</td><td>0.648</td></tr><tr><td>+ Noise Decoupling</td><td>21.49</td><td>0.617</td><td>0.351</td></tr><tr><td>+ Noise Decoupling w/o Channel-Attn (replace \(\hat{A}_S^i\) with \(A_S^i\))</td><td>19.32</td><td>0.538</td><td>0.577</td></tr><tr><td>+ Noise Decoupling w/o Cross-Attn (replace \(\hat{F}_S^i\) with \(F\))</td><td>20.95</td><td>0.608</td><td>0.374</td></tr><tr><td>+ Noise Decoupling + Visual Modulator (Full FDI Adapter)</td><td>24.14</td><td>0.699</td><td>0.280</td></tr></table>
+
+sented in Table 5. We begin with the initial stable diffusion by adding the proposed components of our FDI adapter. We observe that high-quality image restoration requires both degradation guidance from the FDI and feature regularization from the visual modulator. The output tends to retain some corrupted content without noise decoupling, and it generates more artifacts without the visual modulator.
+
+# 5. Conclusion
+
+Image restoration is challenging, even with the advances offered by Diffusion-based Text-to-Image (T2I) models. We explore the linear separability of features in the pretrained BLIP-2, and its potential across various image restoration tasks. Combining this with the multi-scale features subject to the input reference images, we propose a novel Feature Difference Instruction (FDI), to efficiently inject the degradation-aware guidance into the pretrained text-to-image models. Taking advantage of inversion-based diffusion and reference-based FDI adapter, we achieved a flexible tuning-based pipeline, namely DiffRes, to efficiently restore arbitrary hybrid degradations in a unified framework. To enable all-in-one and mixed-degradation learning, we further introduce an additional classifier providing adjustable weights for the composable FDI adapter. Extensive experiments on 10 restoration tasks and other translation applications demonstrate the effectiveness and superiority of our unified network over strong baselines. We hope that our method takes the research community one step closer to transferring the knowledge of vision-language models to more CV tasks in the future.
+
+# Acknowledgment
+
+This work was supported in part by the National Natural Science Foundation of China (U2336212), Fundamental Research Funds for the Zhejiang Provincial Universities (226-2024-00208), and the Earth System Big Data Platform of the School of Earth Sciences, Zhejiang University. Chao Wang is supported by the CSIRO Data61 top-up scholarship.
+
+# References
+
+[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In ECCV, 2020. 5, 1  
+[2] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPR workshops, 2017. 5, 1  
+[3] Yang Ai, Huaibo Huang, Xiaogiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In CVPR, 2024. 2, 3, 4  
+[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 2  
+[5] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In CVPR, 2017. 4  
+[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2  
+[7] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In ICCV, 2023. 6  
+[8] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 2  
+[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 3  
+[10] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In ECCV, 2020. 1  
+[11] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In ICCV, 2021. 1  
+[12] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge
+
+learning and multi-contrastive regularization: Toward a unified model. In CVPR, 2022. 1, 3  
+[13] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 2  
+[14] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In ECCV, 2025. 6, 7  
+[15] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Focal network for image restoration. In ICCV, 2023. 6  
+[16] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In ICLR, 2017. 4  
+[17] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, 2024. 2, 3  
+[18] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators. TPAMI, 43 (1):33-47, 2019. 6  
+[19] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In CVPR, 2017. 1, 3, 5  
+[20] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022. 2  
+[21] J. Han, W. Li, P. Fang, C. Sun, J. Hong, M. A. Armin, L. Petersson, and H. Li. Blind image decomposition. In ECCV, 2022. 1, 3, 5  
+[22] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. TPAMI, 33(12):2341-2353, 2010. 1, 3  
+[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 5  
+[24] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 5  
+[25] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 4, 5  
+[26] Jiaxi Jiang, Kai Zhang, and Radu Timofte. Towards flexible blind JPEG artifacts removal. In ICCV, 2021. 6  
+[27] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 1, 3  
+[28] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 4  
+[29] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In CVPR, 2018. 6  
+[30] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In CVPR, 2022. 2, 3
+
+[31] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, 2022. 1, 6, 7  
+[32] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS, 2024. 3  
+[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 2  
+[34] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 3  
+[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2, 3  
+[36] Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retina model. TIP, 27(6):2828-2841, 2018. 1  
+[37] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In CVPR, 2019. 5, 8  
+[38] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In CVPR, 2020. 1, 3  
+[39] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. 2  
+[40] Xinqi Lin, Jingwen He, Zihan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diffir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3  
+[41] Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xi-angyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. In ECCV, 2022. 6  
+[42] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. TIP, 27(6):3064–3073, 2018. 1, 3, 5, 6  
+[43] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5  
+[44] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Image restoration with mean-reverting stochastic differential equations. arXiv preprint arXiv:2301.11699, 2023. 6  
+[45] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for universal image restoration. In ICLR, 2024. 1, 2, 3, 6, 7  
+[46] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Photo-realistic image restoration in the wild with controlled vision-language models. In CVPR, 2024. 2, 3, 8
+
+[47] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. TIP, 24(11): 3345-3356, 2015. 5, 1  
+[48] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 5, 1  
+[49] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2  
+[50] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2, 5, 6, 1, 3  
+[51] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 5, 1  
+[52] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3  
+[53] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI, 45(8):10346-10357, 2023. 3  
+[54] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In CVPR, 2023. 3  
+[55] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In NeurIPS, 2023. 6, 7  
+[56] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In CVPR, 2024. 2, 3  
+[57] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In CVPR, 2018. 5, 6, 1  
+[58] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In AAAI, 2020. 5, 1  
+[59] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In CVPR, 2017. 5, 1  
+[60] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In CVPR, 2021. 5, 1  
+[61] 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 ICML, 2021. 2, 3
+
+[62] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3  
+[63] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In CVPR, 2019. 1  
+[64] Wenqi Ren, Jinshan Pan, Hua Zhang, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks with holistic edges. IJCV, 128:240-259, 2020. 1, 3  
+[65] Yulin Ren, Xin Li, Bingchen Li, Xingrui Wang, Mengxi Guo, Shijie Zhao, Li Zhang, and Zhibo Chen. Moe-diffir: Task-customized diffusion priors for universal compressed image restoration. In ECCV, 2025. 4  
+[66] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3  
+[67] Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. Learning to deblur using light field generated and real defocus images. In CVPR, 2022. 6  
+[68] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In SIGGRAPH, 2022. 3  
+[69] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2, 3  
+[70] Kai Shang, Mingwen Shao, Chao Wang, Yuanshuo Cheng, and Shuigen Wang. Multi-domain multi-scale diffusion model for low-light image enhancement. In AAAI, 2024. 1  
+[71] H Sheikh. Live image quality assessment database release 2. http://live.ece.utexas.edu/research/quality, 2005. 5, 1  
+[72] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4  
+[73] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. TIP, 32:1927-1941, 2023. 6  
+[74] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5100-5111, 2019. 2  
+[75] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 2017. 5, 1  
+[76] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxim: Multi-axis mlp for image processing. In CVPR, 2022. 6  
+[77] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In CVPR, 2022. 1, 5, 6, 7, 8
+
+[78] A Vaswani. Attention is all you need. NeurIPS, 2017. 4  
+[79] Chao Wang, Zhedong Zheng, Ruijie Quan, Yifan Sun, and Yi Yang. Context-aware pretraining for efficient blind image decomposition. In CVPR, 2023. 1, 3  
+[80] Chao Wang, Zhedong Zheng, Ruijie Quan, and Yi Yang. Depth-aware blind image decomposition for real-world adverse weather recovery. In ECCV, 2025. 1, 3  
+[81] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 2, 3  
+[82] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2  
+[83] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In BMVC, 2018. 5, 1  
+[84] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, 2021. 1  
+[85] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In CVPR, 2022. 6  
+[86] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 6  
+[87] Jie Xiao, Xueyang Fu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Image de-raining transformer. TPAMI, 45(11):12978-12995, 2022. 6  
+[88] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. TPAMI, 42(6):1377-1393, 2019. 5, 1  
+[89] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 3  
+[90] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In ECCV, 2020. 1, 3, 6  
+[91] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 6  
+[92] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 6, 7  
+[93] MD Zeiler. Visualizing and understanding convolutional networks. In ECCV, 2014. 4  
+[94] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-
+
+degradation learning for image restoration. In CVPR, 2023. 6  
+[95] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. In ICLR, 2024. 6  
+[96] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. TIP, 30:7419-7431, 2021. 6  
+[97] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2  
+[98] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In CVPR, 2021. 2  
+[99] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 5

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f4c5a08d061405ca9ad8d651b43e49a667f874d12a3caa664b5d943b314f508
+size 944476

CVPR/2025/Adapting Text-to-Image Generation with Feature Difference Instruction for Generic Image Restoration/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e23f90b9eaa66edbd17a95b64c0b3399be2bbfd2f9801a2dae55d9ce5ae5203d
+size 517026

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/2aff36b8-d0fd-49cd-ab28-36f875e41f15_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32557d52d1330175d396272560eadfca55770ad2d3b1553fa06df8d22ff98f99
+size 78562

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/2aff36b8-d0fd-49cd-ab28-36f875e41f15_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d154b2d7c511b69e8ff4829931c63a52d3e63b348b8e8b05d2714a624aeea8f7
+size 95017

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/2aff36b8-d0fd-49cd-ab28-36f875e41f15_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cfb8a036f7d07b267f1e18ea36b3942b87be5772b79b6c9a71bca32f29ed6e2a
+size 795829

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/full.md

@@ -0,0 +1,339 @@
+# Adapting to Observation Length of Trajectory Prediction via Contrastive Learning
+
+Ruiqi Qiu
+
+2200891@stu.neu.edu.cn
+
+Siqi Luo
+
+2300925@stu.neu.edu.cn
+
+Jun Gong *
+
+gongjun@ise.neu.edu.cn
+
+Bowen Zhang
+
+2470807@stu.neu.edu.cn
+
+Xinyu Zhang
+
+2200981@stu.neu.edu.cn
+
+Yi Cen
+
+2370875@stu.neu.edu.cn
+
+Northeastern University, China
+
+# Abstract
+
+The ability to adapt to varying observation lengths is crucial for human trajectory prediction tasks, particularly in scenarios with limited observation lengths or missing data. Existing approaches mainly focus on introducing novel architectures or additional structural components, which substantially increase model complexity and present challenges for integration into existing models. We argue that current network architectures are sufficiently sophisticated to handle the Observation Length Shift problem, with the key challenge lying in improving feature representation for trajectories with limited lengths. To tackle this issue, we introduce a general and effective contrastive learning approach, called Contrastive Learning for Length Shift (CLLS). By incorporating contrastive learning during the training phase, our method encourages the model to extract length-invariant features, thus mitigating the impact of observation length variations. Furthermore, to better accommodate length adaptation tasks, we introduce a lightweight RNN network that, combined with CLLS, achieves state-of-the-art performance in both general prediction and observation length shift tasks. Experimental results demonstrate that our approach outperforms existing methods across multiple widely-used trajectory prediction datasets.
+
+# 1. Introduction
+
+Human trajectory prediction tasks involve forecasting potential future trajectories of individuals while considering various influential factors, with significant applications in autonomous driving, human-computer interaction, and intelligent security. Early attempts at human trajectory prediction relied on manually designed features based on prior knowledge [10, 20, 24]. Subsequently, data-driven models
+
+began to be widely applied in this field. Although these works have achieved notable performance, they mainly focus on idealized experimental conditions. In real-world scenarios, obtaining complete observable trajectories is often impractical under extreme circumstances, considering upstream tracking failures or unexpected events. Many existing datasets lack annotations for instantaneous pedestrians [40], and discard trajectory data with short durations during preprocessing, resulting in poor performance with limited observation lengths, potentially leading to significant safety risks in practical applications.
+
+While training data plays a crucial role in the Observation Length Shift problem, many studies still focus on structural improvements to accommodate more diverse observation lengths [14, 15, 22, 31, 36]. These methods require additional model structures or specialized networks to handle abnormal observation lengths and do not truly enable existing models to adapt to varying lengths.
+
+Previous work [4] has indicated that the last two frames before prediction contribute  $80.3\%$  and  $8.3\%$  respectively to prediction results. In most cases, relying solely on the last two frames of information is sufficient to make accurate predictions of future trajectories, as demonstrated by models specifically trained on limited observation length [14, 36]. We think that existing networks suffer from overfitting to complete observation trajectories, leading to a performance decline when face with varying observation lengths. Therefore, within existing network architecture, it becomes critical to ensure that the model can effectively extract length invariant information with different observation lengths.
+
+To address the above limitations, we propose a contrastive learning method applicable to different networks: Contrastive Learning for Length Shift (CLLS). This method enables models to achieve better performance under any limited observation length without additional inference
+
+structures. The method comprises data augmentation based on random length masking, nonlinear mapper, and contrastive loss. Since many network structures involve a time dimension, we use random length masking methods to generate input trajectories of different observable lengths. The generated masked trajectories, along with random rotations and scaling, form positive pairs for contrastive learning. Additionally, we propose a lightweight RNN network for Length Shift(RNLS), that selectively extracts meaningful time dimension information through iteration. Experimental results demonstrate that our method outperforms existing length adaptation approaches across various model types. Our proposed RNN network, combined with the CLLS method, achieves state-of-the-art performance across multiple datasets.
+
+As mentioned above, our contributions can be summarized as follows:
+
+- We propose a Contrastive Learning for Length Shift (CLLS) approach applicable to various models, improving their prediction performance for trajectories of different observation lengths.  
+- We introduce a lightweight RNN model for Length Shift, called RNLS, to better extract key temporal information selectively.  
+- We investigated how contrastive learning can enhance the predictive capabilities of models across varying observation lengths.
+
+# 2. RELATED WORK
+
+# 2.1. Trajectory Prediction
+
+Early trajectory prediction relied on manual feature design [10, 20, 24], but these methods struggled with complex social behaviors. Subsequently, RNN-based models [1, 8, 11, 19, 28, 35] gain widespread adoption in this field. Social LSTM [1] introduced recurrent networks for trajectory prediction, while Social GAN [8] used GANs for multimodal trajectory generation. Ivanovic and Pavone [11] applied CVAE to model pedestrian uncertainty, and Mangalam et al. [19] improved performance by incorporating pedestrian intent. To model social relationships, many studies used topological graphs with CNN [21, 29] or Attention [34, 37, 38] mechanisms. Social-STGCNN [21] used CNNs for relationship graphs, while SGCN [29] introduced sparse directed graphs to reduce redundancy. Recent work leverages large language models for trajectory prediction, such as Kuo et al. [12] utilize language intermediate representations for trajectory prediction, while Bae et al. [3] employ prompt-based methods to achieve numerical predictions for trajectory forecasting.
+
+# 2.2. Contrastive Learning in Trajectory Prediction
+
+Contrastive learning has been widely used in trajectory prediction. To address long-tail problems, Makansi et al. [18] differentiate positive and negative samples based on difficulty levels to address long-tail problems, while Wang et al. [33] employed clustering and specialized prediction heads for long-tail samples. Similarly, contrastive learning has been employed to enhance the model's ability to capture features of different motion patterns. Cheng et al. [6] treated same-person trajectories as positive samples and different-person trajectories as negative samples to learn latent motion habits. Halawa et al. [9] distinguish positive and negative samples based on action membership, promoting the learning of distinct action features. Liu et al. [17] used collision danger points as negative samples to reduce collisions. However, these methods are only applicable to high-dimensional trajectory features of equal length and do not address the problem under varying observation lengths.
+
+# 2.3. Trajectory Prediction for Length Shift
+
+Variable-length trajectory prediction has gained increasing attention in recent years. Monti et al. [22] employ knowledge distillation to transfer knowledge from networks trained on complete trajectories to student networks handling limited observation lengths. Sun et al. [31] incorporate scene semantic information under limited observation lengths and proposed a Mixture-of-Experts (MOE) architecture to enhance prediction performance under extremely limited observations. Li et al. [14] propose bidirectional diffusion to estimate past and future trajectories when observation length is limited. Xu and Fu [36] identify that position encoding and layer normalization are the root causes of length bias in Transformer structures and proposed using different subnetworks for predictions at different lengths. Li et al. [15] design a length-agnostic knowledge distillation (LaKD) module to dynamically transfer knowledge across trajectories. While these methods have contributed significantly to improving model performance under limited observation lengths, a truly unified model capable of adapting to arbitrary observation lengths has yet to emerge.
+
+# 3. Approach
+
+We define the trajectory prediction task as follows: In a scene with  $N$  agents, the observation sequence of length  $t_{o}$  for the i-th pedestrian is denoted as  $x^{(i)} = \{x_1^{(i)},\dots ,x_{t_o}^{(i)}\}$ , where  $x_{t}\in \mathbb{R}^{t_{o}\times d}$ . Our objective is to make the model's predicted sequence  $\hat{y}^{(i)}\in \mathbb{R}^{t_p\times 2}$  as close as possible to the ground truth  $y^{(i)}$ . Besides, in length shift task, we additionally defined the unobservable length as  $t_{unobs}$ .
+
+In the following, we first introduce Contrastive Learning for Length Shift (CLLS) in Section 3.1. This approach constructs positive sample pairs for contrastive learning
+
+![](images/f1671c978452a7acaa800265af0f54540df531f538e75568b81de4a23fec84d1.jpg)  
+Figure 1. An illustration of our Approach. (a) The CLLS method consists of data augmentation, nonlinear projector, and contrastive loss, with the encoder comes from the base model. The method is only executed during the training phase, meaning it incurs no additional inference cost. (b) RNLS comprises a trajectory encoder, social attention, feature predictor, and trajectory decoder. The target trajectory features output by the target encoder follow two paths: one flows through a projector to participate in contrastive loss computation, while the other is fused with the features output by the social attention via Add & Norm. This fused representation is then processed by the feature predictor and trajectory decoder to generate the final trajectory.
+
+through data augmentation techniques such as random length masking, rotation, and scaling, thereby achieving the extraction of length-invariant features. Subsequently, in Section 3.2, we propose A RNN Network for Length Shift (RNLS), which integrates CLLS into a lightweight RNN network.
+
+# 3.1. Contrastive Learning for Length Shift(CLLS)
+
+To enable a single model to adapt to varying observation lengths, we propose that an appropriate training methodology and dataset design can be employed. An intuitive idea involves utilizing contrastive learning to extract length-invariant features. Drawing inspiration from SimCLR [7], which was proposed for unsupervised image learning, we present a novel contrastive learning method for trajectory prediction called Contrastive Learning for Length Shift (CLLS). This method enhances the encoder's ability to extract invariant features from trajectories of different lengths. The core idea is to minimize the feature space discrepancy between differently augmented trajectories from the same sample. The method consists of several key components: Data Augmentation, Nonlinear Projector, and Contrastive Loss. The primary goal is to enable the model to learn consistent representations regardless of trajectory length variations.
+
+Data Augmentation Based on Length Shift. The key to contrastive learning is the construction of positive and negative samples. During the training phase, we employ a random length masking approach to construct limited ob
+
+servation trajectories. Specifically, starting from an original trajectory of length  $t_{o}$ , we mask  $t_{unobs}$  unobservable frames using the first observable frame. The augmented data is denoted as  $\tilde{x}$ , which can be expressed by the following formula:
+
+$$
+\left\{ \begin{array}{l l} \tilde {x} _ {t} = x _ {t _ {u n o b s} + 1}, & t = 1, \dots , t _ {u n o b s} \\ \tilde {x} _ {t} = x _ {t}, & t = t _ {u n o b s} + 1, \dots , t _ {o} \end{array} \right. \tag {1}
+$$
+
+where  $t_{unobs}$  is randomly selected from  $[1, t_o - 2]$ , and  $t_{unobs}$  denotes the length of unobservable frames. For different samples within a batch, random masking of varying lengths is applied. Additionally, for the same sample, different lengths are masked in each epoch. To enhance model performance, we additionally incorporated random rotation and scaling as data augmentation methods. In practical applications, for trajectories shorter than  $t_o$ , the first observable frame can be used to pad the sequence to achieve the required length of  $t_o$ .
+
+Nonlinear Projector maps the encoded feature representations to a space where the contrastive loss applied. Chen et al. [7] mentioned that the additional nonlinear projector can learn to be invariant to data transformation, so as to avoid the loss of key information due to comparative learning. We employ an MLP as a hidden layer to obtain  $z = \sigma (Wh)$ , where  $z\in \mathbb{R}^{t_{o}d}$ ,  $\sigma$  represents the ReLU activation function, and  $h$  denotes the encoded hidden layer features.
+
+Contrastive Loss. For a given trajectory, three augmented versions are generated using random length mask-
+
+ing, scaling, and rotation, forming a set of positive samples. Meanwhile, augmented trajectories from other trajectories within the same batch constitute the negative samples. A batch contains  $B$  samples. The augmented samples from the same observation trajectory are denoted as a positive sample set  $\Phi$ , while the other  $3(B - 1)$  samples serve as negative samples with respect to the target trajectory. Cosine similarity is used to measure the similarity between sample features, denoted as  $sim(z_i,z_j) = \frac{z_i^\top z_j}{||z_i||||z_j||}$ . The conatrastive loss function is defined as:
+
+$$
+\mathcal {L} _ {c l, \Phi} = - \log \frac {\sum_ {i , j \in \Phi , i \neq j} \exp \left(\operatorname {s i m} \left(z _ {i} , z _ {j}\right) / \tau\right)}{\sum_ {k = 1 , k \neq i} ^ {3 B} \exp \left(\operatorname {s i m} \left(z _ {i} , z _ {k}\right)\right) / \tau)} \tag {2}
+$$
+
+Here,  $\tau$  represents the temperature parameter, which controls the emphasis on hard samples and helps achieve a more uniform distribution in the feature space [32].
+
+# 3.2. A RNN Network For Length Shift (RNLS)
+
+To address the discrepancy between observation and prediction trajectory lengths, the main approaches include using RNN [28, 35] or Transformer [38] for autoregressive generation, or employing one-shot methods [2, 21, 29] that generate complete trajectories in a single shot. With the success of GPT series large language models[5, 23], autoregressive methods have demonstrated powerful capabilities in generative tasks with variable lengths. We think that RNN-based recursive prediction paradigms can better accommodate the Observation Length Shift problem. Limited observable sequences can enhance the model's comprehension of temporal relationships and focus more attention on adjacent time frames, thereby minimizing the impact of unobservable information on prediction outcomes. Consequently, we introduce an RNN network for Length Shift (RNLS).
+
+The model comprises four key modules: (1) a target trajectory encoder for encoding the target's historical trajectory; (2) a social attention module for modeling interactions between the target and neighboring pedestrians; (3) a future prediction module using RNNs to iteratively predict future features; and (4) a trajectory decoder for generating the final predicted trajectory, as illustrated in Fig. 1.
+
+Trajectory Encoder. For each frame in the observation sequence, in addition to the target's current coordinate position  $(p_t^x,p_t^y)$ , we incorporate the cosine value of the motion angle  $\theta_{t}$  to represent spatial orientation changes and the temporal index  $t$  to express temporal characteristics:
+
+$$
+x _ {t} = \left[ p _ {t} ^ {x}, p _ {t} ^ {y}, \cos \theta_ {t}, t \right] \tag {3}
+$$
+
+All temporal data, after normalization, is fed into a MLP layer with ReLU activation functions and encoded to obtain the  $i$ -th target trajectory feature encoding  $T_{target}^{(i)} \in \mathbb{R}^{t_o \times d}$ :
+
+$$
+T _ {t a r g e t} ^ {(i)} = M L P \left(x ^ {(i)}\right) \tag {4}
+$$
+
+Social Attention is utilized to model the interactions between the prediction target and its neighbors. Moreover, incorporating social relations can mitigate potential performance degradation in contrastive learning caused by similar target trajectories within the same batch. To model the relationship at time  $t$ , we follow previous work [35] and compute the euclidean distance  $d_t$  between neighbors and the target at time  $t$ , the cosine of the motion angle  $\alpha_t$ , the cosine of the azimuth angle  $\beta_t$ , and the sequence index  $t$  to form the input vector, which is then encoded through MLP:
+
+$$
+R _ {t} = \left[ d _ {t}, \cos \alpha_ {t}, \cos \beta_ {t}, t \right] \tag {5}
+$$
+
+$$
+T _ {n e i g h b o r} ^ {(i)} = M L P \left(R ^ {(i)}\right) \tag {6}
+$$
+
+Here,  $T_{\text{neighbor}}^{(i)} \in \mathbb{R}^{t_o \times n \times d}$ , where  $n$  is the number of observable neighbors. We followed the approach in [34] by selecting the nearest  $n$  pedestrians to the target as neighbors. To model the influence of different neighbors at different timestamps on the target, following the concept from [38], we flatten the neighbor and time dimensions to obtain  $T_{\text{neighbor}}^{(i)} \in \mathbb{R}^{t_o n \times d}$ . Using neighbor features as Key and Value, and target features as Query, we employ cross-attention to enable the Query to retrieve key features from different pedestrians at different timestamps, thoroughly exploring the relationships between pedestrian movements at different times:
+
+$$
+N ^ {(i)} = \operatorname {C r o s s A t t n} \left(T _ {\text {t a r g e t}} ^ {(i)}, T _ {\text {n e i g h b o r}} ^ {' (i)}\right) \tag {7}
+$$
+
+$$
+E ^ {(i)} = \operatorname {L a y e r N o r m} \left(N ^ {(i)} + T _ {\text {t a r g e t}} ^ {(i)}\right) \tag {8}
+$$
+
+Feature Predictor. The observed trajectory sequence features  $E_{t}^{(i)}$  are passed through the RNN's hidden layer  $h_{t}^{(i)}$  to iteratively transmit sequence information, then the  $h_{t_o}^{(i)}$  is fed into another RNN to iteratively generate future predicted sequences. Here, GRU modules are consistently used for all RNN operations:
+
+$$
+E _ {t, n e w} ^ {(i)}, h _ {t} ^ {(i)} = G R U \left(E _ {t - 1} ^ {(i)}, h _ {t - 1} ^ {(i)}\right), t = 1, 2, \dots , t _ {o} \tag {9}
+$$
+
+$$
+E _ {t} ^ {(i)}, h _ {t} = G R U \left(E _ {t - 1} ^ {(i)}, h _ {t - 1} ^ {(i)}\right), t = t _ {o} + 1, \dots , t _ {o} + t _ {p} \tag {10}
+$$
+
+where  $h_0$  is initialized as a zero matrix. It should be noted that the  $E_{t, new}^{(i)}$  output from the first GRU does not participate in the next iteration. The GRU only transmits historical feature information through the state matrix  $h_t$  at this stage. During the prediction phase, we directly use the previous iteration's output  $E_t^{(i)}$  as the input for the next  $E_{t+1}^{(i)}$ .
+
+Trajectory Decoder. The feature dimensions from the previous stage's output are directly decoded by MLP:
+
+$$
+\hat {y} ^ {\prime (i)} = M L P \left(E ^ {(i)}\right) \tag {11}
+$$
+
+This output  $\hat{y}^{(i)}\in \mathbb{R}^{t_p\times 2k}$  is reshaped to obtain the final multi-modal trajectory  $\hat{y}^{(i)}\in \mathbb{R}^{t_p\times k\times 2}$ .
+
+Loss Functions. The loss function consists of two parts: contrastive loss and prediction loss. First, the trajectory encoder output is passed through an MLP with ReLU activation function to compute the contrastive loss  $\mathcal{L}_{cl}$ . Then, the prediction loss function is calculated as the average of the L2 norm of the difference between predicted and actual coordinates along the temporal dimension. The complete loss function is obtained by summing these two components:
+
+$$
+\mathcal {L} _ {p r e d} = \frac {1}{3 B t _ {p}} \sum_ {i = 1} ^ {3 B} \sum_ {t = 1} ^ {t _ {p}} \| y _ {t} ^ {(i)} - \hat {y} _ {t} ^ {(i)} \| _ {2} \tag {12}
+$$
+
+$$
+\mathcal {L} = \mathcal {L} _ {p r e d} + \omega \mathcal {L} _ {c l} \tag {13}
+$$
+
+$\omega$  controls the ratio between the two loss functions. It should be noted that the prediction loss employs a winnertakes-all process [27], where only the result closest to the groundtruth will be back-propagated.
+
+# 4. Experiments
+
+# 4.1. Datasets and Evaluation Protocols.
+
+We evaluated our method on three public datasets: ETHUCY, SDD, and NBA, which include pedestrian trajectories, social interactions, and scene information across diverse urban settings.
+
+ETH-UCY [13, 25] is a standard benchmark for pedestrian trajectory prediction, comprising five video clips from four city scenes: eth, hotel, univ, and zara. The dataset uses real-world coordinates and fixed annotation intervals. We adopted the "leave-one-out" strategy for training and testing, consistent with prior works.
+
+SDD [26] is widely used for tasks like object detection, tracking, and trajectory prediction. It includes 60 bird's-eye view videos from Stanford University, featuring over 11,000 labeled individuals (pedestrians, bicycles, vehicles), 185,000 individual interactions, and 40,000 individual-scene interactions. Compared to ETH-UCY, SDD offers richer interactions and more complex scene structures.
+
+NBA [16, 39] presents a challenging benchmark with complex human interactions. It contains player coordinates at 25 FPS from the 2012-2013 NBA season, along with sports events like passing and shooting.
+
+We used the widely tested Average Displacement Error (ADE) and Final Displacement Error (FDE) to measure the performance of our network. Note that we used the best-of-20 validation method, which means that the prediction with the lowest ADE and FDE among 20 sampled predictions will be used to measure the accuracy of the model. We have:
+
+$$
+E r r _ {t} ^ {(i)} = \left\| p _ {t} ^ {(i)} - \hat {p} _ {t} ^ {(i)} \right\| \tag {14}
+$$
+
+$$
+A D E ^ {(i)} = \frac {1}{t _ {p}} \sum_ {t = t _ {o} + 1} ^ {t _ {o} + t _ {p}} E r r _ {t} ^ {(i)} \tag {15}
+$$
+
+$$
+F D E ^ {(i)} = \operatorname {E r r} _ {t _ {o} + t _ {p}} ^ {(i)} \tag {16}
+$$
+
+# 4.2. Implementation Details.
+
+Model Training. All MLPs in the RNLS are implemented with 2-layer FC networks, each followed by a ReLU activation function. The nonlinear projector in CLLS uses a single FC layer with ReLU activation. Following Xu et al. [35], we process data with 12 future frames and 8 observation frames. All networks are trained using the Adam optimizer, with feature dimension set to 128, the number of neighbors  $n$  set to 20, batch size of 512,  $\omega = 0.1$ , and  $\tau = 0.07$ . Experiments are conducted on a single RTX 3090 GPU. Isolated Training (IT) setting indicates training the model directly using specified limited observation lengths to predict the next 12 frames.
+
+# 4.3. Comparisons with Other Length Shift Methods.
+
+Instantaneous trajectory prediction. To thoroughly validate our proposed CLLS method, we compare against four: IT, DTO [22], MOE [31], and BCDiff [14]. Following previous work [14], we evaluate on Trajectory++ [28], PCC-SNet [30], SGCN [29], and SocialVAE [35] models with observation length set to 2 frames in the validation set. In the CLLS setup, we added a projector after the encoder of four models and incorporated the output features of augmented trajectories into the contrastive loss computation. This contrastive loss is subsequently added with the pre-existing loss functions to partake in the training process.
+
+As shown in Tab. 1, our method outperforms existing models on Trajectory++, SGCN, and SocialVAE for instantaneous observation length tasks, and surpasses the IT method on most datasets. It is noteworthy that our model did not perform well on PCCSNet, which may be attributed to: 1) interference from augmented samples in PCCSNet's memory bank structure, and 2) PCCSNet lacks explicit modeling of interactions among neighbors. Since other negative samples in the same batch may follow paths similar to the positive samples, the absence of social relations could lead to performance degradation. Additionally, our model performed poorly on the ZARA1 dataset in Trajectory++, possibly due to the left-turn intersection in the ZARA scene and Trajectory++'s preprocessing step of subtracting mean coordinates, which may disrupt the motion distribution during rotation augmentation.
+
+Comparison across varying observation lengths. To evaluate our model's performance across various observation lengths, we conducted comparative experiments with BCDiff [14] using the Trajectory++ [28] model, as shown in
+
+<table><tr><td>Model</td><td>Methods</td><td>ETH</td><td>HOTEi</td><td>UNIV</td><td>ZARA1</td><td>ZARA2</td><td>AVG</td><td>SDD</td></tr><tr><td rowspan="5">Trajectory++</td><td>IT</td><td>0.76/1.43</td><td>0.30/0.56</td><td>0.36/0.74</td><td>0.22/0.42</td><td>0.18/0.34</td><td>0.36/0.70</td><td>13.07/22.88</td></tr><tr><td>MOE</td><td>0.64/1.12</td><td>0.20/0.33</td><td>0.33/0.62</td><td>0.22/0.42</td><td>0.17/0.32</td><td>0.31/0.56</td><td>11.71/19.54</td></tr><tr><td>DTO</td><td>0.70/1.23</td><td>0.22/0.45</td><td>0.32/0.62</td><td>0.22/0.42</td><td>0.17/0.33</td><td>0.33/0.61</td><td>12.32/20.79</td></tr><tr><td>BCDiff</td><td>0.61/1.09</td><td>0.22/0.45</td><td>0.28/0.53</td><td>0.22/0.41</td><td>0.18/0.33</td><td>0.29/0.53</td><td>11.56/19.32</td></tr><tr><td>CLLS</td><td>0.57/0.83</td><td>0.12/0.18</td><td>0.25/0.47</td><td>0.25/0.51</td><td>0.17/0.33</td><td>0.27/0.46</td><td>10.05/17.42</td></tr><tr><td rowspan="5">PCCSNet</td><td>IT</td><td>0.34/0.65</td><td>0.14/0.25</td><td>0.31/0.63</td><td>0.23/0.46</td><td>0.16/0.37</td><td>0.24/0.47</td><td>9.19/17.71</td></tr><tr><td>MOE</td><td>0.31/0.57</td><td>0.13/0.21</td><td>0.25/0.53</td><td>0.20/0.41</td><td>0.14/0.31</td><td>0.20/0.41</td><td>8.40/16.08</td></tr><tr><td>DTO</td><td>0.33/0.64</td><td>0.14/0.24</td><td>0.31/0.62</td><td>0.22/0.46</td><td>0.15/0.35</td><td>0.23/0.46</td><td>8.93/16.92</td></tr><tr><td>BCDiff</td><td>0.30/0.56</td><td>0.13/0.20</td><td>0.25/0.52</td><td>0.18/0.37</td><td>0.14/0.31</td><td>0.19/0.39</td><td>8.32/15.87</td></tr><tr><td>CLLS</td><td>0.28/0.54</td><td>0.12/0.20</td><td>0.32/0.67</td><td>0.20/0.42</td><td>0.15/0.32</td><td>0.21/0.43</td><td>8.80/17.11</td></tr><tr><td rowspan="5">SGCN</td><td>IT</td><td>0.88/1.66</td><td>0.55/1.16</td><td>0.38/0.71</td><td>0.30/0.54</td><td>0.25/0.46</td><td>0.47/0.91</td><td>15.40/25.69</td></tr><tr><td>MOE</td><td>0.74/1.41</td><td>0.45/0.85</td><td>0.38/0.71</td><td>0.29/0.54</td><td>0.25/0.45</td><td>0.42/0.79</td><td>14.45/24.88</td></tr><tr><td>DTO</td><td>0.80/1.56</td><td>0.49/1.02</td><td>0.38/0.71</td><td>0.30/0.54</td><td>0.25/0.46</td><td>0.44/0.86</td><td>14.99/25.07</td></tr><tr><td>BCDiff</td><td>0.66/1.18</td><td>0.34/0.62</td><td>0.38/0.70</td><td>0.30/0.54</td><td>0.25/0.44</td><td>0.39/0.72</td><td>13.67/23.92</td></tr><tr><td>CLLS</td><td>0.68/1.20</td><td>0.30/0.53</td><td>0.36/0.69</td><td>0.28/0.52</td><td>0.24/0.44</td><td>0.37/0.68</td><td>11.57/19.64</td></tr><tr><td rowspan="5">SocialVAE</td><td>IT</td><td>0.64/1.10</td><td>0.21/0.34</td><td>0.27/0.51</td><td>0.22/0.39</td><td>0.18/0.34</td><td>0.30/0.54</td><td>9.56/16.10</td></tr><tr><td>MOE</td><td>0.57/1.01</td><td>0.17/0.29</td><td>0.26/0.44</td><td>0.22/0.36</td><td>0.17/0.32</td><td>0.28/0.48</td><td>9.12/14.98</td></tr><tr><td>DTO</td><td>0.61/1.06</td><td>0.18/0.31</td><td>0.25/0.43</td><td>0.22/0.38</td><td>0.17/0.33</td><td>0.29/0.50</td><td>9.28/15.58</td></tr><tr><td>BCDiff</td><td>0.53/0.91</td><td>0.17/0.27</td><td>0.24/0.40</td><td>0.21/0.37</td><td>0.16/0.26</td><td>0.26/0.44</td><td>9.05/14.86</td></tr><tr><td>CLLS</td><td>0.47/0.71</td><td>0.15/0.22</td><td>0.22/0.38</td><td>0.20/0.34</td><td>0.14/0.24</td><td>0.24/0.38</td><td>9.02/13.26</td></tr></table>
+
+Table 1. ADE/FDE metrics of different length-adaptive methods on the ETH-UCY and SDD datasets with an observation length of 2 frames. Bold means the best. IT represents the baseline model trained directly on 2 frames observations to predict the subsequent 12 frames.
+
+Tab. 2. Results demonstrate that our model consistently outperforms BCDiff across all observation lengths. Compared to BCDiff, our CLLS method maintains superior performance across all observation lengths while introducing no additional inference computation to the baseline model, truly achieving adaptation to arbitrary observation lengths with a single model.
+
+# 4.4. Comparisons with Other Models
+
+As shown in Tabs. 3 and 4, our model achieves state-of-the-art performance on the ETH, ZARA1, and ZARA2 datasets within the ETH-UCY benchmark, surpassing existing methods in terms of average metrics. On the more diverse SDD dataset, we exceed the second-best performance by  $7.7\%$  in ADE. For the NBA dataset, our model surpasses the second-best performance by  $9.7\% / 12.4\%$  in Rebound-ing and  $4.7\% / 6.8\%$  in Scoring, demonstrating strong generalization capabilities across multiple datasets.
+
+# 4.5. Ablation Study
+
+Selection of projector. As shown in Tab. 5, we evaluated several projector: 1) no proj: using encoded features directly for contrastive loss; 2) linear: using linear projector
+
+via Fc layer; 3) MLP: single FC layer with ReLU activation; 4) CNN: aggregating features through convolution kernels; 5) Self-Attention: self-attention mechanism. While no projection shows optimal performance under complete observation length, MLP and Self-Attention outperform the no-projection setting as unobservable length increases.
+
+Impact of different data augmentation methods. We investigated the impact of different data augmentation methods, as illustrated in Tab. 6. Length masking proves crucial among the three augmentation methods, with its absence leading to significant performance degradation in both baseline performance and length adaptation. Combining either rotation or scaling with length masking significantly improves the model's predictive capability across different observation lengths. The combination of all three methods yields further improvements, surpassing the IT training setting.
+
+CLLS Hyperparameter Analysis. Tab. 7 presents experimental results for hyperparameters  $\omega$  and  $\tau$  in the contrastive loss. Variations in  $\alpha$  affect the model's emphasis on prediction accuracy versus length-invariant feature extraction tasks, with this distinction becoming more pronounced as the unobservable length increases. Additionally, accord-
+
+<table><tr><td>Dataset</td><td>Methods</td><td>Tunobs = 1</td><td>Tunobs = 2</td><td>Tunobs = 3</td><td>Tunobs = 4</td><td>Tunobs = 5</td><td>Tunobs = 6</td></tr><tr><td rowspan="3">ETH/UCY</td><td>Trajectory++ &amp; BCDiff</td><td>0.31/0.55</td><td>0.30/0.54</td><td>0.30/0.53</td><td>0.29/0.53</td><td>0.32/0.57</td><td>0.34/0.60</td></tr><tr><td>Trajectory++ &amp; CLLS</td><td>0.30/0.51</td><td>0.30/0.50</td><td>0.29/0.50</td><td>0.29/0.49</td><td>0.28/0.48</td><td>0.28/0.48</td></tr><tr><td>RNLS &amp; CLLS</td><td>0.20/0.33</td><td>0.20/0.33</td><td>0.21/0.33</td><td>0.21/0.34</td><td>0.21/0.35</td><td>0.22/0.36</td></tr><tr><td rowspan="3">SDD</td><td>Trajectory++ &amp; BCDiff</td><td>11.86/19.78</td><td>11.81/19.71</td><td>11.68/19.50</td><td>11.56/19.32</td><td>11.92/20.01</td><td>12.33/20.98</td></tr><tr><td>Trajectory++ &amp; CLLS</td><td>9.94/17.10</td><td>9.96/17.16</td><td>10.00/17.24</td><td>10.08/17.34</td><td>9.93/17.09</td><td>10.05/17.42</td></tr><tr><td>RNLS &amp; CLLS</td><td>7.44/11.42</td><td>7.41/11.39</td><td>7.41/11.46</td><td>7.48/11.55</td><td>7.62/11.92</td><td>8.02/12.70</td></tr></table>
+
+Table 2. ADE/FDE metrics on the ETH-UCY and SDD datasets with varying observation lengths. Bold means the best.  $T_{\text{unobs}}$  means unobservation lengths.  
+
+<table><tr><td>Models</td><td>ETH</td><td>HOTEL</td><td>UNIV</td><td>ZARA1</td><td>ZARA2</td><td>AVG</td><td>SDD</td></tr><tr><td>Social GAN</td><td>0.60/1.19</td><td>0.52/1.02</td><td>0.44/0.84</td><td>0.22/0.43</td><td>0.29/0.58</td><td>0.41/0.81</td><td>27.3/41.4</td></tr><tr><td>Social-STGCNN</td><td>0.64/1.11</td><td>0.49/0.85</td><td>0.44/0.79</td><td>0.34/0.53</td><td>0.30/0.48</td><td>0.44/0.75</td><td>20.8/33.2</td></tr><tr><td>PECNet</td><td>0.54/0.87</td><td>0.18/0.24</td><td>0.35/0.60</td><td>0.22/0.39</td><td>0.17/0.30</td><td>0.29/0.48</td><td>10.0/15.9</td></tr><tr><td>Trajectory++</td><td>0.55/0.98</td><td>0.21/0.34</td><td>0.35/0.64</td><td>0.22/0.42</td><td>0.16/0.31</td><td>0.30/0.54</td><td>11.4/20.1</td></tr><tr><td>AgentFormer</td><td>0.45/0.75</td><td>0.14/0.22</td><td>0.25/0.45</td><td>0.18/0.30</td><td>0.14/0.24</td><td>0.23/0.39</td><td>8.7/14.9</td></tr><tr><td>SocialVAE</td><td>0.41/0.58</td><td>0.13/0.19</td><td>0.21/0.36</td><td>0.17/0.29</td><td>0.13/0.22</td><td>0.21/0.33</td><td>8.1/11.7</td></tr><tr><td>EqMotion</td><td>0.40/0.61</td><td>0.12/0.18</td><td>0.23/0.43</td><td>0.18/0.32</td><td>0.13/0.23</td><td>0.21/0.35</td><td>7.9/11.9</td></tr><tr><td>LED</td><td>0.39/0.58</td><td>0.11/0.17</td><td>0.26/0.43</td><td>0.18/0.26</td><td>0.13/0.22</td><td>0.21/0.33</td><td>8.5/11.7</td></tr><tr><td>LMTraj-SUP</td><td>0.41/0.51</td><td>0.12/0.16</td><td>0.22/0.34</td><td>0.20/0.32</td><td>0.17/0.27</td><td>0.22/0.32</td><td>7.8/10.1</td></tr><tr><td>RNLS &amp; CLLS (Ours)</td><td>0.38/0.55</td><td>0.12/0.18</td><td>0.23/0.41</td><td>0.17/0.29</td><td>0.13/0.21</td><td>0.20/0.32</td><td>7.2/11.0</td></tr></table>
+
+Table 3. ADE/FDE metrics of existing methods on the ETH-UCY and SDD datasets under the full 8-frame observation length.  
+
+<table><tr><td></td><td>Rebounding</td><td>Scoring</td></tr><tr><td>Trajectory++</td><td>0.98/1.93</td><td>0.73/1.46</td></tr><tr><td>BiTraP</td><td>0.83/1.72</td><td>0.74/1.49</td></tr><tr><td>SGNet-ED</td><td>0.78/1.55</td><td>0.58/1.30</td></tr><tr><td>SocialVAE</td><td>0.72/1.37</td><td>0.64/1.17</td></tr><tr><td>RNLS &amp; CLLS (Ours)</td><td>0.65/1.20</td><td>0.61/1.09</td></tr></table>
+
+Table 4. ADE/FDE metrics of existing methods on the NBA dataset under the full 8-frame observation length.  
+
+<table><tr><td>Methods</td><td>Tunobs = 0</td><td>Tunobs = 2</td><td>Tunobs = 4</td><td>Tunobs = 6</td></tr><tr><td>no proj</td><td>7.17/10.99</td><td>7.20/11.09</td><td>7.38/11.40</td><td>8.23/12.94</td></tr><tr><td>linear</td><td>7.20/10.95</td><td>7.26/11.25</td><td>7.31/11.37</td><td>8.15/12.87</td></tr><tr><td>MLP</td><td>7.22/10.98</td><td>7.24/11.16</td><td>7.31/11.50</td><td>7.97/12.54</td></tr><tr><td>CNN</td><td>7.23/11.17</td><td>7.33/11.47</td><td>7.43/11.76</td><td>8.03/12.94</td></tr><tr><td>Attention</td><td>7.19/11.13</td><td>7.21/11.22</td><td>7.36/11.53</td><td>7.97/12.71</td></tr></table>
+
+ing to [32] changes in  $\tau$  influence the model's attention to difficult samples. Considering ADE/FDE across different lengths, we selected  $\omega = 0.1$  and  $\tau = 0.07$  as the final
+
+Table 5. Ablation studies of nonlinear projector on the SDD dataset.  $T_{unobs}$  means unobservation lengths. Lower is better.  
+
+<table><tr><td>Methods</td><td>Tunobs = 0</td><td>Tunobs = 2</td><td>Tunobs = 4</td><td>Tunobs = 6</td></tr><tr><td>IT</td><td>7.64/11.87</td><td>7.71/12.14</td><td>7.76/12.24</td><td>7.98/12.68</td></tr><tr><td>w/o mask</td><td>7.59/11.92</td><td>10.62/17.69</td><td>14.04/23.56</td><td>44.47/75.38</td></tr><tr><td>w/o resize</td><td>7.49/11.62</td><td>7.61/11.70</td><td>7.63/11.80</td><td>8.23/13.03</td></tr><tr><td>w/o rotate</td><td>7.57/11.84</td><td>7.47/11.79</td><td>7.61/11.96</td><td>8.16/12.90</td></tr><tr><td>Full</td><td>7.25/11.16</td><td>7.27/11.27</td><td>7.33/11.44</td><td>8.05/12.08</td></tr></table>
+
+Table 6. Ablation studies of data augmentation methods on the SDD dataset.  $T_{unobs}$  means unobservation lengths. IT means the baseline model trained directly on  $t_o - T_{unobs}$  frames observations to predict the subsequent 12 frames.
+
+model parameters.
+
+# 4.6. Qualitative Analysis
+
+T-SNE Visualization for Feature Distribution. We proposed CLLS to enable extraction of length-invariant features under varying observation lengths. To validate this, we applied T-SNE dimensionality reduction to trajectory encoding features for different observation lengths, denoted as  $z_{\text{unobs},t_{\text{unobs}}} \in \mathbb{R}^{B \times 2}$ , where  $B$  represents the dataset size. We visualized the differences between these features and those with unobservable length 6 ( $z_{\text{unobs},6}$ ), as shown in Fig. 2. Compared to the original Src model, our CLLS
+
+![](images/9b9470639ffc9f3a808564e2165063b3791ef285c886d75f6cfc8056592f4b90.jpg)  
+(a)  $z_{unobs,0} - z_{unobs,6}$
+
+![](images/8a27b4d42ddb7e5ec0dba6da3e31ce0222979b11e8dda694f6b81cae6499a513.jpg)  
+(b)  $z_{unobs,2} - z_{unobs,6}$  
+Figure 2. Scatter plots of T-SNE dimensionality reduction on encoding feature differences with various observation lengths on the SDD dataset.  $z_{\text{unobs},t_{\text{unobs}}}$  represents the encoded features after projector of trajectories with unobservable length  $t_{\text{unobs}}$ . CLLS and Src represent the original RNLS model and the model enhanced by contrastive learning.
+
+![](images/6fea96ba97ae7f7d2b02ed331c407be3138a41c17c555e46889c4f374535aa37.jpg)  
+(c)  $z_{unobs,4} - z_{unobs,6}$
+
+<table><tr><td>ω</td><td>τ</td><td>Tunobs=0</td><td>Tunobs=2</td><td>Tunobs=4</td><td>Tunobs=6</td></tr><tr><td>10</td><td>0.07</td><td>7.29/11.15</td><td>7.29/11.20</td><td>7.41/11.57</td><td>8.04/12.73</td></tr><tr><td>1</td><td>0.07</td><td>7.21/11.16</td><td>7.27/11.27</td><td>7.33/11.44</td><td>8.05/12.08</td></tr><tr><td>0.1</td><td>0.07</td><td>7.22/10.98</td><td>7.24/11.16</td><td>7.31/11.50</td><td>7.97/12.54</td></tr><tr><td>0.01</td><td>0.07</td><td>7.13/11.05</td><td>7.18/11.22</td><td>7.28/11.39</td><td>8.04/12.58</td></tr><tr><td>0.1</td><td>0.01</td><td>7.15/11.08</td><td>7.22/11.31</td><td>7.39/11.58</td><td>8.00/12.69</td></tr><tr><td>0.1</td><td>0.1</td><td>7.16/11.03</td><td>7.24/11.18</td><td>7.35/11.30</td><td>7.98/12.58</td></tr><tr><td>0.1</td><td>0.5</td><td>7.16/11.20</td><td>7.15/11.31</td><td>7.25/11.41</td><td>7.91/12.64</td></tr></table>
+
+Table 7. Ablation experiments on hyperparameters of the contrastive loss on the SDD dataset.  $T_{unobs}$  means unobservation lengths. Lower is better.
+
+![](images/ccb002636e92e8b5f72a561505a30ab33e1a618975ed9b2d1a8d2f6716681e12.jpg)  
+(a) Src Model.  
+Figure 3. Line plots of mean encoding features along the temporal dimension for trajectories with different unobservation lengths in the SDD dataset.  $CLLS$  and  $Src$  represents the original RNLS model and the model enhanced by contrastive learning.
+
+![](images/86e01b1cb2fb815d0d286f74cf9b3dd5f7a281eaab9a26049aa05a14f6618ce9.jpg)  
+(b) CLLS Model.
+
+model demonstrates features more closely aligned with the x-axis zero point, indicating that CLLS helps extract more consistent features across different observation lengths, facilitating adaptation to the Length Shift problem.
+
+What Improves the Model's Adaptability to Different Lengths? We analyzed the cause by examining differences in temporal feature encoding across various unobservable trajectory lengths. Specifically, we computed the average feature values at each frame for both the original Src model
+
+and our CLLS model. As illustrated in Fig. 3, in the Src model, feature values gradually decrease for unobservable frames, showing an inflection point at the first observable frame, followed by increasing feature values. The CLLS model exhibits two characteristics: 1) It assigns higher feature values to the first observable frame and the subsequent frame, as evidenced by sudden changes in frames 5 and 6 for unobs 4; 2) After the first two observable frames, features show a decline before rising again. This indicates that through contrastive learning, the model pays additional attention to truncation positions.
+
+# 4.7. Limitations.
+
+The contrastive learning approach introduces certain invariance in encoder feature extraction, potentially sacrificing some fine-grained features and affecting basic model performance. Besides, length-based, scaling, and rotation data augmentation methods are ineffective for map-based data, which is somewhat constrained by the model's processing approach.
+
+# 5. Conclusion
+
+This paper introduces CLLS, a contrastive learning method for the Length Shift problem. It constructs positive samples via random length masking, rotation, and scaling, using encoded features from a nonlinear projector for contrastive loss calculation. We also propose a lightweight RNN combining MLP and GRU with CLLS, achieving SOTA performance on ETH-UCY, SDD, and NBA datasets. Experiments demonstrate CLLS's superiority over existing length-adaptive methods in both instantaneous and multi-observation prediction tasks. Ablation studies validate the nonlinear projector design, data augmentation strategies, and hyperparameter selection. Feature visualization further demonstrates CLLS's effectiveness. We hope our research provides valuable insights for addressing length shift problems.
+
+# References
+
+[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 961–971, 2016. 2  
+[2] Inhwan Bae and Hae-Gon Jeon. A set of control points conditioned pedestrian trajectory prediction. AAAI, 37(5):6155-6165, 2023. 4  
+[3] Inhwan Bae, Junoh Lee, and Hae-Gon Jeon. Can language beat numerical regression? language-based multimodal trajectory prediction, 2024. 2  
+[4] Stefan Becker, Ronny Hug, Wolfgang Hubner, and Michael Arens. Red: A simple but effective baseline predictor for the trajnet benchmark. In Computer Vision - ECCV 2018 Workshops: Munich, Germany, September 8-14, 2018, Proceedings, Part III, pages 138-153, 2019. 1  
+[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners, 2020. 4  
+[6] Guangyi Chen, Junlong Li, Nuoxing Zhou, Liangliang Ren, and Jiwen Lu. Personalized trajectory prediction via distribution discrimination, 2022. 2  
+[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations, 2020. 3  
+[8] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2255–2264, 2018. 2  
+[9] Marah Halawa, Olaf Hellwich, and Pia Bideau. Action-based contrastive learning for trajectory prediction, 2022. 2  
+[10] Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics. Phys. Rev. E, 51(5):4282-4286, 1995. 1, 2  
+[11] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2375-2384, 2019. 2  
+[12] Yen-Ling Kuo, Xin Huang, Andrei Barbu, Stephen G. McGill, Boris Katz, John J. Leonard, and Guy Rosman. Trajectory prediction with linguistic representations. In 2022 International Conference on Robotics and Automation (ICRA), pages 2868-2875, 2022. 2  
+[13] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. Computer Graphics Forum, 26(3):655-664, 2007. 5
+
+[14] Rongqing Li, Changsheng Li, Dongchun Ren, Guangyi Chen, Ye Yuan, and Guoren Wang. Bcdiff: Bidirectional consistent diffusion for instantaneous trajectory prediction. In Advances in Neural Information Processing Systems, pages 14400-14413, 2023. 1, 2, 5  
+[15] Yuhang Li, Changsheng Li, Ruilin Lv, Rongqing Li, Ye Yuan, and Guoren Wang. Lakd: Length-agnostic knowledge distillation for trajectory prediction with any length observations. In Advances in Neural Information Processing Systems, pages 28720–28744, 2024. 1, 2  
+[16] Kostya Linou, Dzmitryi Linou, and Martijn de Boer. Nba-player-movements. https://github.com/linouk23/NBAPlayer-Movements, 2016. 5  
+[17] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations, 2021. 2  
+[18] Osama Makansi, Özgün Cicek, Yassine Marrakchi, and Thomas Brox. On exposing the challenging long tail in future prediction of traffic actors, 2021. 2  
+[19] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction, 2020. 2  
+[20] Ramin Mehran, Alexis Oyama, and Mubarak Shah. Abnormal crowd behavior detection using social force model. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 935-942, 2009. 1, 2  
+[21] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction, 2020. 2, 4  
+[22] Alessio Monti, Angelo Porrello, Simone Calderara, Pasquale Coscia, Lamberto Ballan, and Rita Cucchiara. How many observations are enough? knowledge distillation for trajectory forecasting, 2022. 1, 2, 5  
+[23] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback, 2022. 4  
+[24] S. Pellegrini, A. Ess, K. Schindler, and L. van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th International Conference on Computer Vision, pages 261-268, 2009. 1, 2  
+[25] S. Pellegrini, A. Ess, K. Schindler, and L. van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th International Conference on Computer Vision, pages 261-268, 2009. 5  
+[26] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning social etiquette: Human trajectory understanding in crowded scenes. In Computer Vision - ECCV 2016, pages 549-565, 2016. 5  
+[27] Christian Rupprecht, Iro Laina, Robert DiPietro, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D. Hager. Learning in an uncertain world: Representing ambiguity through multiple hypotheses. In Proceedings of the
+
+IEEE International Conference on Computer Vision (ICCV), 2017. 5  
+[28] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data, 2021. 2, 4, 5  
+[29] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgen:sparse graph convolution network for pedestrian trajectory prediction. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8990-8999, 2021. 2, 4, 5  
+[30] Jianhua Sun, Yuxuan Li, Hao-Shu Fang, and Cewu Lu. Three steps to multimodal trajectory prediction: Modality clustering, classification and synthesis. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13230-13239, 2021. 5  
+[31] Jianhua Sun, Yuxuan Li, Liang Chai, Hao-Shu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6457-6466, 2022. 1, 2, 5  
+[32] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2495-2504, 2021. 4, 7  
+[33] Yuning Wang, Pu Zhang, Lei Bai, and Jianru Xue. Fend: A future enhanced distribution-aware contrastive learning framework for long-tail trajectory prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1400-1409, 2023. 2  
+[34] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In Computer Vision - ECCV 2022, pages 682-700, 2022. 2, 4  
+[35] Pei Xu, Jean-Bernard Hayet, and Ioannis Karamouzas. Socialvae: Human trajectory prediction using timewise latents. In Computer Vision - ECCV 2022, pages 511-528, 2022. 2, 4, 5  
+[36] Yi Xu and Yun Fu. Adapting to length shift: Flexilength network for trajectory prediction, 2024. 1, 2  
+[37] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Computer Vision - ECCV 2020, pages 507-523, 2020. 2  
+[38] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M. Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9813-9823, 2021. 2, 4  
+[39] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE International Conference on Data Mining, pages 670-679, 2014. 5  
+[40] Xian Zhong, Xu Yan, Zhengwei Yang, Wenxin Huang, Kui Jiang, Ryan Wen Liu, and Zheng Wang. Visual exposes you: Pedestrian trajectory prediction meets visual intention. IEEE Transactions on Intelligent Transportation Systems, 24(9): 9390-9400, 2023. 1

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00be3f5cab32d7bbcb32a1fce41e9bc163d5113ce872d3033cfcffa3b988eb25
+size 752962

CVPR/2025/Adapting to Observation Length of Trajectory Prediction via Contrastive Learning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee9312e6d31b7e78f6e1dd64c02d354c1f029e58a49d550c9ace05b138eab947
+size 365231

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/e3af0000-7a87-46f4-897e-0e5ba81f78d7_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f28f0d3a1957ee2496bedfc0714d37bce6a454b353a7e40ce257208ca880f681
+size 74652

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/e3af0000-7a87-46f4-897e-0e5ba81f78d7_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c07c38edd9e3580ce222d50a0bf405d805a63e2a81c229c577f80cdc7785219d
+size 89957

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/e3af0000-7a87-46f4-897e-0e5ba81f78d7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ef1a2a1307afa12d517b2c72472d7043dff46609f0d75698d9407c41ca70352
+size 4384884

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/full.md

@@ -0,0 +1,326 @@
+# Adapting to the Unknown: Training-Free Audio-Visual Event Perception with Dynamic Thresholds
+
+Eitan Shaar $^{1*}$   
+Ariel Shaulov $^{2*}$   
+Gal Chechik $^{1,3}$   
+Lior Wolf $^{2}$
+
+$^{1}$ Bar-Ilan University  $^{2}$ Tel-Aviv University  $^{3}$ NVIDIA
+
+# Abstract
+
+In the domain of audio-visual event perception, which focuses on the temporal localization and classification of events across distinct modalities (audio and visual), existing approaches are constrained by the vocabulary available in their training data. This limitation significantly impedes their capacity to generalize to novel, unseen event categories. Furthermore, the annotation process for this task is labor-intensive, requiring extensive manual labeling across modalities and temporal segments, limiting the scalability of current methods. Current state-of-the-art models ignore the shifts in event distributions over time, reducing their ability to adjust to changing video dynamics. Additionally, previous methods rely on late fusion to combine audio and visual information. While straightforward, this approach results in a significant loss of multimodal interactions. To address these challenges, we propose AudioVisual Adaptive Video Analysis  $(AV^2A)$ , a model-agnostic approach that requires no further training and integrates a score-level fusion technique to retain richer multimodal interactions.  $AV^2A$  also includes a within-video label shift algorithm, leveraging input video data and predictions from prior frames to dynamically adjust event distributions for subsequent frames. Moreover, we present the first training-free, open-vocabulary baseline for audio-visual event perception, demonstrating that  $AV^2A$  achieves substantial improvements over naive training-free baselines. We demonstrate the effectiveness of  $AV^2A$  on both zero-shot and weakly-supervised state-of-the-art methods, achieving notable improvements in performance metrics over existing approaches. Our code is available on Github.
+
+# 1. Introduction
+
+Video event recognition, which involves the temporal localization and classification of events using both audio and
+
+![](images/89c528b87b17d498a85ae9385ada4aa0a9c6e3369fcfbbcce4de42d4b272325c.jpg)  
+Figure 1. Overview of the AVVP task. Audio-visual event perception focuses on predicting the temporal boundaries of events within a video that are exclusively visible (shown in blue), exclusively audible (shown in red), or both audible and visible (shown in purple).
+
+visual cues (see Figure 1), is essential for a wide range of applications, including video content analysis, autonomous driving, multimedia indexing, and personalized content recommendation systems [5, 27, 30]. Despite the recognized importance of this task, current approaches face substantial limitations that hinder their effectiveness and adaptability in open-world scenarios.
+
+A primary limitation in previous state-of-the-art methods [2, 5, 10, 12, 18, 19, 21] is their inability to dynamically adjust to evolving events in the video, leading to an oversight of changes in event distribution over time. This issue stems from the reliance on fixed thresholds for event prediction, which restricts the model's capacity to adapt to evolving patterns and contextual shifts within the video sequence. Moreover, the current state-of-the-art methods often employ a late fusion strategy, where audio and visual information are processed independently and then merged by intersecting the predicted labels from both modalities to obtain audio-visual predictions. While this approach simplifies model design, it limits the potential to capture the rich multimodal interactions needed for complex event detection [5, 18].
+
+Another major limitation of existing methodologies [2, 5, 10, 12, 18, 19, 21] is their reliance on fixed-vocabulary datasets. This reliance significantly restricts these mod
+
+els' ability to generalize to novel or evolving event categories—a capability that is essential for open-world environments where the range of possible event types is inherently diverse and continuously evolving [2, 30]. Addressing this challenge requires developing models that can move beyond pre-defined vocabularies and adapt to new events without extensive re-training. Lastly, the annotation process is labor-intensive, time-consuming, and costly, making it difficult to scale existing approaches to larger datasets or to adapt them for open-world deployment [12].
+
+To address these challenges, we propose a model-agnostic approach requiring no additional training, termed Audio-Visual Adaptive Video Analysis  $(\mathrm{AV}^2\mathrm{A})$ . Our method integrates a score-level fusion mechanism that combines audio and visual cues at an initial processing stage, thereby preserving richer multimodal interactions essential for accurate and context-aware event recognition. This approach enhances the alignment of audio and visual information, improving the model's capacity to recognize complex event patterns. Additionally,  $\mathrm{AV}^2\mathrm{A}$  incorporates a novel within-video label-shift mechanism designed to operate within video contexts, leveraging both input video data and predictions from prior frames to dynamically adjust event distributions for subsequent frames, achieved via dynamic thresholds that adapt over time. By aligning model expectations with recurring or evolving events, this mechanism enables  $\mathrm{AV}^2\mathrm{A}$  to capture temporal dependencies. In scenarios where events frequently recur or evolve, the dynamic thresholds capability ensures that the model remains adaptable and responsive to changing contexts. Furthermore, by leveraging foundation models [4, 22, 37] we introduce the first training-free, open-vocabulary baseline through  $\mathrm{AV}^2\mathrm{A}$ . This approach circumvents the limitations of fixed vocabularies and reduces the need for additional data annotations. Our primary contributions are as follows:
+
+- We present  $\mathrm{AV}^2\mathrm{A}$ , a model-agnostic approach that incorporates a score-level fusion strategy to combine audio and visual cues, preserving critical multimodal interactions. Additionally, our approach applies a within video label-shift mechanism that dynamically adjusts the predicted event distribution based on prior context, enhancing temporal adaptability.  
+- We introduce the first training-free approach specifically tailored for open-world audio-visual event perception, capable of recognizing and adapting to unseen events without the constraints of a fixed vocabulary.  
+- Through comprehensive evaluations on the AVE [27] and LLP [28] datasets, we demonstrate that our approach achieves state-of-the-art results in both zero-shot and training-free settings, showcasing its robustness and versatility.  
+- We illustrate the efficacy of our label-shift mechanism through experimental applications on existing state-of
+
+the-art weakly supervised methods, demonstrating consistent improvements across all metrics without additional training. These findings underscore the potential of leveraging temporal patterns directly within existing models.
+
+# 2. Related Work
+
+Weakly-supervised Audio-Visual Event Perception. To fully exploit the potential of both audio and visual modalities for weakly-supervised video analysis, [27] introduce the audio-visual event (AVE) localization task to leverage both modalities for weakly-supervised video understanding. In AVE, the model detects when an event is both audible and visible, pinpointing its temporal boundaries. Approaches often use attention strategies [11, 31, 32, 34] or cross-modal interactions [14, 23-25], while others incorporate regularization like background suppression [33] and positive sample propagation [36]. Unlike AVE, the audio-visual video parsing (AVVP) task [28] accounts for misalignment between audio and visual signals. To address this, multimodal multiple instance learning (MMIL [28]) frameworks have been employed [2, 5, 10, 12, 18, 19, 21], enabling the model to identify relevant instances across modalities without the need for strict temporal synchronization. These methods typically include techniques such as audio-visual track swapping [30] and cross-modality enhancement [12]. Currently, most methods target either AVE or AVVP separately due to the different properties between the two tasks. Furthermore, all of the above methods are restricted to a closed vocabulary, requiring extensive training to recognize events, which limits their ability to generalize to new, unseen events.
+
+Audio & Visual Zero-Shot Methods. Audio-visual zero-shot learning was first introduced in [20] with the Coordinated Joint Multimodal Embedding (CJME) model, which aligns video, audio, and text in a shared feature space using triplet loss. Building on this, [15] developed the Audio-Visual Generalized Zero-shot Learning Network (AVGZSLNet), which employs a cross-modal decoder and composite triplet loss to enhance the alignment of audio and video embeddings with text. The Audio-Visual Cross-Attention (AVCA) framework [17] further facilitates information exchange between modalities, achieving state-of-the-art performance in audio-visual zero-shot classification and introducing three novel benchmarks for AV-ZSL. Additionally, the Temporal Cross-Attention Framework (TCAF) [16] extends AVCA by leveraging temporal information. While [3] shows promising signs for zero-shot audio-visual event perception, the field remains largely underexplored, with significant opportunities for deeper exploration and advancement.
+
+![](images/c785d5e1641193fd919a3855dfcadc27d746bc8d133c2927ac7ad068a7a4cfd4.jpg)  
+Figure 2. Overview of  $\mathrm{AV}^2\mathrm{A}$ . The process begins with category selection, where the input video clip passes through a video-level score-level fusion module (blue) to select relevant categories based on a threshold  $\tau_{f}$ . These categories guide segment-level score-level fusion, where a dynamic threshold module (orange) updates thresholds  $\pmb{\tau}_{t}$  via our label-shift technique, using the soft confusion matrix  $M$  from prior predictions  $\mathcal{Y}_1,\dots ,\mathcal{Y}_{t - 1}$  and segment scores  $P_{av}^{1},\ldots ,P_{av}^{t - 1}$ , cosine similarity between segments and  $P_{av}^{t}$ . Finally, predicted candidates are validated against a confidence threshold  $\tau_r$ , retaining only those above it. This figure illustrates the process for audio-visual events; audio and visual events are handled similarly.
+
+Label Shift Estimation. Label shift estimation refers to the task of inferring the label distribution in a target domain using labeled source data and unlabeled samples from the target domain [6]. Earlier methods [1, 26, 35] often rely on explicit modeling of the conditional probability  $P(X = x \mid Y = y)$ , a requirement that becomes impractical for high-dimensional data, such as images. More recently [13] proposed BBSE, a method that leverages a black box predictor  $\hat{f}$  to estimate label shift between source and target distributions. Using predictions from  $\hat{f}$ , BBSE constructs a confusion matrix  $M = p_{s}(\hat{y}, y) \in \mathbb{R}^{k \times k}$  on source data with  $k$  classes, where  $y$  is the true label and  $\hat{y}$  is the predicted label, which may be a hard label ( $\hat{y} = \arg \max \hat{f}(x)$ ) or a soft probability vector,  $\hat{f}(x) = [p(\hat{y} = 1|x), \dots, p(\hat{y} = k|x)]$ , which is the distribution over classes predicted by  $\hat{f}$  for a given input  $x$ . The label shift estimate  $\hat{w}$  is then calculated as  $\hat{w} = M^{-1}\hat{\mu}$ , where  $M$  is the estimated confusion matrix and  $\hat{\mu}$  represents the distribution of predicted labels in the target data, derived by applying  $\hat{f}$  to the target samples. This correction factor  $\hat{w}$  aligns the target label distribution with the source distribution. In contrast to prior approaches, we explore the integration of input data as a component within our label shift estimation method. To the best of our knowledge, this work is the first to adapt label shift estimation within video contexts.
+
+# 3. Method
+
+We present a framework for audio-visual event perception that is entirely training-free. Central to our approach is a dy
+
+namic thresholds mechanism, which adjusts threshold values over time-based on the input video data and previous segment predictions, see Figure 2. The method enables precise detection of audio, visual, and audio-visual events.
+
+Problem Setup And Notations. Given an input video divided into  $T$  non-overlapping segments, our objective is to temporally localize and identify audio, visual, and audio-visual events present for each segment  $1 \leq t \leq T$ . Each segment  $t$  can contain zero or more events, which makes this a multi-label problem. The events to be identified are drawn from a set of  $C$  possible categories, defined for each benchmark. The Given video consists of audio  $A$  and video frames  $V$ :
+
+$$
+\begin{array}{l} A = \left\{a _ {1}, a _ {2} \dots , a _ {T} \right\}, \quad V = \left\{v _ {1}, v _ {2}, \dots , v _ {T} \right\} \\ a _ {t} \in \mathbb {R} ^ {1 \times F}, v _ {t} \in \mathbb {R} ^ {3 \times H \times W} \\ \end{array}
+$$
+
+where  $a_{t}$  and  $v_{t}$  are the corresponding audio and video visual frame at segment  $t$ ,  $H$  and  $W$  are the height and width for each frame, and  $F$  is number of frequency bins for each audio snippet  $a_{t}$ .
+
+We define two types of vector scores: video-level vector scores and segment-level vector scores. Both types are calculated by a foundation model (FM), combined with a sigmoid function  $\sigma$  to address the multi-label classification problem. Our method operates in an open-vocabulary setting, where the label set  $C$  is defined per video at inference time, even ones not seen during hyper-parameter tuning, unlike the baselines which are restricted to labels from
+
+a predefined set. Instead,  $C$  consists of user-defined textual descriptions, such as "an image of a military helicopter" or "the sound of helicopter rotor blades".
+
+Video-level scores are represented by  $P_{a}$  and  $P_{v}$  for audio and visual data, from the entire video's audio  $A$  and visual  $V$ .
+
+$$
+P _ {a} = \sigma (F M (C, A)), \quad P _ {v} = \sigma (F M (C, V)) \tag {1}
+$$
+
+$$
+P _ {a}, P _ {v} \in [ 0, 1 ] ^ {| C |}
+$$
+
+Segment-level vector scores capture more localized information. These are denoted as  $P_{a}^{t}$  and  $P_{v}^{t}$  for audio and visual scores, respectively, in each segment  $t$  based on segment-specific audio  $a_{t}$  and video frame  $v_{t}$ .
+
+$$
+P _ {a} ^ {t} = \sigma (F M (C, a _ {t})), \quad P _ {v} ^ {t} = \sigma (F M (C, v _ {t})) \tag {2}
+$$
+
+$$
+P _ {a} ^ {t}, P _ {v} ^ {t} \in [ 0, 1 ] ^ {| C |}
+$$
+
+Each element in both types of score vectors represents the score that a specific category  $c$  occurs.
+
+Effective audio-visual event recognition is based on the ability to leverage complementary information from both audio and visual modalities. Our score-level fusion is straightforward, we integrate audio and video scores using a weighted sum to create an audio-visual representation.
+
+$$
+P _ {a v} = \alpha P _ {a} + (1 - \alpha) P _ {v} \tag {3}
+$$
+
+$$
+P _ {a v} ^ {t} = \alpha P _ {a} ^ {t} + (1 - \alpha) P _ {v} ^ {t} \tag {4}
+$$
+
+$$
+P _ {a v}, P _ {a v} ^ {t} \in [ 0, 1 ] ^ {| C |}
+$$
+
+where  $P_{av}, P_{av}^{t}$  are the audio-visual vector scores at the video-level and at segment-level  $t$  respectively, and  $\alpha \in [0,1]$  serves as a scalar weight that determines the contribution of the audio score  $P_{a}, P_{a}^{t}$  relative to the visual score  $P_{v}, P_{v}^{t}$ .
+
+For generality, we will focus on identifying and localizing audio-visual events from this point forward. It is important to note that the processes for localizing and identifying audio and visual events are conducted in the same manner, using  $\alpha = 1$  for audio events and  $\alpha = 0$  for visual events in Equations 3 and 4.
+
+In our framework, we first want to ensure that only the relevant event categories are retained. We apply a selection step for the video-level vector  $P_{av}$ , discarding categories that do not exceed a predefined threshold  $\tau_f$ . This ensures us that only the most relevant event categories  $\hat{C}_{av}$  are preserved for further analysis.
+
+$$
+\hat {\mathcal {C}} _ {a v} = \left\{c \mid P _ {a v} [ c ] > \tau_ {f}, \quad \forall c \in C \right\}, \tag {5}
+$$
+
+where  $P_{av}[c]$  represents the audio-visual classification score for category  $c$  is being presented in the video.
+
+<table><tr><td>Method</td><td>α</td><td>τ0</td><td>τr</td><td>τf</td><td>λ</td></tr><tr><td>AV2A (CLIP [22] + CLAP [4])</td><td>0.45</td><td>0.75</td><td>0.75</td><td>0.5</td><td>1.0</td></tr><tr><td>AV2A (LanguageBind [37])</td><td>0.5</td><td>0.75</td><td>0.75</td><td>0.55</td><td>2.5</td></tr><tr><td>AV2A (MGN-MA [18])</td><td>0.3</td><td>0.55</td><td>-</td><td>-</td><td>0.5</td></tr><tr><td>AV2A (JoMoLD [2])</td><td>0.4</td><td>0.5</td><td>-</td><td>-</td><td>0.5</td></tr><tr><td>AV2A (CMPAE [5])</td><td>0.45</td><td>0.5</td><td>-</td><td>-</td><td>0.5</td></tr></table>
+
+Table 1. Hyperparameters for each baseline using  ${\mathrm{{AV}}}^{2}\mathrm{\;A}$  .
+
+To adapt to the content of the video domain, we apply thresholds for each category  $c \in \hat{\mathcal{C}}_{av}$  that change over time. We initialize the threshold for each category with the same predefined value, denoted as  $\tau^0$ , which serves as the initial threshold at segment  $t = 1$ .
+
+$$
+\boldsymbol {\tau} _ {t = 1} = \left[ \begin{array}{c c} \tau^ {0} & \forall c \in \hat {\mathcal {C}} _ {a v} \end{array} \right]
+$$
+
+Then, we predict events for segment  $t$  in the following way:
+
+$$
+\mathcal {Y} _ {t} = \left[ y _ {c} ^ {t} \mid \forall c \in \hat {\mathcal {C}} _ {a v} \right] \tag {6}
+$$
+
+$$
+\text {W h e r e} \quad y _ {c} ^ {t} = \left\{ \begin{array}{l l} 1 & \text {i f} P _ {a v} ^ {t} [ c ] > \boldsymbol {\tau} _ {t} [ c ], \\ 0 & \text {o t h e r w i s e} \end{array} \right.
+$$
+
+where  $P_{av}^{t}[c]$  represents the audio-visual score of category  $c$  being presented at segment  $t$  and  $\pmb{\tau}_{t}[c]$  is the category threshold for category  $c$  at segment  $t$ .
+
+To modify our dynamic thresholds, we estimate for each segment  $t$  the ratio  $\hat{w}_t$  between the predictions generated from segments  $1, \ldots, t-1$  and those from segment  $t$  across each category  $c$ .
+
+To perform this estimation, we first define a soft confusion matrix  $M$ , where each entry represents the score that a given segment  $t$  has the true category  $c$  but it's predicted category is  $\hat{c}$ , averaged over all samples. Let  $M_p \in \mathbb{R}^{(t-1) \times |\hat{C}_{av}|}$  be the matrix of prediction scores where each row is given by  $P_{av}^t$  (Eq. 4), which is the prediction score vector for segment  $t$ . We also define  $M_y \in \mathbb{R}^{(t-1) \times |\hat{C}_{av}|}$  as the matrix of the predicted categories where each row  $\mathcal{V}_t$  is predictions for segment  $t$ . The soft confusion matrix  $M \in \mathbb{R}^{|\hat{C}_{av}| \times |\hat{C}_{av}|}$  is defined as:
+
+$$
+M = \frac {1}{t - 1} \left(M _ {p}\right) ^ {T} \cdot M _ {y},
+$$
+
+i.e., each entry  $M_{\hat{c},c}$  in  $M$  is given by:
+
+$$
+M _ {\hat {c}, c} = \frac {1}{t - 1} \sum_ {i = 1} ^ {t - 1} P _ {a v} ^ {i} [ \hat {c} ] \cdot y _ {c} ^ {i}.
+$$
+
+$M_{\hat{c},c}$  captures the average score up to segment  $t - 1$  for instances where the model's predicted score for the category  $\hat{c}$  aligns with segments where the predicted category is  $c$ . Additionally, we propose incorporating the input data as a
+
+<table><tr><td></td><td colspan="2">Audio</td><td colspan="2">Visual</td><td colspan="2">Audio Visual</td><td colspan="2">Type@AV</td><td colspan="2">Event@AV</td></tr><tr><td></td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td></tr><tr><td rowspan="2">CLIP [22] + CLAP [4] with AV2A</td><td>17.6</td><td>15.1</td><td>18.8</td><td>17.8</td><td>30.0</td><td>25.6</td><td>22.1</td><td>19.5</td><td>18.3</td><td>16.2</td></tr><tr><td>32.7 ± .01</td><td>28.5 ± .01</td><td>43.5 ± .01</td><td>41.5 ± .01</td><td>55.1 ± .06</td><td>48.2 ± .04</td><td>43.8 ± .04</td><td>39.4 ± .01</td><td>34.7 ± .01</td><td>31.2 ± .01</td></tr><tr><td rowspan="2">LanguageBind [37] with AV2A</td><td>20.3</td><td>17.5</td><td>21.3</td><td>20.0</td><td>24.3</td><td>21.0</td><td>22.0</td><td>19.5</td><td>20.8</td><td>18.4</td></tr><tr><td>40.9 ± .09</td><td>35.9 ± .12</td><td>57.4 ± .04</td><td>54.4 ± .09</td><td>59.1 ± .1</td><td>52.3 ± .13</td><td>52.4 ± .04</td><td>47.5 ± .04</td><td>43.4 ± .04</td><td>38.9 ± .06</td></tr></table>
+
+Table 2. Comparison of training-free methods on the LLP dataset, reporting AVVP metrics.
+
+scaling factor for  $\hat{w}_t$ . We use cosine similarity between the visual features of the current segment  $t$ , and the previous segment  $t - 1$  as our scaling factor. The estimated ratio  $\hat{w}_t$  is computed as follows:
+
+$$
+\hat {w} _ {t} = M ^ {- 1} P _ {a v} ^ {t} \cdot \frac {x _ {t} \cdot x _ {t - 1}}{\| x _ {t} \| \| x _ {t - 1} \|}, \tag {7}
+$$
+
+where  $M$  is the estimated soft confusion matrix,  $P_{av}^{t}$  is the score vector for segment  $t$  and  $x_{t}$  is denoted as the corresponding visual feature vector produced from the foundation model. The estimated ratio  $\hat{w}_{t}$ , computed independently for each test video at inference time, quantifies the change in the score for each category  $c$  between consecutive segments  $t - 1$  and  $t$ . Specifically, the ratio captures how the score vector  $P_{av}^{t}$  for the current segment  $t$  has shifted relative to the previous segments. The cosine similarity serves as a scaling factor, providing a measure of alignment between the feature vectors of adjacent segments. This ensures that if two segments are highly similar in content, any score changes are emphasized less, while more significant score changes in dissimilar segments are weighted more heavily. Thus, the ratio  $\hat{w}_{t}$  not only reflects changes in category scores but also incorporates segment continuity, offering a refined measure of temporal dynamics in the event patterns across segments.
+
+Following this, we update the category thresholds  $\pmb{\tau}_{t}$  for each segment  $t$ . This update employs an exponential decay function to gradually reduce the thresholds, ensuring that the adjustments become smaller over time and preventing excessive reductions as the process progresses. Here,  $\pmb{\tau}_{1}$  represents the initial threshold vector, while  $\mathbf{z}$  is a vector tracking the cumulative count of occurrences for each category  $c \in \hat{C}_{av}$  up to time  $t - 1$ . The exponential decay,  $e^{-\lambda \mathbf{z}}$  is applied element-wise, meaning that each component  $z_{i}$  in  $\mathbf{z}$  undergoes a separate transformation  $e^{-\lambda z_i}$ , allowing each category's threshold adjustment to be scaled individually according to its occurrence frequency. Here,  $\lambda$  is the decay constant in the exponential decay function, controlling the rate at which threshold adjustments decrease.
+
+$$
+\boldsymbol {\tau} _ {t} = \boldsymbol {\tau} _ {t - 1} - \boldsymbol {\tau} _ {1} e ^ {- \lambda \mathbf {z}} \cdot \hat {w} _ {t} \tag {8}
+$$
+
+$$
+\boldsymbol {\tau} _ {t}, \mathbf {z}, w _ {t} \in \mathbb {R} ^ {1 \times | \hat {\mathcal {C}} _ {a v} |},
+$$
+
+After obtaining all  $\mathcal{Y}_1, \ldots, \mathcal{Y}_T$ , we define a candidate  $S_k$ , which is the predicted audio-visual localization for event category  $c$ , where  $k$  represents the index of the multiple predicted candidates.  $S_k$  is constructed to capture the maximal-length interval during which the score  $P_{av}^t[c]$  remains above  $\pmb{\tau}_t[c]$  to each segment  $t$ , meaning that  $y_c^t = 1$  holds true continuously throughout the entire candidate duration, such as:
+
+$$
+S _ {k} = \left(i _ {c} ^ {n}, j _ {c} ^ {n}, c\right) \quad \text {s u c h t h a t} \quad \exists n \leq T, \exists c \in \hat {\mathcal {C}} _ {a v} ,
+$$
+
+Where:
+
+$$
+i _ {c} ^ {n} = \min \{t \mid y _ {c} ^ {t} = 1, t > j _ {c} ^ {n - 1} \}
+$$
+
+$$
+j _ {c} ^ {n} = \max  \left\{t \mid y _ {c} ^ {t} = 1, t \geq i _ {c} ^ {n} \right\}
+$$
+
+$$
+\text {s u c h} \quad y _ {c} ^ {t} = 1, \quad \forall t \in \left[ i _ {c} ^ {n}, j _ {c} ^ {n} \right] \quad .
+$$
+
+Here,  $i_c^n$  and  $j_c^n$  denote the start and end indices of the candidate  $S_k$ , respectively, and  $c$  represents the event category associated with the candidate  $S_i$  and  $n$  is the number of candidates for the same category  $c$ .
+
+To further refine the predicted candidates, each candidate  $S_{k}$  is validated based on its associated audio  $A'$  and video frames  $V'$ , where:
+
+$$
+A ^ {\prime} = \{a _ {i}, \dots , a _ {j} \}, \quad V ^ {\prime} = \{v _ {i}, \dots , v _ {j} \}.
+$$
+
+Here,  $i$  and  $j$  represent the start and end timestamps of the predicted candidate. The candidate scores are computed as:
+
+$$
+P _ {a} ^ {\prime} = \sigma \left(F M \left(\hat {\mathcal {C}} _ {a v}, A ^ {\prime}\right)\right), \quad P _ {v} ^ {\prime} = \sigma \left(F M \left(\hat {\mathcal {C}} _ {a v}, V ^ {\prime}\right)\right) \tag {9}
+$$
+
+where  $P_{a}^{\prime}$  and  $P_{v}^{\prime}$  denote the audio and visual scores for all categories  $\hat{C}_{av}$  between timestamps  $i$ ,  $j$ . We compute the audio-visual score  $P_{av}^{\prime}$ , which is computed as described in Equation 3 with  $P_{a}^{\prime}$  and  $P_{v}^{\prime}$ . We apply a confidence threshold,  $\tau_{r}$ , to the audio-visual score  $P_{av}^{\prime}[c]$  for the predicted category  $c$  associated with a candidate. If  $P_{av}^{\prime}[c]$  surpasses  $\tau_{r}$ , the candidate is deemed valid; otherwise, it is discarded. Our framework employs a total of  $T \cdot |\hat{C}_{av}| + 2$  thresholds, specifically  $\{\tau_1, \dots, \tau_T\}, \tau_f$ , and  $\tau_r$ . The output of our framework consists of these valid candidates.
+
+<table><tr><td></td><td colspan="2">Audio</td><td colspan="2">Visual</td><td colspan="2">Audio-Visual</td><td colspan="2">Type@AV</td><td colspan="2">Event@AV</td></tr><tr><td></td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td></tr><tr><td>MGN-MA [18]</td><td>60.7</td><td>51.0</td><td>55.5</td><td>52.4</td><td>50.6</td><td>44.4</td><td>55.6</td><td>49.3</td><td>57.2</td><td>49.2</td></tr><tr><td>with AV2A&#x27;s dynamic thresholds</td><td>60.8</td><td>51.0</td><td>55.6</td><td>52.7</td><td>51.9</td><td>45.5</td><td>56.1</td><td>49.7</td><td>57.3</td><td>49.3</td></tr><tr><td>JoMoLD [2]</td><td>61.3</td><td>53.9</td><td>63.8</td><td>59.9</td><td>57.2</td><td>49.6</td><td>60.8</td><td>54.5</td><td>59.9</td><td>52.5</td></tr><tr><td>with AV2A&#x27;s dynamic thresholds</td><td>61.4</td><td>53.9</td><td>64.1</td><td>61.0</td><td>57.7</td><td>50.9</td><td>61.1</td><td>55.2</td><td>60.0</td><td>52.7</td></tr><tr><td>CMPAE [5]</td><td>64.4</td><td>57.5</td><td>66.1</td><td>63.3</td><td>59.5</td><td>52.4</td><td>63.3</td><td>57.7</td><td>62.9</td><td>56.3</td></tr><tr><td>with AV2A&#x27;s dynamic thresholds</td><td>64.7</td><td>58.0</td><td>66.3</td><td>63.4</td><td>59.9</td><td>52.7</td><td>63.6</td><td>58.0</td><td>63.0</td><td>56.7</td></tr></table>
+
+Table 3. Performance of state-of-the-art weakly supervised methods and their improvements with  ${\mathrm{{AV}}}^{2}\mathrm{\;A}$  ’s dynamic thresholds mechanism on the LLP dataset.  
+
+<table><tr><td>Method</td><td>Accuracy(%)</td></tr><tr><td>DG-SCT [3]</td><td>64.7</td></tr><tr><td rowspan="2">CLIP [22] + CLAP [4] with AV²A</td><td>38.2 ± 2.87</td></tr><tr><td>68.5 ± 0.16</td></tr><tr><td rowspan="2">LanguageBind [37] with AV²A</td><td>32.9 ± 1.27</td></tr><tr><td>72.8 ± 0.14</td></tr></table>
+
+Table 4. Comparison with zero-shot methods on AVE datasets, reporting AVEL metrics.
+
+# 4. Experiments
+
+We assess the performance of  $\mathrm{AV^2A}$  on two widely adopted datasets, AVE [27] and LLP [28], through comprehensive experiments demonstrating its improvements in both training-free and zero-shot settings. Additionally, we apply our novel dynamic thresholds technique to state-of-the-art weakly supervised methods, achieving significant performance enhancements without additional training. An extensive ablation study further highlights the contribution of each component in our approach.
+
+# 4.1. Experimental setup
+
+Datasets. We evaluate our approach on two datasets, AVE [27] and LLP [28].
+
+The Audio-Visual Event (AVE) dataset [27] consists of 4,143 YouTube videos spanning 28 categories derived from AudioSet [7], with each video clip lasting 10 seconds. The AVE [27] dataset focuses on recognizing audiovisual events that are both visible and audible across multiple time segments within a video, that is to say, it provides synchronized audio and visual signals, along with precise frame-level annotations that indicate the temporal locations of these events.
+
+The Look, Listen, and Parse (LLP) dataset [28] is a comprehensive audio-visual benchmark designed for studying event perception in unconstrained, real-world videos.
+
+It comprises 11,849 YouTube video clips, each lasting 10 seconds and covering 25 diverse event categories that involve various audio-visual interactions. The dataset provides video-level event labels, including both single and multiple overlapping events. The LLP dataset suits weakly-supervised learning, providing video-level annotations without precise event boundaries.
+
+Evaluation Metrics. Consistent with previous work, we evaluate using F1-scores for audio, visual, and audio-visual events at both segment and event levels. Additionally, we report aggregate metrics, Type@AV and Event@AV, computed at segment and event levels as well. For a comprehensive explanation of these metrics, refer to [28]. For the AVE dataset, in line with [27], we use overall accuracy as the evaluation metric.
+
+For hyperparameter optimization, we performed a search over 200 samples from the LLP validation set for the training-free methods, while using the full validation set of LLP for the weakly-supervised methods. Since only  $\mathrm{AV}^2\mathrm{A}$  's dynamic threshold mechanism is applied, optimizing the  $\tau_{r}$  and  $\tau_{f}$  values is unnecessary for the MGN-MA [18], JoMoLD [2], and CMPAE [5] methods. All hyperparameters used for each baseline are listed in Table 1.
+
+# 4.2. Comparison with State-of-the-art Methods
+
+Baselines. In our approach to training-free compression, we employ three foundational models to create two training-free baseline methods, upon which we implement  $\mathrm{AV^2A}$ . The first baseline employs the LanguageBind model [37], which provides both text-visual and text-audio similarity measurements. The second baseline combines the CLIP [22] and CLAP [4] models, where CLIP captures text-visual similarities, and CLAP supplies text-audio similarities; we refer to this baseline as CLIP-CLAP. For the evaluation of the CLIP-CLAP and LanguageBind baselines, a single static threshold is applied to predict categories across segments, creating two naive baselines. For weakly-supervised compression, we adopt three methodologies: JoMoLD [2],
+
+![](images/d103a532068cd21fdfa29f5df213d3342e4836188f58d2fdde787961736473df.jpg)  
+(a)  
+Figure 3. Performance analysis of  $\mathrm{AV^2A}$ , based on the LanguageBind [37] model, showcasing predictions on audio-visual events, specifically those occurring simultaneously in both audio and video. Comparisons highlight (a) improvements and (b) failure-cases relative to state-of-the-art weakly supervised baselines.
+
+![](images/91cc9de8215cd3d5a41604792e85c6e9b139982a84ab8ea0dedeeb12f62a22ee.jpg)  
+(b)
+
+which performs joint modality learning for improved cross-modal representations; MGN-MA [18], which leverages multi-granularity networks to enhance modality alignment; and the state-of-the-art method CMPAE [5], that integrates adaptive encoding to capture complex, high-quality cross-modal associations.
+
+AVVP Evaluation. In this experiment, we employ two baselines, CLIP-CLAP and LanguageBind, which are constructed as described in Section 4.2. As shown in Table 2, applying  $\mathrm{AV}^2\mathrm{A}$  to the baselines CLIP-CLAP and LanguageBind results in metric improvements across the board. Specifically, when  $\mathrm{AV}^2\mathrm{A}$  is integrated with the naive CLIP-CLAP baseline, it achieves a  $21.7\%$  improvement in segment-level Type@AV, which measures average segment performance across audio, visual, and audio-visual settings. Additionally,  $\mathrm{AV}^2\mathrm{A}$  enhances performance by  $19.9\%$  in
+
+event-level Type@AV, which evaluates the average event performance across these setups. Furthermore, AV $^2$ A improves the naive LanguageBind baseline by  $30.4\%$  and  $28\%$  in segment-level and event-level Type@AV, respectively.
+
+We further validate  $\mathrm{AV}^2\mathrm{A}$  's dynamic thresholds mechanism on the AVVP benchmark by applying it on top of current state-of-the-art weakly supervised methods, specifically, CMPAE [5], JoMoLD [2], and MGN-MA [18]. In this experiment, we apply the dynamic thresholds mechanism (Eq. 8) to these methods, as they already employ a relevant class selection stage that closely aligns with our approach; however, these methods used static thresholds to predict events over time. The results, presented in Table 3, indicate that our dynamic thresholds outperform the static thresholds technique without additional training, establishing a new state-of-the-art in video parsing. To imple
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Audio</td><td colspan="2">Visual</td><td colspan="2">Audio Visual</td><td colspan="2">Type@AV</td><td colspan="2">Event@AV</td></tr><tr><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td><td>segment</td><td>event</td></tr><tr><td>\( AV^2A(LanguageBind [37]) \)</td><td>40.9</td><td>35.9</td><td>57.4</td><td>54.4</td><td>59.1</td><td>52.3</td><td>52.4</td><td>47.5</td><td>43.4</td><td>38.9</td></tr><tr><td>w/o cosine similarity</td><td>40.8</td><td>35.9</td><td>57.3</td><td>54.3</td><td>58.8</td><td>51.8</td><td>52.3</td><td>47.3</td><td>43.3</td><td>38.8</td></tr><tr><td>w/o dynamic thresholds</td><td>40.7</td><td>35.1</td><td>56.9</td><td>55.2</td><td>55.1</td><td>47.2</td><td>52.2</td><td>47.2</td><td>43.0</td><td>37.8</td></tr><tr><td>w/o refine segments</td><td>38.9</td><td>34.1</td><td>53.6</td><td>50.9</td><td>57.9</td><td>51.1</td><td>50.2</td><td>45.4</td><td>41.4</td><td>37.0</td></tr><tr><td>w/o relevant class selection</td><td>33.4</td><td>29.3</td><td>40.2</td><td>38.4</td><td>39.7</td><td>34.3</td><td>37.8</td><td>34.0</td><td>34.8</td><td>31.4</td></tr></table>
+
+Table 5. Ablation study results on the LLP dataset.
+
+ment our dynamic threshold mechanism in the audio-visual regime, it is essential to apply our score-level fusion stage (Eq.4), as the audio-visual scores required for the dynamic thresholds mechanism are unavailable in the methods standard form due to their late-fusion strategy.
+
+AVE Evaluation. As shown in Table 4, we compare  $\mathrm{AV}^2\mathrm{A}$  with three zero-shot baselines: the CLIP-CLAP baseline [4, 22], the LanguageBind baseline [37], and the state-of-the-art DG-SCT method [3]. Notably,  $\mathrm{AV}^2\mathrm{A}$  outperforms DG-SCT [3], achieving improvements of  $3.8\%$  and  $8.1\%$  when built upon the CLIP-CLAP and Language-Bind baselines, respectively. To ensure a fair zero-shot comparison, we used the same hyperparameters identified during tuning on the LLP validation set, as detailed in Section 4.2. Please note that we did not apply  $\mathrm{AV}^2\mathrm{A}$  on the DG-SCT baseline, as the trained checkpoints and features were not made available.
+
+Method Efficiency. In our experiments, inference times across all weakly-supervised methods were of the same order of magnitude, with our method achieving  $\sim 2.49$  seconds per video. However, our primary efficiency advantage lies in the fact that  $\mathrm{AV^2A}$  requires no additional training, unlike the weakly-supervised baselines. For comparison, the weakly-supervised methods trainable parameters are as follows: MGN-MA [18] has  $\sim 4.3M$ , JoMoLD [2] has  $\sim 4.5M$ , and CMPAE [5] has  $\sim 5.6M$ . In addition to the trainable parameters, the above methods employ three pretrained, fixed models to generate input features: ResNet152 [8], VGGish [9], and R(2+1)D [29]. This procedure adds  $\sim$  2.68 seconds to the inference time per video.
+
+# 4.3. Ablation study
+
+We conducted an ablation study using the LanguageBind baseline [37] to investigate the contributions of each component of our method. The results, detailed in Table 5, illustrate the impact of these components on performance. The first row presents our complete method, with the LanguageBind model [37]. The second row of the table represents the model configuration without cosine similarity in the label shift method, effectively demonstrating the method's per
+
+formance when excluding the cosine similarity between adjacent frames. The third row highlights the effect of omitting the dynamic threshold mechanism, employing fixed thresholds instead. Additionally, the fourth row reports the outcome when the refinement of segments is removed, resulting in candidate event segments being used as final predictions. Lastly, we examined the model without the relevant class selection process, using the entire class set from the outset. Note that in each row, only the specified stage is pruned, while the remainder of the pipeline remains unchanged. The findings indicate that each component of our method contributes positively to the overall performance, reinforcing the importance of their inclusion.
+
+# 5. Limitations
+
+Our reliance on foundation models for extracting audiovisual similarities may introduce biases and limit the accuracy of short-duration or rapid events (Figure 3). These models, trained on longer, distinct events, struggle with subtle changes. Since  $\mathrm{AV}^2\mathrm{A}$  requires no training or fine-tuning, improvements in foundation models should enhance audiovisual perception.
+
+# 6. Conclusion
+
+In this paper, we introduce the first training-free, open-vocabulary approach for audio-visual perception via  $\mathrm{AV}^2\mathrm{A}$ , a model-agnostic method for robust event perception.  $\mathrm{AV}^2\mathrm{A}$  improves weakly-supervised baselines by replacing static thresholds with a within-video label shift mechanism, enhancing adaptability to temporal variations in event distribution. Applied to zero-shot and weakly-supervised settings,  $\mathrm{AV}^2\mathrm{A}$  achieves new state-of-the-art results without additional training.
+
+# 7. Acknowledgments
+
+This work was supported by the Tel Aviv University Center for AI and Data Science (TAD), and by a Vatat data-science grant to GC and the Bar-Ilan University data-science institute.
+
+# References
+
+[1] Yee Seng Chan and Hwee Tou Ng. Word sense disambiguation with distribution estimation. In IJCAI, pages 1010-5, 2005. 3  
+[2] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In European Conference on Computer Vision, pages 431-448. Springer, 2022. 1, 2, 4, 6, 7, 8  
+[3] Haoyi Duan, Yan Xia, Zhou Mingze, Li Tang, Jieming Zhu, and Zhou Zhao. Cross-modal prompts: Adapting large pretrained models for audio-visual downstream tasks. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 8  
+[4] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2, 4, 5, 6, 8  
+[5] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18827-18836, 2023. 1, 2, 4, 6, 7, 8  
+[6] Saurabh Garg, Yifan Wu, Sivaraman Balakrishnan, and Zachary Lipton. A unified view of label shift estimation. Advances in Neural Information Processing Systems, 33:3290-3300, 2020. 3  
+[7] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 6  
+[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 8  
+[9] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 ieee international conference on acoustics, speech and signal processing (icassp), pages 131-135. IEEE, 2017. 8  
+[10] Xun Jiang, Xing Xu, Zhiguo Chen, Jingran Zhang, Jingkuan Song, Fumin Shen, Huimin Lu, and Heng Tao Shen. Dhhn: Dual hierarchical hybrid network for weakly-supervised audio-visual video parsing. In Proceedings of the 30th ACM International Conference on Multimedia, pages 719–727, 2022. 1, 2  
+[11] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In Proceedings of the Asian Conference on Computer Vision, 2020. 2  
+[12] Yan-Bo Lin, Hung-Yu Tseng, Hsin-Ying Lee, Yen-Yu Lin, and Ming-Hsuan Yang. Exploring cross-video and cross-modality signals for weakly-supervised audio-visual video
+
+Parsing. Advances in Neural Information Processing Systems, 34:11449-11461, 2021. 1, 2  
+[13] Zachary Lipton, Yu-Xiang Wang, and Alexander Smola. Detecting and correcting for label shift with black box predictors. In International conference on machine learning, pages 3122-3130. PMLR, 2018. 3  
+[14] Tanvir Mahmud and Diana Marculescu. Ave-clip: Audioclip-based multi-window temporal transformer for audio visual event localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5158–5167, 2023. 2  
+[15] Pratik Mazumder, Pravendra Singh, Kranti Kumar Parida, and Vinay P Namboodiri. Avgzslnet: Audio-visual generalized zero-shot learning by reconstructing label features from multi-modal embeddings. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3090–3099, 2021. 2  
+[16] Otniel-Bogdan Mercea, Thomas Hummel, A Sophia Koepke, and Zeynep Akata. Temporal and cross-modal attention for audio-visual zero-shot learning. In European Conference on Computer Vision, pages 488-505. Springer, 2022. 2  
+[17] Otniel-Bogdan Mercea, Lukas Riesch, A Koepke, and Zeynep Akata. Audio-visual generalised zero-shot learning with cross-modal attention and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10553–10563, 2022. 2  
+[18] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. Advances in Neural Information Processing Systems, 35: 34722-34733, 2022. 1, 2, 4, 6, 7, 8  
+[19] Shentong Mo and Yapeng Tian. Semantic-aware multimodal grouping for weakly-supervised audio-visual video parsing. In ECCV Workshop, 2022. 1, 2  
+[20] Kranti Parida, Neeraj Matiyali, Tanaya Guha, and Gaurav Sharma. Coordinated joint multimodal embeddings for generalized audio-visual zero-shot classification and retrieval of videos. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3251–3260, 2020. 2  
+[21] Kranthi Kumar Rachavarapu, Kalyan Ramakrishnan, et al. Weakly-supervised audio-visual video parsing with prototype-based pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18952-18962, 2024. 1, 2  
+[22] 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. 2, 4, 5, 6, 8  
+[23] Janani Ramaswamy. What makes the sound?: A dual-modality interacting network for audio-visual event localization. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4372-4376. IEEE, 2020. 2  
+[24] Janani Ramaswamy and Sukhendu Das. See the sound, hear the pixels. In Proceedings of the IEEE/CVF Winter Confer
+
+ence on Applications of Computer Vision, pages 2970-2979, 2020.  
+[25] Varshanth Rao, Md Ibrahim Khalil, Haoda Li, Peng Dai, and Juwei Lu. Dual perspective network for audio-visual event localization. In European Conference on Computer Vision, pages 689–704. Springer, 2022. 2  
+[26] Amos Storkey. When training and test sets are different: characterizing learning transfer. 2008. 3  
+[27] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of the European conference on computer vision (ECCV), pages 247–263, 2018. 1, 2, 6  
+[28] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 436–454. Springer, 2020. 2, 6  
+[29] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 8  
+[30] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1326–1335, 2021. 1, 2  
+[31] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6292–6300, 2019. 2  
+[32] Yiling Wu, Xinfeng Zhang, Yaowei Wang, and Qingming Huang. Span-based audio-visual localization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1252–1260, 2022. 2  
+[33] Yan Xia and Zhou Zhao. Cross-modal background suppression for audio-visual event localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19989-19998, 2022. 2  
+[34] Hanyu Xuan, Zhenyu Zhang, Shuo Chen, Jian Yang, and Yan Yan. Cross-modal attention network for temporal inconsistent audio-visual event localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 279–286, 2020. 2  
+[35] Kun Zhang, Bernhard Scholkopf, Krikamol Muandet, and Zhikun Wang. Domain adaptation under target and conditional shift. In International conference on machine learning, pages 819-827. Pmlr, 2013. 3  
+[36] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audiovisual event line. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8436-8444, 2021. 2  
+[37] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 2, 4, 5, 6, 7, 8

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e8364fd18b1d329dee583c472a959f22efd4b204e8bd54bb8782b01ef2f5ff
+size 583424

CVPR/2025/Adapting to the Unknown_ Training-Free Audio-Visual Event Perception with Dynamic Thresholds/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c473bf3ad06b70b07582af22d4ac06780dba89f9a214a7a9fb130082ef6906e2
+size 452229

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/753ef088-ca7a-48f3-810f-83fa0edbc376_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8c714b91e99c0d807d851012069a3ca05e9d8345e765e14c5f790a2d009c8ff
+size 85779

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/753ef088-ca7a-48f3-810f-83fa0edbc376_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03adad00e405ecfd4949591fa322207bb38269b7deb3c92c9f21fa6fe6fec21c
+size 107847

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/753ef088-ca7a-48f3-810f-83fa0edbc376_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db3007bdae8cf9b8295babc07acff4b9639332c30ed68249436d87ff90aba514
+size 2966745

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/full.md

@@ -0,0 +1,306 @@
+# Adaptive Dropout: Unleashing Dropout across Layers for Generalizable Image Super-Resolution
+
+Hang Xu Jie Huang Wei Yu Jiangtong Tan Zhen Zou Feng Zhao*
+
+MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition  
+University of Science and Technology of China
+
+{xuhang0609, hj0117, patrick914y, jttan, zouzhen}@mail.ustc.edu.cn, fzhao956@ustc.edu.cn
+
+# Abstract
+
+Blind Super-Resolution (blind SR) aims to enhance the model's generalization ability with unknown degradation, yet it still encounters severe overfitting issues. Some previous methods inspired by dropout, which enhances generalization by regularizing features, have shown promising results in blind SR. Nevertheless, these methods focus solely on regularizing features before the final layer and overlook the need for generalization in features at intermediate layers. Without explicit regularization of features at intermediate layers, the blind SR network struggles to obtain well-generalized feature representations. However, the key challenge is that directly applying dropout to intermediate layers leads to a significant performance drop, which we attribute to the inconsistency in training-testing and across layers it introduced. Therefore, we propose Adaptive Dropout, a new regularization method for blind SR models, which mitigates the inconsistency and facilitates application across intermediate layers of networks. Specifically, for training-testing inconsistency, we re-design the form of dropout and integrate the features before and after dropout adaptively. For inconsistency in generalization requirements across different layers, we innovatively design an adaptive training strategy to strengthen feature propagation by layer-wise annealing. Experimental results show that our method outperforms all past regularization methods on both synthetic and real-world benchmark datasets, also highly effective in other image restoration tasks. Code is available at https://github.com/xuhang07/Adpative-Dropout.
+
+# 1. Introduction
+
+Blind SR [7, 13, 18, 37, 43] aims to recover high-quality images from low-quality images without knowing the degradation kernel. However, even with better real-world degradation simulation [52, 57, 68] and more expressive model
+
+![](images/ea92a1c0cc6e0ceab4f08dd15040214086aba36c8b7f25228895d70a2f4ecf25.jpg)  
+Figure 1. Placing dropout before the final convolutional layer can enhance the generalization performance of blind SR models, whereas placing it elsewhere has the opposite effect due to the inconsistency introduced. Our method effectively addresses the inconsistency allowing it to be easily applied to intermediate layers.
+
+architectures [2, 34, 35], blind SR models still encounter a severe over-fitting problem. Therefore, appropriate regularization terms [12, 53, 59] or training strategies [3, 49] are urgently needed to enhance the generalization of blind SR models. In the past, Kong et al. [29] first introduced regularization into blind SR, adding dropout before the last convolutional layer of the model, which effectively improved the model's generalization ability. Subsequently, Wang et al. [56] argued that dropout could harm the recovery of high-frequency image information and proposed Simple-Align, which feeds low-quality images obtained from different degradations of the same image into the network and aligns the features before the last convolutional layer, encouraging the model to be degradation-invariant.
+
+It is worth noting that the above regularization methods all target the features before the last convolutional layer, while the features at intermediate layers are only implicitly impacted [25, 45, 46]. However, we find this indirect impact hard to regularize the features at intermediate layers to be degradation-invariant [19, 24, 38], and instead exacerbates the imbalance between channels, indicating the necessity of explicit regularization for intermediate features. Therefore,
+
+we explore a new solution that regularizes the intermediate features on the shoulder of Dropout [29], which further elevates the generalization ability of blind SR models.
+
+We rethink the dropout technique for this task. In previous discussions [29, 48], it is not suitable for dropout to be directly applied in regression problems like SR. But Kong et al. [29] improved the model's generalization ability by adding dropout before the last convolutional layer, and indicated that adding it elsewhere led to performance drop. This raises the question of what causes this phenomenon. To analyze the reason, we conduct detailed experiments and attribute the performance decline to the inconsistency brought by directly applying dropout to intermediate layers. Specifically, as shown in Figure 2(b), directly applying dropout to intermediate layers introduces training-testing inconsistency, which harms the expressiveness of features considering the mean shift in the distribution fitted during training. Moreover, we observe that different layers of the network impact generalization differently, with shallow layers influencing general feature extraction and deep layers influencing degradation handling. Among them, general feature representation can consistently play a role in content reconstruction, while degradation representation needs regularization to perform well on unseen data. However, directly applying dropout to intermediate layers doesn't take into account the inconsistency in generalization requirements across different layers. These two types of inconsistency essentially require us to strike a balance between the fitting ability and generalization ability of the network.
+
+In this paper, we introduce Adaptive Dropout for blind SR models to enhance generalization, consisting of the adaptive dropout format and the adaptive training strategy as shown in Figure 3. The key design of our method lies in its consideration of the balance between feature expressiveness and degradation generalization. Concretely, we redesign the format of dropout to adaptively integrate the dropouted feature with the original feature, which effectively mitigates training-testing inconsistency. Furthermore, considering the differing generalization requirements across different layers, we introduce the adaptive training strategy which adaptively helps blind SR models achieve a balance between fitting ability and generalization ability while satisfying the generalization requirements across different layers by layer-wise annealing. It is worth noting that, since our method is easy to integrate with Dropout [29], we unify them and treat them as a whole to help blind SR models improve generalization. We show the superiority of our proposed method in Figure 1 and summarize our main contributions as follows:
+
+- We specifically analyze the inconsistency introduced by directly applying dropout to intermediate layers, including training-testing inconsistency and the inconsistency in the generalization requirements across different layers.
+
+- We introduce Adaptive Dropout that adaptively integrates the dropouted feature with the un-dropouted one. Moreover, we introduce an adaptive training strategy that achieves an elegant trade-off between fitting and generalizing across different layers by introducing different adaptive weight formats and layer-wise annealing.  
+- We validate the effectiveness of Adaptive Dropout on multiple synthetic datasets and real-world datasets. It is also highly effective in other image restoration tasks.
+
+# 2. Related Work
+
+# 2.1. Blind Super-Resolution
+
+To deal with complicated degradation kernels [11] in the real world, blind SR [51, 65] is introduced into the field of super-resolution, which aims to recover high-resolution images even though the degradations are unknown. Compared with implicit degradation modeling [28, 47, 70], which requires large datasets during training, explicit degradation modeling [32, 62] performs better, where the degradations are added into images as data augmentation during training. DASR [36] introduces a new degradation scheme containing four types of degradation (bicubic, blur, noise, JPEG). Then in Real-ESRGAN [57], applying the degradation scheme twice has proved helpful in generating low-resolution images closer to the real world, and has become a regular part of data augmentation when training blind SR models. However, just training blind SR models with a large degradation pool brings a severe problem: the SR network learns to overfit the distribution of specific degradations rather than the distribution of natural images [29, 65, 68]. The problem forces researchers to find a proper training strategy to help blind SR models generalize better.
+
+# 2.2. Regularization methods for blind SR
+
+For models of another field [20, 22, 23], regularization [12, 53, 59], normalization [5, 60, 61], and other modules [8, 14, 21] are widely used as training strategies to improve the network's generalization performance. However, for blind SR models, few researchers pay attention to the training strategies. Previously, researchers focused more on improving the network's generalization ability to unknown degradations through complex degradation paradigms [52, 57, 68] and network designs [2, 34, 35], while overlooking the fact that training strategies are also an important factor affecting the network's generalization performance. Recently, Kong et al. [29] demonstrated that dropout can also be applied to blind SR models and should be added before the last convolutional layer while adding dropout to other locations (such as within the residual blocks) leads to a decline in model performance. Simple-Align [56] shows the side-effect of dropout and provides a better choice as regularization, where two low-resolution images obtained from the same high-resolution image but with different degradations
+
+![](images/0f3ffc21edd85d21237c709889ced704a9228e29d0d1829844c5e9b4cc7da9f9.jpg)  
+(a) imbalance between feature channels at intermediate layers
+
+![](images/4bacc1ec879de55736b680b1379ca3917af07206c59fffe68ef2b93a6fa685c7.jpg)  
+(b) training-testing inconsistency accumulation  
+(c) differing responses of features at different layers to dropout  
+Figure 2. The source of our motivation. (a) shows the imbalance between shallow feature channels, which is not well constrained by past methods. Mask channel index refers to the position of the channel we mask when adopting channel ablation. (b) compares the results of adding dropout to more blocks, indicating the accumulation of training-testing inconsistency along layers. (c) demonstrates the different behaviors of shallow and deep features when faced with dropout, as summarized in the text. We attribute this to the differing generalization requirements of different layers. Specifically, deep features more prominently represent information related to training degradations. Regularizing degradation representations alleviates the network's overfitting to training degradations, but directly applying dropout in deep layers also impairs the expressiveness of feature representations, leading to the performance shown in the chart. Shallow features, which contain more general representations, exhibit more consistent performance across wider degradations after being regularized.
+
+![](images/69fc28fc1af3196a7ff28632d48785b9101b3176ee908fa2c17190066c279506.jpg)
+
+are fed into the network, and the statistics of the outputs before the last convolutional layer are aligned across different dimensions. However, we argue that dropout is still a great choice and can be added into all blocks in a different way.
+
+# 3. Motivation
+
+# 3.1. Intermediate layers need explicit regularization
+
+Both Dropout [29] and Simple-Align [56] regularize features before the final convolutional layer which directly impact the quality of restoration. So applying constraints to these features can effectively enhance the model's generalization ability. Meanwhile, regularizations targeting the final layer also serve as implicit constraints on intermediate layers. However, we argue this indirect constraint is insufficient to generalize the features at intermediate layers.
+
+We evaluate the balance between channels in the intermediate layers under different regularization with channel ablation [44] and show the results in Figure 2(a). Specifically, we sequentially occlude each channel to obtain a new feature map, which is then normalized to ensure the conservation of total energy, input into subsequent layers, and calculate the PSNR. Surprisingly, adding regularization only to the final layer doesn't improve the generalization of the features at intermediate layers, while stronger regularization exacerbates the imbalance between channels. In contrast, we explicitly regularize the features at intermediate layers resulting in a smaller performance decrease when conducting channel ablation [44]. This indicates the necessity of explicit regularization of the features at intermediate layers.
+
+As a widely used regularization method, we want to incorporate dropout into intermediate layers of Blind SR networks to explicitly regularize features there. Although Kong et al. [29] argue that dropout can only be applied before the last convolutional layer and not to intermediate lay
+
+ers, they lack an in-depth analysis of the reasons behind this. Therefore, we delve deeper by analyzing the issues brought about by dropout, seeking a reasonable approach to explicitly regularize features at intermediate layers.
+
+# 3.2. Inconsistency when directly applying dropout
+
+In Dropout [29], the impact of applying dropout to the network's intermediate layers was analyzed, leading to the following conclusions: (1) Applying dropout to the intermediate layers of the network rather than the final layer results in a performance decline. (2) Applying more dropout comes to more performance declines. (3) Applying dropout closer to the output layer results in better model performance.
+
+Overall, we attribute these three conclusions to the inconsistency brought by applying dropout to the intermediate layers, including that between the training and testing phases, and that among different layers of the network. Considering that the performance of Blind SR networks in unseen degradations and unseen scenarios is a joint effect of their fitting and generalization ability, improving generalization requires ensuring fitting ability. However, the inconsistency severely undermines the network's fitting ability. Below, we separately explain the causes of the inconsistency and provide corresponding solutions.
+
+Point 1: The training-testing inconsistency. For conclusion (1) and (2) brought by Dropout [29], We attribute them to the training-testing inconsistency [27, 33] caused by applying dropout directly at intermediate layers. Moreover, when applying more dropout to intermediate layers, this inconsistency accumulates along the direction of information propagation, leading to the representations obtained from training not being well utilized during inference. In other words, the feature representations learned are losing their expressiveness. We illustrate the accumulation of inconsistency in 2, where the inconsistency is manifested as the shift
+
+![](images/05fc19e39c76333b434ae7235b4917223d07c97cae855f92f02c2d3a4465092e.jpg)  
+Figure 3. Overall schema of our proposed adaptive dropout. Adaptive dropout consists of two components. (a) The adaptive dropout format adaptively couples the features before and after dropout, thereby achieving a balance between the network's fitting ability and generalization ability while alleviating the training-testing inconsistency. (b) Considering the different generalization requirements of features at different layers, we design an adaptive training strategy and its two variants. Regarding Implicit Adaptive Dropout, it adaptively adjusts the weight  $w$  through a module, with different designs for shallow and deep layers. Regarding Explicit Adaptive Dropout, it initializes  $w$  in all blocks and explicitly anneals it block-by-block during training, as shown in the algorithm beside.
+
+in features at intermediate layers after applying dropout.
+
+First, we introduce the training-testing inconsistency brought by dropout. We take channel dropout as an example. Consider a feature map  $x_{k} \in \mathbb{R}^{c \times h \times w}$ . When applying channel dropout to  $x_{k}$ , it is multiplied by a dropout mask  $a_{k} \in \mathbb{R}^{c}$  which comes from a Bernoulli distribution during training:  $\hat{x}_{k} = a_{k}x_{k}$ . During training and testing, the mean of  $\hat{x}_{k}$  remains the same, but there is a variance shift. Then after passing through the activation function, the variance shift leads to the mean shift which harms the expressiveness of the feature presentations. This issue does not exist when dropout is added before the last convolutional layer, with no activation function following it. Assume we have  $x_{k}$  with a mean  $\mu$  and variance  $\sigma^2$ :
+
+$$
+\mathbf {v a r} [ \hat {x _ {k}} ] = \left\{ \begin{array}{l l} \frac {1}{1 - p} (\mu^ {2} + \sigma^ {2}) - \mu^ {2}, & \text {d u r i n g t r a n i n g} \\ \sigma^ {2}, & \text {d u r i n g i n f e r e n c e} \end{array} \right.
+$$
+
+Then, the relative deviation of the variance of  $\hat{x_k}$  during
+
+training can be expressed as  $s_k$ :
+
+$$
+\begin{array}{l} s _ {k} = \frac {\operatorname {v a r} _ {\text {t r a i n i n g}} \left[ \hat {x} _ {k} \right]}{\operatorname {v a r} _ {\text {i n f e r e n c e}} \left[ \hat {x} _ {k} \right]} - 1 \\ = \frac {1}{1 - p} \left(\frac {\mu^ {2}}{\sigma^ {2}} + 1\right) - \frac {\mu^ {2}}{\sigma^ {2}} - 1 \\ = \frac {p}{1 - p} (\frac {\mu^ {2}}{\sigma^ {2}} + 1) \\ \end{array}
+$$
+
+So reducing  $s_k$  can decrease the training-testing inconsistency, while also reducing the inconsistency accumulation.
+
+Solution 1: Adaptive dropout format. Integrating the dropouted feature with the original feature is a natural thought, considering we need to enhance the generalization of features while ensuring their expressiveness. Therefore, we introduce an adaptive dropout format to Blind SR, as shown in Figure 3(a). the basic format can be expressed as:
+
+$$
+f (x) = w x + (1 - w) d r o p o u t (x, p) \tag {1}
+$$
+
+Where  $x$  represents the feature map, dropout refers to dropout operations,  $w$  represents the weighted coefficients, and  $p$  represents the dropout rate. We can calculate the relative deviation of the variance introduced by the adaptive
+
+![](images/6ead3c4c68cb992345bce3d30fda67540f089cd7bd0d5e393e3b067790c63412.jpg)  
+Figure 4. Our exploration for mitigating the inconsistency in generalization requirements across different layers. The left and right figures respectively show the model performance using different regularizations on Set5 and Set14. With consideration of how to better retain the fitting ability of SR models, a learnable weighted structure yields better results than the standard one.
+
+dropout format in the same way:
+
+$$
+\begin{array}{l} \operatorname {v a r} \left[ x _ {k} ^ {a d} \right] = \frac {(1 - w) ^ {2}}{1 - p} \left(\mu^ {2} + \sigma^ {2}\right) - \mu^ {2} \\ s _ {k} ^ {a d} = \frac {(1 - w) ^ {2} p}{1 - p} (\frac {\mu^ {2}}{\sigma^ {2}} + 1) \\ \end{array}
+$$
+
+Compared with standard dropout which can be viewed as a special case of adaptive dropout format when  $w = 0$ , adaptive dropout format reduces the variance shift by  $(1 - w)^2$  leading to less training-testing inconsistency and inconsistency accumulation when applying it into multiple blocks.
+
+Point 2: The inconsistency in generalization requirements across different layers. For conclusion (3) brought by Dropout [29], we attribute this to the differing generalization requirements of shallow and deep layers. Prior works [40, 64] have found that representations extracted from the shallow layers are more generalized, while the representations extracted from the deep layers show strong task relevance. For blind SR, the extraction of general features is crucial for image reconstruction, regardless of the degradation the low-quality image has undergone. However, task-specific representations may not perform well on test images, considering that these representations may overfit the degradations of the train set without regularization. Therefore, providing deeper features with greater perturbation without compromising fitting ability helps the model avoid overfitting to specific degradations in the training pool. To support this point, we verify the performance on Set5 with different degradations when dropout is directly applied at different layers, presenting the main results in Figure 2(c). It can be observed that after being regularized, deep features result in differing performance across various degradations, whereas shallow features exhibit the opposite behavior, indicating their differing generalization requirements.
+
+Solution 2: Adaptive training strategy. In the basic format of adaptive dropout,  $x$  represents the complete transmission of information,  $dropout(x)$  represents the regularizations on features and  $w$  represents the trade-off between fitting ability and generalization ability. Considering the differing generalization requirements of different layers, it is reasonable to set different  $w$  for each layer or block.
+
+![](images/7b36687589461f62ab5cdf1162f752af44215816cae81dfad4320b0438d6ac40.jpg)  
+Figure 5. Our exploration for the adaptive training strategy. The graph on the left shows the trend of changes in  $w$  across different layers when we use the learnable weight structure. The graph on the right illustrates the performance differences between vector  $w$  and value  $w$ , where vector  $w$  performs better on more complex datasets while value  $w$  performs better on others.
+
+![](images/df7396a0148d212d11b8fca11f973cbdbb0e2602d1864a03615eea57da777c01.jpg)
+
+However, there is no clear guidance on how to apply different perturbations to different layers. So we design an experiment where we couple  $x$  and dropout(x) with learnable weights  $w$  and  $1 - w$  for each block, to help uncover the patterns guiding the application of adaptive dropout format. Also, we simply use an adaptive dropout format with  $w = 0.5$  as a comparison.
+
+Understandably, this learnable weighted structure yields more significant effects than the simple application of the adaptive dropout format as illustrated in Figure 4, for the former one ensures the network's fitting ability since the network adjusts  $w$  towards fitting the train set. Then, we present the changes in  $w$  across different blocks in Figure 5(a) to identify the patterns of the network's adaptive adjustment of  $w$ : First,  $w$  in each block gradually approaches 1, indicating that the model itself slowly reduces the perturbation of features, which represents the requirements of SR network for fitting ability. Second, compared to deeper blocks, the  $w$  values in shallower blocks approach 1 more rapidly, representing the requirements for generalization ability across different layers. We can summarize that the network is performing an adaptive annealing process from shallow to deep layers. This layer-wise annealing training strategy, which we refer to as the adaptive training strategy shown in Figure 3(b), aligns with our requirement to address the inconsistency in generalization needs across different layers. We divide the training phase into generalization and fitting stages, referring to the periods when intermediate layer features are perturbed and not perturbed, respectively. The shallow layers of the network enter the fitting stage earlier, ensuring that the model's general feature representations have good expressiveness. The deep layers remain in the generalization stage for a longer time, ensuring that the model does not overfit the degradations of the train set, thereby performing well on unseen degradations. Additionally, the adaptive training strategy effectively addresses the inconsistency accumulation issue, enabling the network to eliminate perturbations from the source.
+
+# 4. Methods
+
+To further improve the generalization of blind SR networks, we propose adaptive dropout, where the adaptive dropout format and adaptive training strategy are the core designs of our method. Below, we separately introduce them in detail.
+
+# 4.1. Adaptive Dropout format
+
+In Equation 1, we presented the basic form of the adaptive dropout format, so here we specifically discuss the hyperparameters, Integration, and application in networks.
+
+Hyper-parameters. The adaptive dropout format has two hyper-parameters,  $w$  and  $p$ . However, we find that setting  $p$  to 0.5 and adjusting  $w$  alone is sufficient to achieve better generalization performance in blind SR networks. Therefore, we simply set  $p$  to 0.5 to simplify our method design.
+
+Integration. Since the proposed adaptive dropout targets the intermediate features of the network, it can effectively be integrated with Dropout [29] which targets features before the last convolutional layer. We still retain the original format of dropout before the last convolutional layer for it doesn't bring the training-testing inconsistency.
+
+Application. In CNN-based SR networks [30, 69], using multiple adaptive dropout within the same block does not improve performance on unseen degradation. We speculate that this might excessively constrain the network's fitting capability within the same training time. For SwinIR [34] where dropout is generally not used under normal settings, we reintroduce this regularization method and replace all dropout with adaptive dropout, which is also applicable to other transformer-based SR networks [9, 15].
+
+# 4.2. Adaptive Training Strategy
+
+Previously, we discovered that the network implicitly employs a layer-wise annealing strategy through our attempt to adaptively adjust the learnable  $w$ . Given this, we argue that explicitly applying this strategy can also yield good results and design two variants for adaptive dropout.
+
+Implicit Adaptive Dropout with different formats of  $w$ . We implicitly adopt the adaptive training strategy by treating  $w$  as an estimizable variable and letting the neural network learn the optimal  $w$  just as our attempt to adaptively adjust the learnable  $w$ . To better weigh the importance of each channel, we use a network similar to channel attention [16], where the weight  $w_{i}$  can be obtained by:
+
+$$
+\begin{array}{l} w _ {i} = \operatorname {s i g m o i d} \left(f _ {2} \left(\operatorname {R e L U} \left(f _ {1} \left(x _ {k}\right)\right)\right)\right) \\ \hat {x} _ {k} = w x _ {k} + (1 - w) \text {d r o p o u t} (x _ {k}) \\ \end{array}
+$$
+
+Where  $f_{1}$  and  $f_{2}$  represent linear projection,  $x_{k}$  means the original feature, and  $\hat{x_k}$  means the coupled feature. According to specific requirements,  $w_{i}$  can take the format of  $[B,C]$ , where  $B$  means the batch size and  $C$  means the number of channels, meaning that different weights are learned for each channel of the feature, referred to as
+
+vector w. It can also take the format of  $[B,1]$ , meaning that the same weight is shared across all channels, referred to as value w. The former introduces more randomness and variability compared to the latter. The impact of this randomness is shown in Figure 5(b): in more complex datasets, Manga109 and Urban100, vector w achieves better results than value w, whereas in simpler datasets, Set14 and BSD100, the opposite is true. Therefore, we use them in different blocks to strike a balance between the two. Considering that we should reduce disturbances in the shallow layers, we apply value w in the shallow layers and vector w in the deeper layers.
+
+Explicit Adaptive Dropout with layer-wise annealing. We explicitly adopt the adaptive training strategy by initializing  $w$  for all blocks and then annealing it block-by-block during training. Specifically, for every  $t$  training iteration, we set the  $w$  to 0 in one block, repeating the operation from shallow to deep layers, where  $t$  is related to the total number of training iterations. In this way, general feature representations can be sufficiently preserved, while degradation-related representations can be sufficiently constrained.
+
+# 5. Experiments
+
+Experimental settings. We follow the multi-degradation settings and experimental configurations outlined in Dropout [29]. For training, we utilize the HR images from the DIV2K dataset [1]. During training, we employ L1 loss and the Adam optimizer [10]. The batch size is set to 16, and the size of the LR images is  $32 \times 32$ . We adopt a cosine annealing strategy to dynamically adjust the learning rate, with an initial learning rate set to  $2 \times 10^{-4}$ . We train for a total of 500,000 iterations. For testing, we use the synthetic datasets Set5 [4], Set14 [63], BSD100 [41], Manga109 [42], and Urban100 [17], as well as the real-world datasets NTIRE 2018 SR challenge data [54], RealSR [6], and DrealSR [58]. We compare our approach with Dropout [29] and Simple-Align [56].
+
+Comparison with baseline models. We present the experimental results for the baseline with different regularization in Table 1. It can be observed that our proposed adaptive dropout significantly outperforms the baseline and comparison methods across all degradation settings in all datasets. Based on the experimental results, we can draw the following conclusions: (1) Both Implicit Adaptive Dropout and Explicit Adaptive Dropout can bring more performance increase compared with Dropout [29], for both of these variants explicitly regularize features at intermediate layers, encouraging models to learn more general features representations. (2) Explicit Adaptive Dropout generally outperforms Implicit Adaptive Dropout, considering that Implicit Adaptive Dropout focuses more on maintaining the model's fitting ability. We show the comparison results on RealSR-ResNet [30] and SwinIR [34] in Figure 6.
+
+![](images/ae0b3a2e555511bd5671585a1845b89f3fa314d50215948aa4073bd056007533.jpg)  
+Figure 6. Comparison with other regularization methods [29, 56].
+
+![](images/3259fa57874df1ba588872d0aa4db3522bfbd041a853bda13c728d9ec5ecfbdc.jpg)
+
+![](images/89459d860335babf3de17727cf2115dbf7265e377ae2f78eb0dfdffaa98ff281.jpg)
+
+![](images/93a23cba36da2499d7bd251bbaac3217906ad5d311398f0a0c11e11297fb5760.jpg)
+
+Table 1. The PSNR (dB) results of models with  $\times 4$ . We test them on eight types of degradations or multi-degradations (b is blur, n is noise, j is jpeg). Red and Blue indicate the best and the second-best performance, respectively.  
+
+<table><tr><td rowspan="2">Models</td><td rowspan="2">Regularization</td><td colspan="2">Set5 [4]</td><td colspan="2">Set14 [63]</td><td colspan="2">BSD100 [41]</td><td colspan="2">Manga109 [42]</td><td colspan="2">Urban100 [17]</td></tr><tr><td>clean</td><td>blur</td><td>clean</td><td>blur</td><td>clean</td><td>blur</td><td>clean</td><td>blur</td><td>clean</td><td>blur</td></tr><tr><td rowspan="4">SRResNet [30]</td><td>None</td><td>24.89</td><td>24.76</td><td>22.60</td><td>22.50</td><td>23.06</td><td>22.99</td><td>18.42</td><td>18.75</td><td>21.24</td><td>21.06</td></tr><tr><td>Dropout [29]</td><td>25.67</td><td>25.34</td><td>23.11</td><td>22.92</td><td>23.37</td><td>23.24</td><td>18.87</td><td>19.14</td><td>21.58</td><td>21.31</td></tr><tr><td>Explicit Adaptive Dropout</td><td>26.11</td><td>25.39</td><td>23.42</td><td>23.20</td><td>23.76</td><td>23.59</td><td>19.40</td><td>19.62</td><td>21.87</td><td>21.48</td></tr><tr><td>Implicit Adaptive Dropout</td><td>26.09</td><td>25.56</td><td>23.47</td><td>23.21</td><td>23.73</td><td>23.50</td><td>19.51</td><td>19.69</td><td>21.90</td><td>21.40</td></tr><tr><td rowspan="4">SwinIR [34]</td><td>None</td><td>26.25</td><td>26.03</td><td>23.76</td><td>23.47</td><td>23.91</td><td>23.83</td><td>19.10</td><td>19.19</td><td>22.18</td><td>21.90</td></tr><tr><td>Dropout [29]</td><td>26.32</td><td>26.08</td><td>23.80</td><td>23.36</td><td>23.90</td><td>23.87</td><td>19.15</td><td>19.30</td><td>22.27</td><td>21.99</td></tr><tr><td>Explicit Adaptive Dropout</td><td>26.54</td><td>26.21</td><td>23.92</td><td>23.56</td><td>24.09</td><td>23.59</td><td>19.22</td><td>19.39</td><td>22.41</td><td>22.19</td></tr><tr><td>Implicit Adaptive Dropout</td><td>26.46</td><td>26.12</td><td>23.95</td><td>23.41</td><td>24.01</td><td>23.50</td><td>19.31</td><td>19.44</td><td>22.32</td><td>22.16</td></tr><tr><td></td><td></td><td>noise</td><td>jpeg</td><td>noise</td><td>jpeg</td><td>noise</td><td>jpeg</td><td>noise</td><td>jpeg</td><td>noise</td><td>jpeg</td></tr><tr><td rowspan="4">SRResNet [30]</td><td>None</td><td>22.02</td><td>23.72</td><td>20.81</td><td>21.84</td><td>20.34</td><td>22.48</td><td>19.74</td><td>18.30</td><td>19.73</td><td>20.60</td></tr><tr><td>Dropout [29]</td><td>21.87</td><td>24.04</td><td>20.61</td><td>22.16</td><td>21.00</td><td>22.70</td><td>18.25</td><td>18.62</td><td>19.58</td><td>20.89</td></tr><tr><td>Explicit Adaptive Dropout</td><td>22.41</td><td>24.38</td><td>21.14</td><td>22.46</td><td>21.48</td><td>23.00</td><td>18.67</td><td>19.07</td><td>19.90</td><td>21.11</td></tr><tr><td>Implicit Adaptive Dropout</td><td>22.41</td><td>24.36</td><td>21.15</td><td>22.51</td><td>21.53</td><td>23.00</td><td>18.76</td><td>19.20</td><td>19.90</td><td>21.15</td></tr><tr><td rowspan="4">SwinIR [34]</td><td>None</td><td>22.96</td><td>24.37</td><td>21.56</td><td>23.04</td><td>22.12</td><td>23.04</td><td>18.71</td><td>18.95</td><td>20.56</td><td>21.32</td></tr><tr><td>Dropout [29]</td><td>23.12</td><td>24.41</td><td>21.59</td><td>23.09</td><td>22.10</td><td>23.08</td><td>18.73</td><td>19.03</td><td>20.67</td><td>21.38</td></tr><tr><td>Explicit Adaptive Dropout</td><td>23.63</td><td>24.68</td><td>21.67</td><td>23.29</td><td>22.24</td><td>23.29</td><td>18.92</td><td>19.27</td><td>20.85</td><td>21.52</td></tr><tr><td>Implicit Adaptive Dropout</td><td>23.49</td><td>24.49</td><td>21.69</td><td>23.19</td><td>22.19</td><td>23.28</td><td>19.02</td><td>19.31</td><td>20.74</td><td>21.40</td></tr><tr><td></td><td></td><td>b+n</td><td>b+j</td><td>b+n</td><td>b+j</td><td>b+n</td><td>b+j</td><td>b+n</td><td>b+j</td><td>b+n</td><td>b+j</td></tr><tr><td rowspan="4">SRResNet [30]</td><td>None</td><td>23.31</td><td>23.44</td><td>21.81</td><td>21.70</td><td>22.27</td><td>22.34</td><td>18.60</td><td>18.53</td><td>20.46</td><td>20.30</td></tr><tr><td>Dropout [29]</td><td>23.51</td><td>23.67</td><td>21.87</td><td>22.01</td><td>22.24</td><td>22.54</td><td>18.94</td><td>18.81</td><td>20.48</td><td>20.53</td></tr><tr><td>Explicit Adaptive Dropout</td><td>23.74</td><td>23.80</td><td>22.15</td><td>22.29</td><td>22.48</td><td>22.81</td><td>19.26</td><td>19.23</td><td>20.57</td><td>20.66</td></tr><tr><td>Implicit Adaptive Dropout</td><td>23.65</td><td>23.73</td><td>22.12</td><td>22.29</td><td>22.44</td><td>22.78</td><td>19.51</td><td>19.29</td><td>20.50</td><td>20.63</td></tr><tr><td rowspan="4">SwinIR [34]</td><td>None</td><td>23.80</td><td>23.84</td><td>22.20</td><td>22.26</td><td>22.61</td><td>22.82</td><td>19.07</td><td>19.02</td><td>20.89</td><td>20.79</td></tr><tr><td>Dropout [29]</td><td>24.00</td><td>23.93</td><td>22.40</td><td>22.24</td><td>22.68</td><td>22.80</td><td>19.12</td><td>18.98</td><td>20.92</td><td>20.91</td></tr><tr><td>Explicit Adaptive Dropout</td><td>24.24</td><td>24.27</td><td>22.91</td><td>22.41</td><td>22.81</td><td>23.05</td><td>19.25</td><td>19.29</td><td>21.00</td><td>20.96</td></tr><tr><td>Implicit Adaptive Dropout</td><td>24.18</td><td>24.30</td><td>22.78</td><td>22.36</td><td>22.80</td><td>23.05</td><td>19.31</td><td>19.37</td><td>20.92</td><td>20.94</td></tr><tr><td></td><td></td><td>n+j</td><td>b+n+j</td><td>n+j</td><td>b+n+j</td><td>n+j</td><td>b+n+j</td><td>n+j</td><td>b+n+j</td><td>n+j</td><td>b+n+j</td></tr><tr><td rowspan="4">SRResNet [30]</td><td>None</td><td>23.21</td><td>22.70</td><td>21.59</td><td>21.44</td><td>22.24</td><td>22.05</td><td>18.25</td><td>18.43</td><td>20.42</td><td>20.10</td></tr><tr><td>Dropout [29]</td><td>23.55</td><td>22.99</td><td>21.86</td><td>21.64</td><td>22.39</td><td>22.17</td><td>18.57</td><td>18.71</td><td>20.64</td><td>20.23</td></tr><tr><td>Explicit Adaptive Dropout</td><td>23.71</td><td>23.10</td><td>22.09</td><td>21.84</td><td>22.59</td><td>22.31</td><td>18.95</td><td>19.04</td><td>20.79</td><td>20.29</td></tr><tr><td>Implicit Adaptive Dropout</td><td>23.74</td><td>23.10</td><td>22.13</td><td>21.84</td><td>22.60</td><td>22.32</td><td>19.07</td><td>19.10</td><td>20.81</td><td>20.27</td></tr><tr><td rowspan="4">SwinIR [34]</td><td>None</td><td>23.67</td><td>22.99</td><td>22.11</td><td>21.82</td><td>22.61</td><td>22.34</td><td>18.79</td><td>18.80</td><td>20.98</td><td>20.45</td></tr><tr><td>Dropout [29]</td><td>23.65</td><td>23.09</td><td>22.17</td><td>21.84</td><td>22.64</td><td>22.33</td><td>18.75</td><td>18.84</td><td>21.12</td><td>20.55</td></tr><tr><td>Explicit Adaptive Dropout</td><td>23.82</td><td>23.18</td><td>22.30</td><td>22.19</td><td>22.78</td><td>22.50</td><td>18.82</td><td>18.91</td><td>21.15</td><td>20.57</td></tr><tr><td>Implicit Adaptive Dropout</td><td>23.76</td><td>23.13</td><td>22.23</td><td>22.19</td><td>22.57</td><td>22.39</td><td>18.87</td><td>19.10</td><td>21.02</td><td>20.56</td></tr></table>
+
+Comparison with Simple-Align. The comparison results between our method and Simple-Align [56] are also shown in Figure 6. Our proposed method outperforms SimpleAlign [56] on all synthetic datasets, with notable improvements on the more complex datasets Manga109 and Urban100, showing better generalization ability on complex scenarios. The detailed data is in supplementary material.
+
+To further demonstrate the effectiveness of our method, we compare it with previous methods on real-world datasets, for low-resolution images collected from the real
+
+world are far more challenging to reconstruct than synthetic low-resolution images. RealSR [6] and DRealSR [58] are the latest two paired datasets for real-world superresolution, and we show the comparison results of these two datasets in Figure 7 and Table 2. We also conduct tests on the realistic NTIRE 2018 SR challenge data [54], and the results are presented in Table 2. Our method outperforms Simple-Align [56] on all five real-world datasets.
+
+Applications in single-degradation task. Adaptive Dropout also works well for generalization tasks of single
+
+Table 2. Comparison with Simple-Align [56] on synthetic datasets and real-world datasets.  
+
+<table><tr><td rowspan="2">Models+Regularization</td><td colspan="5">Synthetic dataset</td><td colspan="5">Real-world dataset</td></tr><tr><td>Set5</td><td>Set14</td><td>BSD100</td><td>Manga109</td><td>Urban100</td><td>RealSR</td><td>DRealSR</td><td>mild</td><td>difficult</td><td>wild</td></tr><tr><td>SRResNet+Simple-Align [56]</td><td>24.05</td><td>22.31</td><td>22.70</td><td>19.11</td><td>20.80</td><td>24.63</td><td>27.14</td><td>17.19</td><td>18.05</td><td>17.81</td></tr><tr><td>SRResNet+Explicit Adaptive Dropout</td><td>24.10</td><td>22.34</td><td>22.73</td><td>19.15</td><td>20.81</td><td>25.07</td><td>27.51</td><td>17.29</td><td>18.18</td><td>17.91</td></tr><tr><td>SRResNet+Implicit Adaptive Dropout</td><td>24.05</td><td>22.32</td><td>22.75</td><td>19.24</td><td>20.84</td><td>25.16</td><td>27.51</td><td>17.32</td><td>18.18</td><td>17.92</td></tr><tr><td>RRDB+Simple-Align [56]</td><td>24.33</td><td>22.51</td><td>22.90</td><td>19.14</td><td>21.05</td><td>24.76</td><td>27.00</td><td>17.14</td><td>18.08</td><td>17.81</td></tr><tr><td>RRDB+Explicit Adaptive Dropout</td><td>24.40</td><td>22.56</td><td>22.97</td><td>19.27</td><td>21.00</td><td>25.39</td><td>27.46</td><td>17.62</td><td>18.33</td><td>18.04</td></tr><tr><td>RRDB+Implicit Adaptive Dropout</td><td>24.38</td><td>22.61</td><td>22.91</td><td>19.30</td><td>20.96</td><td>25.35</td><td>27.53</td><td>17.27</td><td>18.20</td><td>17.90</td></tr></table>
+
+![](images/98699aecd315a7fbf258c628a5241393d8c1dcb04f862d5003e76e75a9de6792.jpg)  
+Figure 7. Comparison with Simple-Align [56] on real-world datasets.
+
+degradation scenarios. We adopt our proposed methods on image denoising, image draining, and image dehazing. For image deraining, we train RCDNet [55] and Restormer [67] on Rain100L [26] and test them on Rain100H [26]. For image denoising, we follow Chen et al. [8], and train SwinIR on Gaussian noise. For image dehazing, we train FFANet [50] on ITS Dataset [31] and test it on OTS Dataset [31]. Results in Figure 3 illustrate that our methods help models generalize better on unseen degradation.
+
+Applications in generative models. We show the results on generative IR in Tab. 4, where our methods work well on both GAN-based and diffusion-based blind SR models.
+
+Table 3. The performance of the model in single degradation tasks when augmented with explicit adaptive dropout.  
+
+<table><tr><td>task</td><td colspan="2">models</td><td>PSNR</td><td>SSIM</td></tr><tr><td rowspan="4">derain</td><td rowspan="2">RCDNet</td><td>None</td><td>17.36</td><td>0.554</td></tr><tr><td>w/ ours</td><td>17.59</td><td>0.565</td></tr><tr><td rowspan="2">Restormer</td><td>None</td><td>29.46</td><td>0.887</td></tr><tr><td>w/ ours</td><td>29.79</td><td>0.898</td></tr><tr><td rowspan="2">dehaze</td><td rowspan="2">FFANet</td><td>None</td><td>20.26</td><td>0.881</td></tr><tr><td>w/ ours</td><td>21.09</td><td>0.883</td></tr><tr><td rowspan="2">denoise</td><td rowspan="2">SwinIR</td><td>None</td><td>26.14</td><td>0.665</td></tr><tr><td>w/ ours</td><td>26.40</td><td>0.681</td></tr></table>
+
+Ablation study. We conduct ablation study on different components of Adaptive Dropout as shown in Table 5. It can be observed that both the adaptive dropout format and the adaptive training strategy contribute to improving the
+
+Table 4. Generative IR with our methods  
+
+<table><tr><td>model</td><td>PSNR↑</td><td>SSIM↑</td><td>LPIPS↓</td><td>MANIQA↑</td></tr><tr><td>ResShift [66]</td><td>22.55</td><td>0.6732</td><td>0.40</td><td>0.2906</td></tr><tr><td>ResShift+ours</td><td>22.69</td><td>0.6754</td><td>0.38</td><td>0.3211</td></tr><tr><td>RealEsgan [57]</td><td>23.19</td><td>0.6812</td><td>0.31</td><td>0.2549</td></tr><tr><td>RealEsgan+ours</td><td>23.64</td><td>0.6881</td><td>0.24</td><td>0.2560</td></tr></table>
+
+Table 5. Ablation Study. The left table demonstrates the effectiveness of our method's components. The right table indicates the optimal hyperparameters for Implicit Adaptive Dropout.  
+
+<table><tr><td>dropout</td><td>strategy</td><td>PSNR</td><td>w</td><td>SRResNet</td><td>RRDB</td></tr><tr><td>None</td><td>X</td><td>25.66</td><td>0.5</td><td>25.83</td><td>26.09</td></tr><tr><td>standard</td><td>X</td><td>24.33</td><td>0.6</td><td>25.94</td><td>26.22</td></tr><tr><td>standard</td><td>✓</td><td>24.89</td><td>0.7</td><td>26.07</td><td>26.43</td></tr><tr><td>adaptive</td><td>X</td><td>25.89</td><td>0.8</td><td>26.00</td><td>26.59</td></tr><tr><td>adaptive</td><td>✓</td><td>26.07</td><td>0.9</td><td>25.86</td><td>26.50</td></tr></table>
+
+![](images/b5b5f9c91804bf35df8886452fe146cbf0c5b01e5459131eb146bfce2601594b.jpg)  
+Figure 8. Quantitative generalization performance compared with other regularization methods.
+
+![](images/71a8f1626a4512a727fdece4541d01d4ad423c78025b6953841762ddf1eacb60.jpg)
+
+model's generalization capability. Moreover, we select the optimal  $w$  for Implicit Adaptive Dropout. More results of ablation study will be presented in supplementary materials. Quantitative analysis of generalization performance. To further compare generalization performance, we use SRGA [39], the first generalization performance metric in the low-level domain, to demonstrate the effectiveness of our proposed method. We show the SRGA under different levels of degradation in Figure 8, where a lower SRGA indicates better generalization performance. Our proposed method achieves lower SRGA on most degradations.
+
+# 6. Conclusion
+
+In this work, we first focus on the need for explicit regularization of intermediate layers in BSR networks to enhance generalization performance, and specifically analyze why dropout cannot be directly applied to intermediate layers. Subsequently, we propose Adaptive Dropout, a regularization method that can be applied anywhere in the network, further improving the network's generalization performance. We hope that this research will draw more attention to the training strategies for SR tasks.
+
+# 7. Acknowledgement
+
+This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC.
+
+# References
+
+[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6  
+[2] Sefi Bell-Kligler, Assaf Shocher, and Michal Irani. Blind super-resolution kernel estimation using an internal-gan. Advances in Neural Information Processing Systems, 32, 2019. 1, 2  
+[3] Irwan Bello, William Fedus, Xianzhi Du, Ekin Dogus Cubuk, Aravind Srinivas, Tsung-Yi Lin, Jonathon Shlens, and Barret Zoph. Revisiting resnets: Improved training and scaling strategies. Advances in Neural Information Processing Systems, 34:22614-22627, 2021. 1  
+[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. British Machine Vision Conference, 2012. 6, 7  
+[5] Nils Bjorck, Carla P Gomes, Bart Selman, and Kilian Q Weinberger. Understanding batch normalization. Advances in Neural Information Processing Systems, 31, 2018. 2  
+[6] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2019. 6, 7  
+[7] Chaofeng Chen, Xinyu Shi, Yipeng Qin, Xiaoming Li, Xiaoguang Han, Tao Yang, and Shihui Guo. Real-world blind super-resolution via feature matching with implicit high-resolution priors. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1329-1338, 2022. 1  
+[8] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 2, 8  
+[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 6  
+[10] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6  
+[11] Runmin Dong, Lichao Mou, Lixian Zhang, Haohuan Fu, and Xiao Xiang Zhu. Real-world remote sensing image superresolution via a practical degradation model and a kernel-aware network. ISPRS Journal of Photogrammetry and Remote Sensing, 191:155-170, 2022. 2
+
+[12] Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. Dropout: A regularization method for convolutional networks. Advances in Neural Information Processing Systems, 31, 2018. 1, 2  
+[13] Jinjin Gu, Hannan Lu, Wangmeng Zuo, and Chao Dong. Blind super-resolution with iterative kernel correction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1604-1613, 2019. 1  
+[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2  
+[15] Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou. Drct: Saving image super-resolution away from information bottleneck. arXiv preprint arXiv:2404.00722, 2024. 6  
+[16] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 6  
+[17] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 6, 7  
+[18] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. Advances in Neural Information Processing Systems, 33:5632-5643, 2020. 1  
+[19] Yukun Huang, Xueyang Fu, Liang Li, and Zheng-Jun Zha. Learning degradation-invariant representation for robust real-world person re-identification. International Journal of Computer Vision, 130(11):2770–2796, 2022. 1  
+[20] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022. 2  
+[21] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, pages 920-928, 2023. 2  
+[22] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 2  
+[23] Deyi Ji, Wenwei Jin, Hongtao Lu, and Feng Zhao. Pptformer: Pseudo multi-perspective transformer for uav segmentation. International Joint Conference on Artificial Intelligence, pages 893–901, 2024. 2  
+[24] Deyi Ji, Feng Zhao, Lanyun Zhu, Wenwei Jin, Hongtao Lu, and Jieping Ye. Discrete latent perspective learning for segmentation and detection. In International Conference on Machine Learning, pages 21719-21730, 2024. 1  
+[25] Deyi Ji, Feng Zhao, Hongtao Lu, Feng Wu, and Jieping Ye. Structural and statistical texture knowledge distillation and learning for segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 1
+
+[26] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8346-8355, 2020. 8  
+[27] Rie Johnson and Tong Zhang. Inconsistency, instability, and generalization gap of deep neural network training. Advances in Neural Information Processing Systems, 36, 2024. 3  
+[28] Asif Hussain Khan, Christian Micheloni, and Niki Martinel. Lightweight prompt learning implicit degradation estimation network for blind super resolution. IEEE Transactions on Image Processing, 2024. 2  
+[29] Xiangtao Kong, Xina Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Reflash dropout in image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6002-6012, 2022. 1, 2, 3, 5, 6, 7  
+[30] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 6, 7  
+[31] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 8  
+[32] Feng Li, Yixuan Wu, Huihui Bai, Weisi Lin, Runmin Cong, and Yao Zhao. Learning detail-structure alternative optimization for blind super-resolution. IEEE Transactions on Multimedia, 25:2825-2838, 2022. 2  
+[33] Xiang Li, Shuo Chen, Xiaolin Hu, and Jian Yang. Understanding the disharmony between dropout and batch normalization by variance shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2682-2690, 2019. 3  
+[34] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1, 2, 6, 7  
+[35] Jingyun Liang, Kai Zhang, Shuhang Gu, Luc Van Gool, and Radu Timofte. Flow-based kernel prior with application to blind super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10601-10610, 2021. 1, 2  
+[36] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 2  
+[37] Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong. Blind image super-resolution: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5461-5480, 2022. 1  
+[38] Haofeng Liu, Heng Li, Huazhu Fu, Ruoxiu Xiao, Yunshu Gao, Yan Hu, and Jiang Liu. Degradation-invariant enhance
+
+ment of fundus images via pyramid constraint network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 507-516. Springer, 2022. 1  
+[39] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. IEEE Transactions on pattern analysis and machine intelligence, 2023. 8  
+[40] Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(12):3071-3085, 2018. 5  
+[41] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 416-423. IEEE, 2001. 6, 7  
+[42] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 6, 7  
+[43] Tomer Michaeli and Michal Irani. Nonparametric blind super-resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 945–952, 2013. 1  
+[44] Ari S Morcos, David GT Barrett, Neil C Rabinowitz, and Matthew Botvinick. On the importance of single directions for generalization. arXiv preprint arXiv:1803.06959, 2018. 3  
+[45] Behnam Neyshabur. Implicit regularization in deep learning. arXiv preprint arXiv:1709.01953, 2017. 1  
+[46] Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. 1  
+[47] Han Oh, Dongjin Kim, Sun Gu Lee, and Daewon Chung. Super-resolution of remote sensing imagery using implicit degradation modeling. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 5146-5149. IEEE, 2023. 2  
+[48] Atilla Özgür and Fatih Nar. Effect of dropout layer on classical regression problems. In 2020 28th Signal Processing and Communications Applications Conference (SIU), pages 1-4. IEEE, 2020. 2  
+[49] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 1  
+[50] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11908-11915, 2020. 8  
+[51] Yi Qin, Haitao Nie, Jiarong Wang, Huiying Liu, Jiaqi Sun, Ming Zhu, Jie Lu, and Qi Pan. Multi-degradation super
+
+resolution reconstruction for remote sensing images with reconstruction features-guided kernel correction. Remote Sensing, 16(16):2915, 2024. 2  
+[52] Hshmat Sahak, Daniel Watson, Chitwan Sahara, and David Fleet. Denoising diffusion probabilistic models for robust image super-resolution in the wild. arXiv preprint arXiv:2302.07864, 2023. 1, 2  
+[53] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 1, 2  
+[54] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, pages 114-125, 2017. 6, 7  
+[55] Hong Wang, Qi Xie, Qian Zhao, Yuexiang Li, Yong Liang, Yefeng Zheng, and Deyu Meng. Rcdnet: An interpretable rain convolutional dictionary network for single image deraining. IEEE Transactions on Neural Networks and Learning Systems, 2023. 8  
+[56] Hongjun Wang, Jiyuan Chen, Yinqiang Zheng, and Tieyong Zeng. Navigating beyond dropout: An intriguing solution towards generalizable image super resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25532-25543, 2024. 1, 2, 3, 6, 7, 8  
+[57] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 1, 2, 8  
+[58] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 6, 7  
+[59] Lijun Wu, Juntao Li, Yue Wang, Qi Meng, Tao Qin, Wei Chen, Min Zhang, Tie-Yan Liu, et al. R-drop: Regularized dropout for neural networks. Advances in Neural Information Processing Systems, 34:10890-10905, 2021. 1, 2  
+[60] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 2  
+[61] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 2  
+[62] Qingsen Yan, Axi Niu, Chaoqun Wang, Wei Dong, Marcin Wozniak, and Yanning Zhang. Kgsr: A kernel guided network for real-world blind super-resolution. Pattern Recognition, 147:110095, 2024. 2  
+[63] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE Transactions on Image Processing, 19(11):2861-2873, 2010. 6, 7
+
+[64] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27, 2014. 5  
+[65] Zongsheng Yue, Qian Zhao, Jianwen Xie, Lei Zhang, Deyu Meng, and Kwan-Yee K Wong. Blind image super-resolution with elaborate degradation modeling on noise and kernel. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2128-2138, 2022. 2  
+[66] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36:13294-13307, 2023. 8  
+[67] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 8  
+[68] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2  
+[69] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 286-301, 2018. 6  
+[70] Yongfei Zhang, Ling Dong, Hong Yang, Linbo Qing, Xiaohai He, and Honggang Chen. Weakly-supervised contrastive learning-based implicit degradation modeling for blind image super-resolution. Knowledge-Based Systems, 249:108984, 2022. 2

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:813a5e8f5711182a47e7006b29ec3eeaa50b643fdcab235cf2763bbc02686c2e
+size 900976

CVPR/2025/Adaptive Dropout_ Unleashing Dropout across Layers for Generalizable Image Super-Resolution/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e72983d3e24e9d0873f8e56bcd7d42a3294002efb558a758f3dbfca800389e2
+size 387661

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/f9180f95-5574-49b3-8480-091edde2c028_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20d327c429cc88eed8411e1cf585a5d661f565bd8e7aa105c5ec2befc1779fa5
+size 83163

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/f9180f95-5574-49b3-8480-091edde2c028_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71a7d309bc33954a2fbcd2db857db8b4168950b49a7817741ce1a999c99ffc1a
+size 111419

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/f9180f95-5574-49b3-8480-091edde2c028_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0c07de08dbe7ba4a4e10912e28b81652579a4f69ce93024290f3587a1e45e5b
+size 2212392

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/full.md

@@ -0,0 +1,371 @@
+# Adaptive Keyframe Sampling for Long Video Understanding
+
+Xi Tang $^{1*}$ , Jihao Qiu $^{1*}$ , Lingxi Xie, Yunjie Tian $^{2}$ , Jianbin Jiao $^{1}$ , Qixiang Ye $^{1}$ $^{1}$ University of Chinese Academy of Sciences  
+ $^{2}$ University at Buffalo, SUNY
+
+{tangxi19,qiujihao19}@mails.ucas.ac.cn {jiaojb,qxye}@ucas.ac.cn
+
+# Abstract
+
+Multimodal large language models (MLLMs) have enabled open-world visual understanding by injecting visual input as extra tokens into large language models (LLMs) as contexts. However, when the visual input changes from a single image to a long video, the above paradigm encounters difficulty because the vast amount of video tokens has significantly exceeded the maximal capacity of MLLMs. Therefore, existing video-based MLLMs are mostly established upon sampling a small portion of tokens from input data, which can cause key information to be lost and thus produce incorrect answers. This paper presents a simple yet effective algorithm named Adaptive Keyframe Sampling (AKS). It inserts a plug-and-play module known as keyframe selection, which aims to maximize the useful information with a fixed number of video tokens. We formulate keyframe selection as an optimization involving (1) the relevance between the keyframes and the prompt, and (2) the coverage of the keyframes over the video, and present an adaptive algorithm to approximate the best solution. Experiments on two long video understanding benchmarks validate that AKS improves video QA accuracy (beyond strong baselines) upon selecting informative keyframes. Our study reveals the importance of information pre-filtering in video-based MLLMs. Our codes are available at https://github.com/ncTimTang/AKS
+
+# 1. Introduction
+
+"You can't manage what you can't measure."
+
+Peter Drucker
+
+Recent years have witnessed a rapid development of multimodal large language models (MLLMs) [14, 22, 41, 60] for open-world visual understanding. Among the large corpus of research in MLLMs, a straightforward and important direction is to generalize them to video data. Compared to still images, videos contain richer and more complex visual
+
+![](images/7b0128813733e98617432ed8d189c6c8bb349bb92222f3d35a23366c8e297fce.jpg)  
+Figure 1. The accuracy of video-based MLLMs heavily relies on the quality of keyframes. The above example shows a long video from VideoMME [9] where keyframes are marked with green stars. The same MLLM (i.e., LLaVA-Video [62]) is used for answering the question. Uniform sampling (the default setting in [62]) finds irrelevant frames (the MLLM mostly performs a random guess), while our algorithm (AKS) finds keyframes and produces the correct answer.
+
+content, thus raising serious challenges to MLLMs including key information retrieval, summarization, logical inference, etc. Many benchmarks [9, 48] have been established to evaluate MLLMs for video understanding.
+
+A typical framework of image-based MLLMs involves encoding the input image into a set of visual tokens and feeding them as the context of LLMs. When this framework was transplanted to videos, especially long videos, a difficulty arose from the limited capacity of MLLMs, i.e., the maximal number of visual tokens that MLLMs can process is much fewer than that of an entire video; in other words, not all video tokens can be perceived by MLLMs. To bridge the gap, recent approaches [20, 49] often sam
+
+pled a small portion of frames from the input video; consequently, the performance of these MLLMs heavily relies on the quality of selected frames (i.e., keyframes). Despite its importance, the keyframe selection algorithm has not been carefully designed, e.g., LLaVA-Video [62] simply applied a uniform sampling strategy which, as shown in Figure 1, is prone to losing important information and thus leads to incorrect outputs of video understanding.
+
+This paper presents a systematic study on keyframe selection and reveals its importance to video understanding and beyond. We formulate keyframe selection as a plug-and-play module before the MLLM's visual encoder; its goal is to maximize the usefulness of the keyframes in video understanding. Intuitively, we propose two key aspects to be considered, namely, (1) relevance (i.e., how the keyframes are related to the question) and (2) coverage (i.e., how the keyframe set covers the useful information in the entire video). Specifically, we quantify the target by (1) computing relevance between each candidate frame and the prompt using a vision-language (VL) model, and (2) estimating coverage by recursively partitioning the video into bins and counting the number of keyframes within each bin. We show that maximizing relevance and coverage alone produces simple baselines for keyframe selection, while a proper tradeoff between them, obtained by the proposed Adaptive Keyframe Sampling (AKS) algorithm, leads to the best practice of video understanding.
+
+We evaluate our approach on LongVideoBench [48] and VideoMME [9], two benchmarks for long video understanding. We investigate three frame-based MLLMs (Qwen2VL [43], LLaVA-OV [15], and LLaVA-Video [62]) as the baseline and insert AKS as an off-the-shelf module to improve the quality of keyframes. Our approach achieves consistent accuracy gain throughout all tests. Specifically, when AKS is integrated with LLaVA-Video-7B, we set new records on these two benchmarks with 7B models. We further validate that the improvement owes to higher-quality keyframes found by AKS, demonstrating that MLLMs become stronger with more informative visual contexts. Our study reveals that pre-filtering visual data is crucial and will be a long-lasting research topic for MLLMs in perceiving high-dimensional data, e.g., long videos, and even 4D data.
+
+# 2. Related Work
+
+Large language models (LLMs) and multimodal LLMs (MLLMs). LLMs [2-5, 7, 38, 42, 54, 59] have marked a new era in AI, showcasing significant potential in unifying various tasks covering language understanding and generation. To extend LLMs for visual understanding, the community has focused on aligning visual and language data within a unified feature space [32]. There are generally two types of approaches, (1) internal adaptation, such as [1], that integrates cross-attention mechanisms within LLMs to
+
+achieve vision-language alignment, and (2) external adaptation, such as [6, 17, 22], that trains additional modules for the same purpose. Consequently, vision foundation models [8, 13, 24, 32, 39, 40, 61] have evolved into multimodal LLMs (MLLMs) [14, 22, 41, 60], enabling them to perform language-guided visual understanding tasks.
+
+Video-based MLLMs. Researchers have extended MLLMs to video understanding. Early efforts in this area include VideoChat [18], Video-ChatGPT [26], VideoLLaMA [56], Video-LLaVA [20], LanguageBind [63], and Valley [25], etc. Different from still images, videos contain rich content that, when encoded as visual tokens, exceed the maximal context capacity of MLLMs. Most of the above methods have sampled video frames to fit MLLMs; some of them, such as Video-ChatGPT [26], introduced more efficient video features. There are also studies on instance-level video understanding have been proposed, such as LEGO [19] for moment retrieval, PG-Video-LLaVA [27] for video grounding, and Artemis [31] for video referring, enriching the corpus of video understanding.
+
+MLLMs for Long Video Understanding. Going one step further, long video understanding faces greater challenges due to the increased difficulty of keyframe selection, leading to significant loss of critical information. While some MLLMs (e.g., Kangaroo [23], LLaVA-Video [62], LongVILA [50] and LongVA [58]) extend the context length of language model to allow more frames to be encoded and processed, many others have designed specific strategies to mitigate this issue.
+
+For example, MovieChat [36] employed both short-term and long-term memory banks to compress and preserve video content. Similarly, MA-LMM [11] and VideoStreaming [30] used a Q-former and a small language model (phi2 [12]), to condense video data, while LongVLM [47] and CrossGET [35] adopted token merging, VCWE [46] adopted token pooling and LongVU [34] adopted token compression to decrease the number of video tokens. These approaches aim to reduce the number of video tokens, but there is often no guarantee that key information in the video can be preserved. To identify the key information, Seeling [10] picked the most dominant frames, while ignoring the relevance to the prompt. VideoAgent [44] and VideoTree [45] opted to select keyframes solely using language (from the LLMs), which can miss important visual cues. This work presents a simple yet effective algorithm that maximally preserves important information for long video understanding.
+
+# 3. Method
+
+# 3.1. Preliminaries
+
+Within a broad range of video understanding tasks, the model receives a video clip and a text instruction as input
+
+![](images/8d5529a56867ed5f020a73dc05b5cf9ff46053041f432e0b07f46ba87b2e6434.jpg)  
+Figure 2. The overall framework of our approach. We insert a plug-and-play module, Adaptive Keyframe Sampling (AKS, marked in green frames) into the MLLM to improve the quality of sampled keyframes. Each red dot indicates a prompt-frame matching score (i.e.,  $r(\mathbf{Q}, \mathbf{F}_t)$ , see Section 3.2). AKS follows a recursive, judge-and-split optimization for keyframe selection (see Section 3.3).
+
+and is required to output a text answer. Without loss of generality, we denote the video as  $\mathbf{V} \in \mathbb{R}^{T \times W \times H \times C}$ , where  $T$  is the number of frames and  $W, H, \text{and } C$  denote the width, height, and number of channels, respectively. We consider each frame an image  $\mathbf{V}_t$  ( $t \in \{1, 2, \dots, T\}$ ) and apply a pre-trained encoder (e.g., the CLIP ViT-L model [32]) to extract a set of visual tokens  $\mathbf{F}_t$  from it. The text instruction (a.k.a., prompt) is denoted as  $\mathbf{Q}$ .
+
+The overall pipeline of our algorithm is illustrated in Figure 2. We use a regular MLLM that addresses video understanding with the template [User: <video-tokens> <text-instruction> Assistant: ], where the video tokens and text instruction are projected into the same feature space using an MLP. For simplicity, we denote the MLLM as a function of  $G(\{\mathbf{F}_t\})$  where we omit the LLM part and only focus on the visual tokens as contexts. With a limited capacity of visual contexts (i.e., the number of video tokens cannot exceed a specific value), the above pipeline encounters difficulty in dealing with long videos where not all video content can be perceived by the MLLM.
+
+A straightforward solution is to select keyframes from the input video for token extraction. In other words, the goal is to design a selection function  $\mathrm{KS}_M(\mathbf{Q},\mathbf{F})$  that outputs an index set,  $\mathcal{I}\subseteq \{1,2,\dots,T\}$  and  $|\mathcal{I}| = M$  , indicating the  $M$  best keyframes (  $M$  is pre-defined according to the context capacity of the MLLM). Video tokens extracted from the keyframes (i.e.,  $\{\mathbf{F}_t\mid t\in \mathrm{KS}_M(\mathbf{Q},\mathbf{F})\}$  compose of the context of the MLLM. As shown in Figure 1, the qual
+
+ity of keyframe selection is crucial for video understanding, but the function  $\mathrm{KS}_M(\mathbf{Q},\mathbf{F})$  has not been well studied in the community. As an example, a recent MLLM for video understanding [62] simply performed uniform sampling for keyframe selection; with the function  $\mathrm{KS}_M(\cdot)$  not using  $\mathbf{Q}$  and  $\mathbf{F}$  at all, it cannot guarantee to find useful information for question answering.
+
+In what follows, we establish two principles of keyframe selection (Section 3.2), after which we will present AKS, our optimization algorithm (Section 3.3).
+
+# 3.2. Principles of Keyframe Selection
+
+The keyframe selection function  $\mathrm{KS}_M(\mathbf{Q},\mathbf{F})$  is to maximize the amount of useful information, i.e.,
+
+$$
+\operatorname {K S} _ {M} (\mathbf {Q}, \mathbf {F}) = \arg \max  _ {| \mathcal {I} | = M} G ^ {\prime} \left(\left\{\mathbf {F} _ {t} \mid t \in \mathcal {I} \right\}\right). \tag {1}
+$$
+
+Here, we assume  $G'(\cdot)$  to be a complementary function of  $G(\cdot)$ , indicating the MLLM's confidence about its output. Eqn (1) is mathematically intractable due to two reasons. First, the optimization involves exponentially many candidates of  $\mathcal{I}$ . Second and more importantly, the function  $G'(\cdot)$  is difficult to estimate because there is no supervision for keyframe selection - even when a training set is available and one can compare the output of  $G(\cdot)$  with the ground-truth answer, it is not guaranteed that a correct answer corresponds to a perfect set of keyframes, and vice versa.
+
+We propose a heuristic method to approximate Eqn (1).
+
+# Q: How many outfits did the male protagonist change in total in the video?
+
+![](images/2c2d3576da6b71771e8c28345863fb3606bb6be4656f9c239dc3d81b6d01fbd5.jpg)  
+A: The male protagonist changed 5 outfitsits in total in the video.  
+Figure 3. An example of adaptive sampling (ADA). 8 keyframes are to be selected from the input video. Each red dot indicates a prompt-frame matching score,  $r(\mathbf{Q}, \mathbf{F}_t)$ . At Level-0 and Level-1, all bins are split into two sub-bins; at Level-2, only the rightmost bin is further partitioned while the top-2 scores are sampled from the other three bins. Level-3 has reached the maximal depth.
+
+Intuitively, a set of keyframes is informative when the following conditions are satisfied. (1) The relevance between each frame and the prompt is high, i.e., the visual data is useful for question answering. (2) The coverage of the selected frames is sufficient to comprehensively answer the question. Note that the coverage is difficult to quantify, and the second principle is closely related to preventing redundant frames (e.g., neighboring frames with almost the same visual content) from being selected, because (when the size of  $|\mathcal{I}|$  is fixed) they can potentially reduce the amount of other useful information and thus harm the coverage of the entire keyframe set.
+
+Following the analysis above, we reformulate the right-hand side of Eqn (1), yielding:
+
+$$
+\mathrm {K S} _ {M} (\mathbf {Q}, \mathbf {F}) = \arg \max  _ {| \mathcal {I} | = M} \sum_ {t \in \mathcal {I}} r (\mathbf {Q}, \mathbf {F} _ {t}) + \lambda \cdot c (\mathcal {I}). \tag {2}
+$$
+
+Here we introduce two quantities,  $r(\mathbf{Q},\mathbf{F}_t)$  as the relationship between the prompt  $\mathbf{Q}$  and the  $t$  -th frame  $\mathbf{F}_t$  ,and  $c(\mathcal{I})$  as the coverage of the entire keyframe set over the time axis.  $\lambda$  is the balancing hyper-parameter.
+
+Computing  $r(\mathbf{Q},\mathbf{F}_t)$ . This involves a vision-language (VL) module to measure whether  $\mathbf{F}_t$  contains information for answering  $\mathbf{Q}$ . Although the target MLLM itself  $G(\cdot)$  can play the role, its high computational cost can bring a major burden. In practice, we choose a cheaper VL model (e.g., CLIP [32] or BLIP ITM [16]) for replacement.
+
+Estimating  $c(\mathcal{I})$ . Measuring coverage is an open problem which is related to the homogeneity of data distribution. In
+
+mathematics, Ripley's  $K$ -function [33] is a popular way to measure homogeneity. Given the timestamp set  $\mathcal{I}$  and any search radius  $r < T$ , the  $K$ -function of  $r$ , denoted as  $\hat{K}(r)$ , is proportional to the number of  $(t_i, t_j)$  pairs satisfying  $|t_i - t_j| < r$ . The distribution of  $\mathcal{I}$  is considered homogeneous (i.e., covering the entire time axis) if  $\hat{K}(r)$  is approximately proportional to  $r^2$ .
+
+To adapt  $K$ -function to computing coverage (closely related but a bit different from homogeneity) as well as reducing computational overhead, we introduce bins with width  $r$  and approximate  $\mathbb{I}(|t_i - t_j| < r)$  as whether  $t_i$  and  $t_j$  fall into the same bin. We perform a recursive partition. At the first level, we set 2 bins with the bin width being  $T/2$ , i.e., the time axis  $[0, T)$  is partitioned into 2 non-overlapping bins,  $[0, T/2)$  and  $[T/2, T)$ . With the numbers of keyframes falling within these bins being  $m_1$  and  $m_2$ ,  $c(\mathcal{I})$  adds a penalty term  $|m_1 - m_2|$  since an uneven distribution implies weak coverage in the bin with fewer keyframes. At the second level, each of  $[0, T/2)$  and  $[T/2, T)$  is further partitioned into two bins, and the same calculation continues. The recursion stops at the  $L$ -th level, where  $L \leqslant \lceil \log_2 M \rceil$  is a hyper-parameter.
+
+# 3.3. Adaptive Keyframe Sampling
+
+With the complex definition of  $c(\mathcal{I})$ , it is difficult to find a closed-form or accurate optimization for Eqn (2). This part discusses an approximation. Compared to the baseline that only relies on  $r(\mathbf{Q}, \mathbf{F}_t)$  scores, we name such methods timestamp-aware optimization for its ability to consider
+
+timesteps for better keyframe selection results.
+
+We first discuss two special cases. (1) When  $\lambda = 0$  (i.e., coverage is neglected), Eqn (2) is solved by simply selecting the top-  $M$  frames of the largest scores. We name this strategy TOP, short for 'top sampling'; as shown in Figure 5, in some cases, it results in all keyframes being located within a small range of time and the MLLM missing important information in other moments. (2) When  $\lambda \rightarrow +\infty$  (i.e., coverage is strictly guaranteed), Eqn (2) is solved by selecting the frame within the highest score in each bin as the keyframe (when the number of bins exceeds  $M$ , the champion frames with the highest scores are preserved). We name this strategy BIN, short for 'binned sampling'. This situation further degenerates to the uniform sampling baseline [62] if a dummy VL model is used for scoring (i.e.,  $r(\mathbf{Q},\mathbf{F}_t)$  is a constant over  $t$ ). We name this strategy UNI, short for 'uniform sampling'.
+
+In other cases  $(0 < \lambda \ll +\infty)$ , we adopt a hierarchical optimization method that follows the definition of  $c(\mathcal{I})$ . At the first level, we determine how to allocate  $M$  keyframes into two bins,  $[0,T / 2)$  and  $[T / 2,T)$ . We recall the relevance scores of all frames,  $r(\mathbf{Q},\mathbf{F}_t)$ , and compute the average scores over all frames (denoted as  $s_{\mathrm{all}}$ ) and over  $M$  frames with the highest scores (denoted as  $s_{\mathrm{top}}$ ). If there is only one keyframe to be selected, or  $s_{\mathrm{top}} - s_{\mathrm{all}}$  surpasses a threshold,  $s_{\mathrm{thr}}$ , we believe that it is important to guarantee the top-scored frames to be sampled (i.e., maximizing the first term of Eqn (2)), so the algorithm directly returns the top- $M$  frames as keyframes. Otherwise, we split the current bin into two sub-bins with the number of keyframes evenly allocated (i.e., maximizing the second term of Eqn (2)), and then recursively call the above programs in the sub-bins. We name this strategy ADA, short for 'adaptive sampling'. Note that the hyper-parameter  $\lambda$  is not explicitly tuned; its role is replaced by  $s_{\mathrm{thr}}$ .
+
+Figure 3 uses an example to show how adaptive sampling (ADA) works. ADA is a compromise between the special cases, TOP and BIN. As we shall see in experiments (see Section 4.4), ADA absorbs the advantages of TOP and BIN and achieves the best practice of video understanding.
+
+# 4. Experiments
+
+# 4.1. Experimental Setup and Details
+
+Dataset and evaluation. We utilize the popular LMMsEval [57] to evaluate the performance of AKS. We use two popular benchmarks LongVideoBench [48], and VideoMME [9], for long video understanding. The length of videos in this dataset can exceed one hour, so the quality of keyframe selection plays a crucial role in visual understanding. We establish AKS beyond three recent video-based MLLMs (see the next paragraph). We do not tune the parameters of these MLLMs, but only change the input
+
+Table 1. Video-based question answering accuracy  $(\%)$  of different approaches on LongVideoBench (LVB) val and VideoMME (V-MME). AKS is applied upon three baseline approaches. Frames and LLM indicate the number of video frames fed into the MLLM and the number of parameters in the LLM part, respectively.  
+
+<table><tr><td>Method</td><td>Frames</td><td>LLM</td><td>LVB val</td><td>V-MME</td></tr><tr><td>GPT-4V [28]</td><td>256</td><td>-</td><td>61.3</td><td>59.9</td></tr><tr><td>GPT-4o [29]</td><td>256</td><td>-</td><td>66.7</td><td>71.9</td></tr><tr><td>Gemini-1.5-Flash [37]</td><td>256</td><td>-</td><td>61.6</td><td>70.3</td></tr><tr><td>Gemini-1.5-Pro [37]</td><td>256</td><td>-</td><td>64.0</td><td>75.0</td></tr><tr><td>VideoLLaVA [20]</td><td>8</td><td>7B</td><td>39.1</td><td>39.9</td></tr><tr><td>MiniCPM-V 2.6 [52]</td><td>64</td><td>8B</td><td>54.9</td><td>60.9</td></tr><tr><td>PLLaVA [49]</td><td>32</td><td>34B</td><td>53.2</td><td>-</td></tr><tr><td>VILA [21]</td><td>-</td><td>40B</td><td>-</td><td>60.1</td></tr><tr><td>Qwen2-VL [43]</td><td>32</td><td>7B</td><td>55.5</td><td>57.6</td></tr><tr><td>Qwen2-VL w/ AKS</td><td>32</td><td>7B</td><td>60.5</td><td>59.9</td></tr><tr><td>LLaVA-OV [15]</td><td>32</td><td>7B</td><td>54.8</td><td>56.5</td></tr><tr><td>LLaVA-OV w/ AKS</td><td>32</td><td>7B</td><td>59.3</td><td>58.4</td></tr><tr><td>LLaVA-Video [62]</td><td>64</td><td>7B</td><td>58.9</td><td>64.4</td></tr><tr><td>LLaVA-Video w/ AKS</td><td>64</td><td>7B</td><td>62.7</td><td>65.3</td></tr></table>
+
+frames into those selected by AKS. To highlight the importance of keyframe selection, we do not use video subtitles to assist question answering. This setting also allows us to weaken the impact of the LLM's strength and maximally focus on visual understanding.
+
+Implementation details. We investigate three video-based MLLMs as our baseline, namely, Qwen2VL [43], LLaVA-OV [15], and LLaVA-Video [62]. LongVideoBench and VideoMME contain multi-choice questions; to answer these questions, these MLLMs followed a similar prompt involving the question (in text), video frames (as tokens), and options (in text). Specifically, as the strongest baseline, LLaVA-Video used SigLIP [55] as its vision encoder and Qwen2-7B [51] as its large language model. With capacities to process up to 32 or 64 video frames, these MLLMs offer basic abilities of video understanding, but they were built upon uniformly sampled keyframes and can miss important information.
+
+To reduce computational costs, we sample the candidate frames from the raw video at 1 frame per second. The prompt (in text) and each  $t$ -th frame (as image) are fed into the text and visual encoders of BLIP [16] to obtain  $\mathbf{Q}$  and  $\mathbf{F}_t$ , after which  $r(\mathbf{Q},\mathbf{F}_t)$  is computed via image-text matching (ITM), i.e., the similarity between  $\mathbf{Q}$  and  $\mathbf{F}_t$ . One can also replace BLIP with other vision-language models (e.g., CLIP [32]); see the ablation in Section 4.4. We use ADA sampling unless otherwise specified.
+
+# 4.2. Comparison to the State-of-the-Art
+
+Quantitative results. We first compare the accuracy of video question answering between our approach and
+
+![](images/87e4ef58bbfc2be99d17857885577043862133052b3de7299c251acbc264695a.jpg)  
+Q: After entering the museum, where did the short-haired woman in a black coat, black long skirt, and black mask visit first?  
+Qwen2-VL-7B  
+Qwen2-VL-7B w/AKS  
+A: Bronze statues.  
+A: Sculptures.
+
+![](images/da0cc0fabf0a0f32973553314f4dee641f3879c98f2c8e82b0dc9db65aa42c7b.jpg)  
+Qwen2-VL-7B w/ AKS
+
+![](images/b0233adb69d5bf98d7c72448d0323f782ab0bbd0f8219dbd2597797d62c12294.jpg)  
+Q: How long does it take to drive from the Earth to the Moon according to the video?  
+Qwen2-VL-7B  
+A:19 days.  
+A:160 days.
+
+![](images/973ac8bd228aa8ea1ec8bb920eec799dc3db3075897605f76009591dad87c5db.jpg)
+
+![](images/344de7c6d13ddf4701370f3b6d93141070f0e01f8907f506eb3c4e63720d09f9.jpg)  
+Q: A white-lettered title says 'How much data do we need?' The word 'dog' appears on the far right. When the image of a shaking dog appears, what changes occur in the black scene?  
+LLaVA-OV-7B  
+LLaVA-OV-7B w/ AKS
+
+![](images/a009ce9c874a76eef82aa2e584cdbc0cc97a44d5f67f30adb729ffa829710fde.jpg)  
+A: The image of the shaking dog gradually enlarges.  
+A: The arrow on the right  
+side is labeled 'not dog'  
+LLaVA-OV-7B w/ AKS
+
+![](images/86a1f88c69c31d508a8d6c71384e6869360029c7ea3cf9153cdc0450b530ce92.jpg)  
+Q: What are the moves in the last scene of this dance?  
+LLaVA-OV-7B  
+A: Passe and then chasse.
+
+![](images/56dc67ee4558b93aae2237af1e89287b962167ca2553ee9f024a0687533fb2dd.jpg)  
+Q: Two men wearing straw hats and grey clothes stand in a grass field holding long knives. What does the house behind them look like?  
+LLaVA-Video-7B  
+LLaVA-Video-7B w/ AKS  
+A:Wooden house.  
+A: Straw hut.
+
+![](images/e595a01ac6da574940c206cf887326c21613639b2f05187be397c741de79407b.jpg)  
+LLaVA-Video-7B w/AKS  
+Figure 4. AKS improves the baseline MLLMs for video understanding. The left three examples come from LongVideoBench while the right three come from VideoMME. Green stars indicate keyframes selected by AKS (note that 64 keyframes are selected for each video).
+
+![](images/0539bd3a383ee61448e5abeee37e13910c46abb3763f696baa7a7113633c0133.jpg)  
+Q: What does the monkey do to the woman?  
+LLaVA-Video-7B  
+A: Climbing on her.  
+A:Pulling her hair.
+
+![](images/92acfbe02502d2e8434992e74d9b4a535a01b54519b03e0230469e32a8c52d69.jpg)  
+A:Kneel down on one  
+knee and lean back.
+
+some recent MLLMs. Results are summarized in Table 1. AKS brings consistent accuracy gain over three baselines, e.g., upon Qwen2VL, the improvement is  $5.0\%$  on LongVideoBench and  $2.3\%$  on VideoMME; even upon LLaVA-Video, the strongest baseline, these numbers are  $3.8\%$  and  $0.9\%$ , respectively. These improvements not only make our method surpass other competitors with a similar computational complexity (i.e., input no more than 64 frames, LLM no larger than 7B), but also allow it to achieve higher levels set by larger models (e.g., with AKS, LLaVA-Video-7B reports  $62.7\%$  on LongVideoBench, which is  $0.8\%$  higher than the LLaVA-Video-72B model without AKS, and  $1.4\%$  and  $1.1\%$  higher than GPT-4V and Gemini1.5-Flash, two proprietary models using 256 input frames). We also compare AKS with other long video understanding methods. Please see the supplementary material for more details.
+
+Qualitative results. In Figure 4, we display representative video understanding results of AKS (based on LLaVA-Video-7B) and others. One can see that the selected keyframes are closely related to the question; this allows the MLLM, with a limited capacity of context, to get a comprehensive view of question-related content and thus obtain the correct answer. As a side comment, we find that VideoMME contains many questions that require a high-level comprehension of the video content in which uniform sampling is a safe choice; nevertheless, AKS still finds more
+
+Table 2. Video-based question answering accuracy (\%) of different sampling strategies. LLaVA-Video-7B with AKS is tested. Please refer to Section 3.3 for the explanations of these abbreviations and Section 4.3 for the analysis of results.
+
+<table><tr><td>Sampling</td><td>LongVideoBench val</td><td>VideoMME</td></tr><tr><td>UNI</td><td>58.9</td><td>64.4</td></tr><tr><td>TOP</td><td>62.4</td><td>63.7</td></tr><tr><td>BIN</td><td>60.2</td><td>65.2</td></tr><tr><td>ADA</td><td>62.7</td><td>65.3</td></tr></table>
+
+informative frames and improves the accuracy, although the gain is smaller than that on LongVideoBench. Please also see the supplementary material for more examples.
+
+# 4.3. Diagnostic on Keyframe Selection
+
+This part aims to diagnose how AKS works and ablates design choices of AKS. We build our test upon the strongest baseline, LLaVA-Video.
+
+MLLMs benefit from better keyframes. To show how keyframe selection impacts video understanding, we test different strategies described in Section 3.3. Table 2 lists the results. Beyond the baseline (i.e., UNI sampling), ADA sampling (our default choice in Section 4.2) achieves the best practice, while each of TOP and BIN samplings is better than the other in one benchmark. Note that the MLLM (i.e., LLaVA-Video-7B) remains unchanged throughout all these tests. In other words, all the improvements owe to
+
+![](images/78e40caa9915e93c0c1b5478203c5f4a3ec5d30d90a6540dcfa2f81fa5dd2a82.jpg)  
+Figure 5. Two examples of how different sampling strategies impact video understanding. The left case comes from LongVideoBench (focusing on one moment) and the right one comes from VideoMME (relying on multiple moments). Each curve shows the  $r(\mathbf{Q}, \mathbf{F}_t)$  score over time, and the yellow circles indicate the position of sampled keyframes. We also annotate the number of true keyframes and the reason for each failure case below the answer.
+
+AKS in selecting higher-quality keyframes.
+
+Visualizing keyframe selection. Figure 5 shows two representative examples and explains how the style of questions varies across LongVideoBench and VideoMME and how it impacts the preference between TOP and BIN. Many questions of LongVideoBench are focused on a simple moment (e.g., 'What is a person doing at a specific time point?'), so TOP sampling (i.e., without constraints in temporal distribution) often works well in locating these keyframes, while BIN sampling limits the number of keyframes within each bin and results in information loss. On the contrary, the questions of VideoMME often require the model to collect information from multiple moments (e.g., 'How many times does something happen?'), so BIN sampling is a safe choice to locate keyframes in different bins, while TOP sampling can lose information in weak peaks. ADA sampling absorbs the advantages of TOP and BIN strategies and adaptively allocates keyframes to the desired position (see the example in Figure 3) – this is why it achieves the best results in both benchmarks.
+
+Figure 6 shows an interesting example that, on the same input video, AKS selects different sets of keyframes based on the prompt. This increases the flexibility that a frozen MLLM can adapt to different scenarios.
+
+# 4.4. Ablative Studies
+
+The frequency of sampling keyframe candidates. To decrease the extra computational cost brought by AKS, we sample fewer keyframe candidates (i.e., one frame per 2/4/8/10 seconds, and compare the results with the standard 1-fps method. Results are summarized in Table 3. On LongVideoBench, even at 0.1 fps, all results are higher
+
+![](images/cbe540af45097accb022ba94b2fae5615df6726177c30327e1fdb2ae83b1ee5c.jpg)  
+Q: How many cubs is the mother bear with when the filmer encounters her downstream of the river?
+
+![](images/691e801bdcb8c520444e6d063ddf7524414b41f98f9c312ebf1457fa32f46d8a.jpg)  
+A: 2 cubs are with the mother bear.
+
+Q: What is the main food of the brown bear in the video?
+
+![](images/105dea74b28bf1cbc53e0cb575a1ba65deeb92fe2b9824763cddcef3cb4bf3a3.jpg)  
+A:Salmon.
+
+Q: Why does the mother brown bear in make her cubs run to higher ground?
+
+![](images/729a92480eacc5d13b8bb7d78c284b0186c620f69ac644db333f40aa43872ec1.jpg)  
+A: Because she spotted two adult male brown bears that might kill the cubs.  
+Figure 6. AKS selects different keyframe sets to answer different questions. All answers are correct.
+
+than the baseline (i.e.,  $57.4\%$ ,  $57.9\%$ ,  $58.9\%$  at 16, 32, 64 frames). VideoMME shows a similar trend, and 0.25 fps seems a safe option to surpass the baseline. It is worth exploring more efficient pre-filtering algorithms towards a better tradeoff between accuracy and efficiency.
+
+The VL model for computing  $r(\mathbf{Q},\mathbf{F}_t)$  scores. We analyze the impact of using different VL models for computing prompt-frame relevance. We study three options, i.e., BLIP [16] (the default choice in this paper), Sevilla (used in [53]), and CLIP [32]. Results are summarized in Table 4. We find that BLIP works better on LongVideoBench while CLIP works better on VideoMME. This is because
+
+Table 3. Question answering accuracy  $(\%)$  w.r.t. different sampling frequencies. LLaVA-Video-7B is used as the MLLM.  
+
+<table><tr><td rowspan="2">Frames of MLLM</td><td rowspan="2">1</td><td colspan="4">Sampling Frequency (fps)</td></tr><tr><td>0.5</td><td>0.25</td><td>0.125</td><td>0.1</td></tr><tr><td colspan="6">LongVideoBench val</td></tr><tr><td>16</td><td>61.6</td><td>60.7</td><td>60.6</td><td>61.1</td><td>59.4</td></tr><tr><td>32</td><td>61.9</td><td>62.1</td><td>59.8</td><td>60.2</td><td>58.5</td></tr><tr><td>64</td><td>62.7</td><td>62.2</td><td>61.8</td><td>60.1</td><td>60.1</td></tr><tr><td colspan="6">VideoMME</td></tr><tr><td>16</td><td>62.2</td><td>63.0</td><td>62.2</td><td>61.0</td><td>61.6</td></tr><tr><td>32</td><td>64.6</td><td>64.7</td><td>65.1</td><td>64.4</td><td>64.4</td></tr><tr><td>64</td><td>65.3</td><td>65.1</td><td>64.9</td><td>64.0</td><td>64.2</td></tr></table>
+
+Table 4. Question answering accuracy  $(\%)$  w.r.t. different VL scorers. LLaVA-Video-7B is used as the MLLM.  
+
+<table><tr><td>Frames</td><td>Uniform</td><td>BLIP</td><td>Sevila</td><td>CLIP</td></tr><tr><td colspan="5">LongVideoBench val</td></tr><tr><td>16</td><td>57.4</td><td>61.6</td><td>59.2</td><td>60.2</td></tr><tr><td>32</td><td>57.9</td><td>61.9</td><td>60.9</td><td>61.9</td></tr><tr><td>64</td><td>58.9</td><td>62.7</td><td>61.5</td><td>62.2</td></tr><tr><td colspan="5">VideoMME</td></tr><tr><td>16</td><td>60.6</td><td>62.2</td><td>63.0</td><td>63.1</td></tr><tr><td>32</td><td>63.9</td><td>64.6</td><td>63.7</td><td>65.0</td></tr><tr><td>64</td><td>64.4</td><td>65.3</td><td>65.1</td><td>65.6</td></tr></table>
+
+Table 5. Ablating  $L$  and  ${s}_{\text{thr }}$  together. Left: LVB,Right: V-MME.  
+
+<table><tr><td>L</td><td>sthr</td><td>0.0</td><td>0.2</td><td>0.4</td><td>0.6</td><td>0.8</td><td>1.0</td></tr><tr><td>1</td><td>62.4/63.8</td><td>62.4/64.0</td><td>62.5/64.2</td><td>62.0/64.1</td><td>61.8/63.8</td><td>61.9/64.0</td><td></td></tr><tr><td>2</td><td>62.4/63.8</td><td>62.0/64.0</td><td>62.4/64.0</td><td>61.8/63.5</td><td>61.7/63.4</td><td>62.0/63.6</td><td></td></tr><tr><td>3</td><td>62.4/63.8</td><td>62.7/64.0</td><td>62.6/54.5</td><td>62.1/64.4</td><td>62.2/64.4</td><td>62.1/64.4</td><td></td></tr><tr><td>4</td><td>62.4/63.8</td><td>62.6/64.1</td><td>62.7/64.3</td><td>62.2/64.9</td><td>62.1/65.0</td><td>62.2/65.0</td><td></td></tr><tr><td>5</td><td>62.4/63.8</td><td>62.6/64.1</td><td>62.2/64.7</td><td>61.7/65.0</td><td>61.3/65.3</td><td>61.7/65.2</td><td></td></tr><tr><td>6</td><td>62.4/63.8</td><td>62.6/64.0</td><td>62.3/64.5</td><td>61.8/65.0</td><td>61.3/65.1</td><td>61.4/65.1</td><td></td></tr></table>
+
+CLIP was trained on generic image-text pairs while BLIP learned from object-level data - correspondingly, questions in LongVideoBench and VideoMME are more focused on objects and global perception, respectively.
+
+ADA hyper-parameters,  $L$  and  $s_{\mathrm{thr}}$ . Lastly, we study the impact of  $L$  and  $s_{\mathrm{thr}}$ . Results are summarized in Table 5. One can see that LongVideoBench prefers smaller  $L$  and  $s_{\mathrm{thr}}$  values than VideoMME. This is because the key information on LongVideoBench is more concentrated (i.e., many questions are related to single moments) while that on VideoMME is more diverse (multi-moment data are required for question answering). AKS offers a flexible ability to switch between different 'modes' and achieves better results in both datasets.
+
+# 4.5. Generalization to Other Tasks
+
+Being an off-the-shelf algorithm, AKS is easily applied to other video understanding tasks. Here we showcase two examples known as video referring and captioning. For this
+
+Q: What is the woman in sunglasses doing in the video?
+
+Referring
+
+![](images/ae56309c0fe82fc62d9188786789c6163b1291268d2104d984c81cfc7f0026f0.jpg)  
+Figure 7. Extending AKS for video referring and captioning. The baseline results are produced upon uniform keyframe sampling. Red and green texts indicate incorrect and correct descriptions.
+
+Baseline: The woman in sunglasses is walking through a lush garden with greenery and flowers.  
+w/AKS: The woman in sunglasses is sitting at a table with a drink, possibly enjoying the view.
+
+Q: What is the shirtless man doing in the video?
+
+Referring
+
+![](images/89c7cee228b0aed1eae054f155ff7fccc05a87d082751324b6356195c038c484.jpg)
+
+Baseline: The shirtless man shirt is leading the donkeys.
+
+w/AKS: The shirtless man is standing in a small, rectangular pool with turquoise water.
+
+Q: Please describe this video about a temple.
+
+Captioning
+
+![](images/7f96803ba199c5a59ea274b7a068e02bf1e6eb2dcb22f0df6bfec667eacdf202.jpg)
+
+Baseline: The video opens with a view of a lush tropical landscape, featuring tall palm trees and a mountain range in the background. The sky is overcast, and there's a sense of tranquility. As the video progresses, the camera pans to reveal more of the forested area, with dense foliage and a few buildings visible in the distance. The colors are vibrant, with greens of various shades dominating the scene. There are no visible texts or subtitles in the video.
+
+w/AKS: The video opens with a blurry shot of a temple structure, followed by a clearer view of the temple's intricate carvings and statues. The temple appears to be made of stone and is adorned with detailed sculptures of mythical creatures and deities. The sky is visible in the background, suggesting that the temple is outdoors. The video then transitions to a close-up of a statue, highlighting its craftsmanship and the texture of the stone. The final frames show a wider shot of the temple, emphasizing its grandeur and the surrounding natural environment.
+
+purpose, we use the LLaVA-Video-7B model, switch the text prompt into 'What is [target] doing in the video?' or 'Please describe this video.', and remove the options. Qualitative results are shown in Figure 7. As seen, to obtain a comprehensive description of long videos, it is crucial to locate keyframes and feed them into the MLLM as visual contexts. AKS benefits from its keyframe selection ability and helps the MLLM to generate much better answers.
+
+# 5. Conclusions
+
+This paper focuses on improving the ability of MLLMs for long video understanding. The main difficulty arises from the limited capacity of MLLMs which urges us to feed informative visual tokens into the model. For this purpose, we present the Adaptive Keyframe Sampling (AKS) algorithm which (1) uses a vision-language model to estimate the relevance and (2) applies an adaptive optimization algorithm to facilitate the coverage of selected keyframes. Quantitative and qualitative studies validate the effectiveness of AKS over different baselines and benchmarks. Our work reveals that a pre-filtering stage brings considerable benefit to video understanding and advocates for further studies in this direction.
+
+# Acknowledgements
+
+This work was supported by National Natural Science Foundation of China (NSFC) under Grant 62225208 and 62450046 and CAS Project for CAS Project for Young Scientists in Basic Research under Grant No.YSBR-117.
+
+# References
+
+[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2  
+[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2  
+[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023.  
+[4] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.  
+[5] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2  
+[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 2  
+[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2  
+[8] 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. arXiv preprint arXiv:2010.11929, 2020. 2  
+[9] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 1, 2, 5  
+[10] Wei Han, Hui Chen, Min-Yen Kan, and Soujanya Poria. Self-adaptive sampling for accurate video question answer
+
+ing on image text models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2522-2534, 2024. 2  
+[11] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 2  
+[12] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. *Microsoft Research Blog*, 2023. 2  
+[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything. arXiv preprint arXiv:2304.02643, 2023. 2  
+[14] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. arXiv preprint arXiv:2308.00692, 2023. 1, 2  
+[15] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 5  
+[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4, 5, 7  
+[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. arXiv preprint arXiv:2301.12597, 2023. 2  
+[18] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 2  
+[19] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Van Tu Vu, Zhida Huang, and Tao Wang. Groundinggpt:language enhanced multi-modal grounding model, 2024. 2  
+[20] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2023. 1, 2, 5  
+[21] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 5  
+[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. arXiv preprint arXiv:2304.08485, 2023.1,2  
+[23] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting
+
+long-context video input. arXiv preprint arXiv:2408.15542, 2024.2  
+[24] 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 Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2  
+[25] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability, 2023. 2  
+[26] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 2  
+[27] Shehan Munasinghe, Rusiru Thushara, Muhammad Maaz, Hanoona Abdul Rasheed, Salman Khan, Mubarak Shah, and Fahad Khan. Pg-video-llava: Pixel grounding large video-language models, 2023. 2  
+[28] OpenAI. Gpt-4v. https://openai.com/index/gpt-4v-system-card/, 2023.5  
+[29] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.5  
+[30] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024. 2  
+[31] Jihao Qiu, Yuan Zhang, Xi Tang, Lingxi Xie, Tianren Ma, Pengyu Yan, David Doermann, Qixiang Ye, and Yunjie Tian. Artemis: Towards referential understanding in complex videos. arXiv preprint arXiv:2406.00258, 2024. 2  
+[32] 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. 2, 3, 4, 5, 7  
+[33] Brian D Ripley. The second-order analysis of stationary point processes. Journal of applied probability, 13(2):255-266, 1976. 4  
+[34] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024. 2  
+[35] Dachuan Shi, Chaofan Tao, Anyi Rao, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Crossget: Cross-guided ensemble of tokens for accelerating vision-language transformers. arXiv preprint arXiv:2305.17455, 2023. 2  
+[36] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 2  
+[37] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a
+
+family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5  
+[38] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2  
+[39] Yunjie Tian, Lingxi Xie, Xiaopeng Zhang, Jiemin Fang, Haohang Xu, Wei Huang, Jianbin Jiao, Qi Tian, and Qixiang Ye. Semantic-aware generation for self-supervised visual representation learning. arXiv preprint arXiv:2111.13163, 2021. 2  
+[40] Yunjie Tian, Lingxi Xie, Zhaozhi Wang, Longhui Wei, Xiaopeng Zhang, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Integrally Pre-Trained Transformer Pyramid Networks. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18610-18620. IEEE, 2023. 2  
+[41] Yunjie Tian, Tianren Ma, Lingxi Xie, Jihao Qiu, Xi Tang, Yuan Zhang, Jianbin Jiao, Qi Tian, and Qixiang Ye. Chatterbox: Multi-round multimodal referring and grounding. arXiv preprint arXiv:2401.13307, 2024. 1, 2  
+[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2  
+[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5  
+[44] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2025. 2  
+[45] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2  
+[46] Hongchen Wei and Zhenzhong Chen. Visual context window extension: A new perspective for long video understanding. arXiv preprint arXiv:2409.20018, 2024. 2  
+[47] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. arXiv preprint arXiv:2404.03384, 2024. 2  
+[48] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. arXiv preprint arXiv:2407.15754, 2024. 1, 2, 5  
+[49] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 5  
+[50] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang
+
+Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 2  
+[51] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5  
+[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 5  
+[53] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 7  
+[54] Aohan Zeng, Xiao Liu, Zhengxiao Du, Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yifan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b: An open bilingual pre-trained model. arXiv preprint arXiv:2210.02414, 2022. 2  
+[55] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 5  
+[56] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 2  
+[57] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, et al. Lmms-eval: Reality check on the evaluation of large multimodal models. arXiv preprint arXiv:2407.12772, 2024. 5  
+[58] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2  
+[59] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2  
+[60] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest. arXiv preprint arXiv:2307.03601, 2023. 1, 2  
+[61] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2022. 2  
+[62] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 1, 2, 3, 5  
+[63] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zong-wei Li, Wancai Zhang, Zhifeng Li, Wei Liu, and Li Yuan.
+
+Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment, 2024. 2

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:562a8a356196f3a051f65c6fa95d8c41172cc9fec7bd82c4d5f25b2e8b82c383
+size 737501

CVPR/2025/Adaptive Keyframe Sampling for Long Video Understanding/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:559df445f8dc368de76631a128e2a4af6d478d8ae1409ad7db8a8901299c10e9
+size 501573

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/dbe47fad-c46f-4091-84d0-ae0a68fd8d7a_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ec1bad256897cf2e2192351812822a59cd082d6fb05fab1ee4e9c2653ca6fe2
+size 75809

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/dbe47fad-c46f-4091-84d0-ae0a68fd8d7a_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35f801ba1779b32e20ef2df1ca6bedd5184cf90eddd28f0febf3226f803db618
+size 95712

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/dbe47fad-c46f-4091-84d0-ae0a68fd8d7a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05d7ccfd808b715a4237ef512d93adb2b46cc0b31ab612ab65b7f636e00a54f2
+size 2006309

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/full.md

@@ -0,0 +1,258 @@
+# Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding
+
+Han Xiao $^{1,2\dagger}$ , Yina Xie $^{2}$ , Guanxin Tan $^{2}$ , Yinghao Chen $^{2}$ , Rui Hu $^{2}$ , Ke Wang $^{1}$ , Aojun Zhou $^{1}$ , Hao Li $^{1}$ , Hao Shao $^{1}$ , Xudong Lu $^{1,2\dagger}$ , Peng Gao $^{4}$ , Yafei Wen $^{2}$ , Xiaoxin Chen $^{2}$ , Shuai Ren $^{2\dagger \boxtimes}$ , Hongsheng Li $^{1,3 \boxtimes}$ $^{1}$ CUHK MMLab  $^{2}$ vivo AI Lab  $^{3}$ CPII under InnoHK  
+ $^{4}$ Shanghai AI Lab & Shenzhen Institute of Advanced Technology, CAS {1155229123@link,hsli@ee}.cuhk.edu.hk shuai. ren@vivo.com
+
+# Abstract
+
+Visual Document Understanding has become essential with the increase of text-rich visual content. This field poses significant challenges due to the need for effective integration of visual perception and textual comprehension, particularly across diverse document types with complex layouts. Moreover, existing fine-tuning datasets for this domain often fall short in providing the detailed contextual information for robust understanding, leading to hallucinations and limited comprehension of spatial relationships among visual elements. To address these challenges, we propose an innovative pipeline that utilizes adaptive generation of markup languages, such as Markdown, JSON, HTML, and TiKZ, to build highly structured document representations and deliver contextually-grounded responses. We introduce two fine-grained structured datasets: DocMark-Pile, comprising approximately 3.8M pretraining data pairs for document parsing, and DocMark-Instruct, featuring 624k fine-tuning data annotations for grounded instruction following. Extensive experiments demonstrate that our proposed model significantly outperforms existing state-of-the-art MLLMs across a range of visual document understanding benchmarks, facilitating advanced reasoning and comprehension capabilities in complex visual scenarios. Our code and models are released at https://github.com/Euphoria16/DocMark.
+
+# 1. Introduction
+
+Text-rich visual content, including multi-page documents and various multimedia formats, is ubiquitous in our daily
+
+lives. Consequently, Visual Document Understanding has emerged as a fundamental task that encompasses document classification, information extraction, and document-related visual question answering. This area presents significant challenges, as it requires seamless integration of visual perception with textual comprehension, along with the ability to navigate the complex interactions between these modalities. Such capabilities not only represent a fundamental cognitive skill in humans, but are also essential for the advancement of artificial intelligence systems.
+
+In recent years, Multimodal Large Language Models (MLLMs) have demonstrated remarkable progress in handling multi-modal tasks. Notably, there has been a growing focus on enhancing the capabilities of MLLMs in visual document comprehension. Existing efforts mainly aim to improve the representation of vision encoders originally designed for natural images, which are usually trained on relatively low-resolution datasets. Such efforts include scaling up the input resolution [6, 13, 26, 29], retraining vision encoders on text-rich images [7, 32, 54], and leveraging compositions of different encoders [25, 27].
+
+The core challenge in performing visual document understanding tasks lies in the efficient representation and effective perception of the visual content and its layout information present in the images. However, unlike Large Language Models (LLMs), whose training data generally include comprehensive descriptions and detailed reasoning, common fine-tuning datasets in the MLLM domain usually contain only questions and brief answers. As shown in Fig. 1, answers derived from PDF documents often consist of only a short phrase or a single number, even though understanding various parts of the document is required to answer the question. Similarly, datasets related to charts often lack a structural interpretation of the image content, resulting in challenges for effective analysis. Fine-tuning on
+
+![](images/34bfee56795dfd40e1b7d3d37f09d26380cc6397cb452e7b13088e405c6adde4.jpg)
+
+![](images/4ae6a0ec47d169112228128003912dc4fb9468b38cae7469df33873d00c47407.jpg)  
+Figure 1. Comparison of existing Data annotations and model Predictions: InternVL2-2B vs. our DocMark-2B. Existing datasets consist primarily of questions and brief answers, often resulting in model hallucinations. In contrast, our model effectively converts documents into structured markup languages, offering critical contextual information that enhances the accurate interpretation of the content.
+
+such data frequently leads to hallucinations and insufficient understanding of the contextual and spatial relationships among various visual elements in the documents. Moreover, visual documents come in a wide range of types, exhibiting different semantic content, multiple levels of granularity, and complex layouts. This variety requires adaptive encoding and perception of document images to fully capture their inherent complexity. These observations raise a critical question: How can we build contextually grounded training data and devise effective training pipelines to generalize across a wide range of document images?
+
+In this paper, we propose the adaptive generation of markup languages by the MLLM to facilitate a comprehensive understanding and provision of explicit context for various visual documents. Document images are often linked to their source codes, termed markup languages, which are both well-structured and easy to read. Examples include Markdown or LaTeX for PDF documents, HTML for webpages, and TikZ for scientific diagrams—all preserving rich semantic and layout information crucial for describing diverse document types. Furthermore, their structured nature and clarity make them readily interpretable by pretrained language models, providing valuable contextual information for enhanced document comprehension. To achieve this goal, we introduce DocMark-Pile, a compre
+
+hensive pretraining dataset of 3.8M paired data, designed to parse documents into their corresponding markup languages. This dataset covers a diverse range of visual document types, including natural scene images, dense documents, receipts, LaTeX formulas, web pages, tables, charts, and mathematical diagrams. The markup languages extend beyond basic text recognition and grounding, enabling code generation in formats such as Markdown, LaTeX, HTML, JSON, and TikZ. Building on the pretrained MLLM, we develop an adaptive generation pipeline by fine-tuning on DocMark-Instruct, a chain-of-thought-like reasoning dataset consisting of 624k annotations. Rather than manually specify which markup language to generate by the pretrained model, we let the model itself identify the most suitable type for each document, extracting useful context as intermediate rationales for answering questions. This approach mimics human-like reasoning, as a person first scans a visual document, then localizes and retrieves contextually relevant information to address specific queries. Similarly, when presented with a question, our model generates the most suitable markup languages, such as HTML for webpages, to extract relevant information and formulate answers. This fine-tuning pipeline enables our model to automatically adapt the appropriate markup language to the document type. Furthermore, it simulates human-like reasoning by disentangling perception from reasoning, encouraging the model to provide essential rationales for accurate answers. Extensive experiments demonstrate that our models substantially outperform existing state-of-the-art MLLMs on several challenging visual document understanding tasks.
+
+Our contributions can be summarized as follows:
+
+1. We propose a novel pipeline that adaptively uses various markup languages to bridge the gap between visual inputs and linguistic understanding and significantly enhance document comprehension capabilities.  
+2. We introduce two fine-grained structured datasets, DocMark-Pile and DocMark-Instruct, tailored for pretraining in specialized document parsing and fine-tuning on contextually-grounded instruction following. These datasets enable our model to effectively process complex document formats, including Plain Text, Markdown, LaTeX, HTML, JSON, and TikZ.  
+3. Extensive experiments demonstrate that our model consistently outperforms state-of-the-art multimodal LLMs across various model sizes, excelling in both markup parsing and downstream reasoning tasks.
+
+# 2. Related Work
+
+# 2.1. Multi-modal large language models (MLLMs).
+
+The field of Natural Language Processing (NLP) has seen remarkable advancements, particularly with the rise of
+
+Large Language Models (LLMs) [39, 40, 43]. The GPT series by OpenAI [39] has harnessed the power of massive model scaling, featuring architectures with billions and even trillions of parameters. To enhance instruction-following abilities, models like InstructGPT [41] and ChatGPT [39] have been developed, exhibiting exceptional fluency and versatility across various open-domain conversation tasks. Moreover, the instruction tuning based on open-source LLMs [51] has gained great popularity. Motivated by the advancements achieved by LLMs, recent research [12, 27, 44] has explored the application of LLMs for multi-modal tasks. Prevalent methods involve adapting the Linear Projector [28] or QFormer [24] to align vision encoder outputs with the textual feature space of LLMs. Nevertheless, existing methods struggle with comprehensive scene understanding due to limitations in visual representations. To mitigate this, several strategies have been proposed: utilizing composition of multiple vision encoders [25, 27], employing visual encoders designed for high-resolution inputs [13], and implementing dynamic resolution strategies to partition high-resolution images into multiple low-resolution segments [6, 26, 29]. While these strategies have shown advancements, they still struggle with complex understanding and reasoning tasks. This shortcoming arises from the fact that the pretraining and finetuning datasets used for MLLM training typically contain limited contextual information and lack a thorough reasoning process. In this paper, we address this gap by designing an adaptive markup generation pipeline aimed at producing explicit contextual information, which is crucial for effectively performing comprehensive reasoning tasks.
+
+# 2.2. Visual Document Understanding
+
+Recent advancements in multimodal large language models (MLLMs) have significantly enhanced visual document understanding tasks by employing autoregressive training on document images. UReader [59] introduces a Shape-adaptive Cropping Module to divide high-resolution images into low-resolution sub-images, effectively capturing rich text information. Similarly, Monkey utilizes a sliding window approach to minimize redundancy across different views of the same document. Models such as DocOwl 1.5 [14], LLaVA-Next [29], and InternVL [6] have explored the utility of OCR pretraining and dynamic partition techniques to address various aspect ratios. LLaVA-Read [63] integrates bi-resolution images and additional OCR encoders to improve textual information encoding. Other approaches, including DocPedia [9] and CogAgent [13], employ diverse encoding strategies to enhance visual representations. Despite these progress, challenges persist in processing complex document formats, particularly in interpreting their layout information. Additionally, existing document instruction datasets are often limited to questions and
+
+short answers, which can lead to hallucinations and restrict deeper contextual understanding. Since visual documents are typically high-structured data tied to their source codes, we develop contextually-grounded training data through markup language parsing and reasoning tasks. By training on structured document representations, our method aims to leverage the code interpretation capabilities of LLMs to further enhance visual document understanding.
+
+# 3. Method
+
+# 3.1. DocMark-Pile: Multi-Task Pretraining for Markup Language Conversion
+
+Visual document understanding encompasses various tasks, such as comprehending paragraphs from PDF documents, interpreting numerical data in charts and tables, and retrieving information from receipts, among others. Most documents are generated from source code or can be converted into structured representations. These structured formats act as markup languages, which define the structure of various elements, helping to convey meaning and organize the content effectively. To enhance the model's ability to understand and interpret a wide range of visual document types, we propose a multi-task pretraining dataset, DocMark-Pile, which focuses on translating documents into their corresponding markup languages. To ensure comprehensive coverage of document understanding scenarios, we utilize six distinct markup languages: plain text, Markdown, LaTeX, HTML, JSON, and TiKZ, as demonstrated in Fig. 2. Each format serves a unique purpose, which we will elaborate on below. For data acquisition, we utilize publicly available datasets as well as synthetic methods while maintaining the consistency between images and markup languages.
+
+# 3.1.1 Specialized Task Design
+
+Document to TXT Plain Text is the simplest format for basic content extraction from documents, as it retains all the textual information contained in the image. We primarily adopt this format for natural photos and regional text images, as these types of images typically contain straightforward textual information that can be extracted without formatting complexities. For natural photos, we utilize public datasets including TextOCR [47], COCO-Text [52], HierText [33], LSVT [49], RCTW-17 [45], and CTW1500 [62]. These datasets cover a wide range of scenarios, including advertising posters, signage, product labels, and image subtitles. Additionally, they often include bounding box annotations for the target text. We leverage these annotations to formulate text grounding tasks that require the model to recognize both the text and its coordinates within the image. For regional text images, we further incorporate datasets such as IC13 [17], IC15 [18], Union14ML [15], and Wor
+
+![](images/5cd7bd3ae78c5396973ee0a7cf92235501664721a3af23003bca29a8a4665bd5.jpg)  
+Figure 2. Overview of our DocMark-Pile dataset for pretraining on various markup language generation tasks. Due to the length of the answer, we have omitted a part of the content for better display.
+
+![](images/25b51b649d847daa48eee4f3836711f85cb154fe644ed1fd3a48b841bf45cd93.jpg)  
+Figure 3. Domain distribution of our DocMark-Pile dataset. (a) Distributions of different markup language parsing tasks. (b) Distributions of different image types.
+
+![](images/58007835f31a6626bc5e7c22f3d3b545a412fb777b7f6c9a1ff66544c9180aa5.jpg)
+
+dArt [55]. These images consist exclusively of regions containing text, which are usually in the form of single lines or curved text. We standardize the training data format by reading the text in the order from left to right and top to bottom, including additional bounding box coordinates for the text grounding task.
+
+Document to Markdown Markup is a lightweight syntax which enables basic formatting options such as bold, italics, headings, and lists. Markdown can be easily converted to and from other formats like HTML and LaTeX. We employ the Markdown format for dense text documents and tables because it effectively represents their structured content. For dense text documents, we primarily source our data from arXiv, where we obtain articles along with their corresponding source code and compiled PDFs. Following the process established in [13], we convert the PDFs into images while simultaneously translating the LaTeX code into Markdown format. Since many articles include complex formulas that cannot be successfully compiled in Markdown, we filter these instances during the conversion process. For tables, we draw from publicly available datasets, including PubTabNet [64], SynthTabNet [38], TabRecSet [58], TableGraph [57], and SciTSR [8]. We select these datasets because their annotations contain the original table content. Since many of these annotations are
+
+provided in HTML format, we convert them into Markdown format to meet the requirements of our task.
+
+Document to LaTeX LaTeX allows for the precise modeling of academic papers, particularly due to its capability to typeset complex mathematical formulas and scientific notation. We gather images converted from mathematics and science textbooks and exam papers, since these images typically contain numerous formulas and equations that must be accurately represented in LaTeX format. Subsequently, we utilize Mathpix to generate annotations for these documents. In addition, we incorporate handwritten mathematical expression recognition datasets, including HME100K [61] and MathWriting [11], which convert handwritten mathematical formulas into LaTeX format. We further enrich these data sources with additional rendered images. We first extract original mathematical expressions from our collected corpora, then use the Python library Matplotlib to render them as formula images. Subsequently, we apply data augmentation techniques, such as affine transformations and the addition of Gaussian noise, to enhance the robustness of the model in recognizing images under various backgrounds and display conditions.
+
+Document to HTML For webpage understanding, we introduce the task of parsing webpages into HTML formats. We utilize WebSight [22], a synthetic dataset consisting of HTML codes and their corresponding screenshots. Additionally, we augment our dataset by scraping real website images and HTML codes. Since many real website screenshots contain sparse content, directly learning from their HTML code often fails to extract useful information due to the noise. To address this, we introduce a webpage summarization task, which focuses on identifying and interpreting the key points and core content within a webpage. Specifically, we extract only the main elements of the page, such as the title, content, and other essential information. By focusing on the summary of key elements, this approach reduces the complexity and noise inherent in real website data, enhancing the quality of the pretraining data.
+
+Document to JSON To facilitate structured information extraction, we utilize JSON (JavaScript Object Notation) format, which represents information as key-value pairs, enhancing both readability and conciseness. This extraction is particularly valuable in scenarios involving cards, receipts, and forms, where summarizing and retrieving important data such as names, dates, and addresses from images is critical. To achieve this, we leverage several public datasets, including CORD [42], POIE [21], SVRD [60], WildReceipt [48], and XFUND [56], whose annotations can be easily converted into JSON format. In addition to textual information, we incorporate structured data from charts. We represent this information in a Python dictionary that encompasses fields such as "title", "source", "x-axis", "y-axis", and "value". For chart-related data, we utilize the PlotQA [37] dataset, which contains meta-information about chart content. Moreover, we augment our dataset by rendering various types of chart images. This involves randomly creating textual labels and numerical values. Then we employ Matplotlib to produce diverse chart images, including bar plots, scatter plots, line plots, and dot plots, using randomly sampled fonts and colors.
+
+Document to TikZ TikZ is a markup language commonly used for creating scientific figures. To construct our dataset, we leverage the open-source training sets of DaTikZv1 [1] and DaTikZv2 [2], which contain TikZ graphics collected from arXiv papers and artificial examples, along with their corresponding TikZ programs. Additionally, we extract scientific drawings from a curated corpus of mathematical and scientific textbooks and employ a trained model from [2] to convert these illustrations into TikZ code. Recognizing that conversion may not always be accurate, we take the necessary step of recompiling the TikZ code and use the re-generated images in our training. We also exclude any instances that fail to compile, ensuring that the TikZ code remains consistent with the figures.
+
+# 3.1.2 Dataset Statistics
+
+DocMark-Pile, a comprehensive dataset for document understanding, comprises 3.8 million examples of diverse markup language parsing tasks and image types.
+
+**Markup Language Categories:** We visualize the distribution of different language parsing tasks in Fig. 2, showcasing the percentage of each task type. This highlights the diversity of our dataset. In our data annotation, we prepend the special tokens such as  $\langle \mathrm{md} \rangle$  and  $\langle \mathrm{md} \rangle$  or  $\langle \mathrm{json} \rangle$  and  $\langle \mathrm{json} \rangle$  to the beginning and end of answers to indicate the corresponding markup language category.
+
+Image Types: Our dataset covers a wide range of image types, encompassing a comprehensive representation of visual documents. For clarity, we categorize text-dense materials like academic papers, textbooks, exam papers, and
+
+receipts under "documents", which constitutes the largest portion of the dataset.
+
+# 3.2. Adaptive Markup Language Generation During Fine-tuning
+
+# 3.2.1 Adaptive Generation Pipeline
+
+Through multi-task pretraining for versatile markup language generation, our model gains enhanced abilities to understand both visual content and layout information in diverse documents. We further fine-tune the model using instruction-tuning datasets, equipping it with the skills to handle downstream tasks such as open-ended question-answering related to document contexts. Previous methods primarily utilize questions and brief answers from public benchmark training sets, limiting their ability for deeper comprehension. In contrast, our model's ability to generate structured markup languages offers a significant advantage. These generated markup languages can serve as auxiliary context cues that facilitate more complex reasoning tasks. Being highly concise and structured, markup languages encode textual information efficiently and can be processed effectively by language models pre-trained on extensive text and code corpora.
+
+Motivated by chain-of-thought reasoning widely employed in the domain of LLMs, we design an adaptive markup language generation pipeline to enhance reasoning abilities. Rather than manually determining the markup language type to generate by the pretrained model as a prefix, we let our model automatically generate the appropriate markup language as an intermediate step to derive answers. This mirrors the human-like chain-of-thought reasoning; when reading documents such as PDFs or textbook pages, humans first scan the content, then localize and retrieve relevant information to answer specific questions.
+
+In our adaptive pipeline, the model automatically determines the appropriate markup language type and identifies contextually relevant information to answer the question. We implement this as a two-round conversational process as follows:
+
+$$
+Q _ {1}: \text {T o a n s w e r t h e q u e s t i o n :} <   Q u e s t i o n >, \quad \text {e x t r a c t t h e r e l e v a n t c o m p e x t f r o m t h e i m a g e}.
+$$
+
+$$
+A _ {1}: <   t y p e > <   C o n t e x t > <   / t y p e >
+$$
+
+$$
+\begin{array}{l} Q _ {2}: \text {B a s e d o n t h e i m a g e a n d e x t r a c t e d c o n t e x t ,} \\ \text {a n s w e r t h e q u e s t i o n : <   Q u e s t i o n >} \end{array} \tag {1}
+$$
+
+$$
+A _ {2}: <   A n s w e r >
+$$
+
+In this pipeline, the model is first asked to provide the essential contextual information to address the specific question  $<  \text{Question}>$ . It then outputs this context as  $<  \text{Context}>$ , serving as an intermediate rationale in the process. Special tokens  $<  \text{type}>$  and  $<  / \text{type}>$  indicate
+
+![](images/43aeef0bbed0874c420c9f991d786b6720b488991bca92a77dd3d7c94430c959.jpg)  
+Figure 4. Overview of our adaptive markup language generation framework. (a) During pretraining, the model learns to parse documents into corresponding markup language representations, enhancing its structural understanding capabilities. (b) In the finetuning phase, the model is optimized to identify and extract relevant information as intermediate rationales to formulate precise answers.
+
+the markup language type. In the second round, the model derives the final answer by utilizing both the image and the context provided in the first round. This structured approach ensures that the model systematically processes and retains critical contextual information, thereby enhancing its ability to generate accurate and contextually grounded answer akin to human thought processes.
+
+# 3.2.2 DocMark-Instruct: CoT Reasoning Data Construction
+
+To enable our model with adaptive generation capabilities, it is essential to curate a dataset formatted as described in template (1). However, existing fine-tuning datasets only contain question-answer pairs and lack the detailed CoT reasoning process. To address this gap, we propose an efficient method for constructing the required data, DocMark-Instruct. Most questions in common training datasets within the document domain are closely related to the textual information present in the images. We observe that these questions can often be answered directly by providing the corresponding markup language to a powerful language model, thus eliminating the need for visual input. Therefore, we first utilize our pretrained model to convert images into their corresponding markup languages. Next, we employ ChatGPT-3.5, a freely accessible large language model known for its robust language understanding and generation capabilities, to create additional annotations. We input the question-answer pairs and the entire markup language text into it, asking it to retrieve the necessary information to answer the questions, finally producing 624k annotations. Our data construction pipeline circumvents the high annotation
+
+costs typically associated with proprietary MLLMs such as GPT4-V, making it a practical and accurate solution for developing comprehensive CoT reasoning datasets.
+
+# 3.2.3 Training and Inference
+
+During the fine-tuning phase, we concatenate the two-round dialogue data for the forward process and apply Cross-Entropy loss on both answers,  $A_{1}$  and  $A_{2}$ . In the inference phase, we first input  $Q_{1}$  to prompt the model to generate the adaptive markup language context. Following this, we formulate  $Q_{2}$  to obtain the final response. Importantly, our training pipeline is also compatible with common training data that lacks the CoT reasoning process. Users can directly ask the model to answer a question by omitting the CoT prompt described in template (1).
+
+# 4. Experiments
+
+# 4.1. Implementation Details
+
+Model Architecture We adopt the InternVL [7] architecture for our model due to its enhanced visual representation capabilities. We use the pretrained InternViT [7] as the visual encoder and utilize InternLM-2B [50] and InternLM-8B [50], known for their robust linguistic understanding abilities. We adopt MLP layers to project the image into visual tokens, which are further concatenated with text tokens. As our work focuses on incorporating adaptive markup language generation to improve visual document understanding, we do not explore additional resolution scaling strategies in this paper. Instead, we employ the dynamic resolution method proposed in InternVL2 [6], which demonstrates efficacy in modeling high-resolution
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2" colspan="2">Text Recognition</td><td rowspan="2">HME100K</td><td colspan="3">ChartQA-SE</td><td colspan="2">KIE</td><td rowspan="2">DaTikZ-test</td></tr><tr><td>AP@strict</td><td>AP@slight</td><td>AP@high</td><td>FUNSD</td><td>SROIE</td></tr><tr><td>Paddle OCR</td><td>140</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CAN-ABM [23]</td><td></td><td></td><td>64.3</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>OneChart [4]</td><td></td><td></td><td></td><td>72.0</td><td>82.9</td><td>85.9</td><td></td><td></td><td></td></tr><tr><td>Nougat [3]</td><td></td><td></td><td></td><td></td><td></td><td></td><td>55.4</td><td>33.6</td><td></td></tr><tr><td>DT-DS1.3B [2]</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.22</td></tr><tr><td>DT-CL7B [2]</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td>0.68</td></tr><tr><td>DocMark-2B</td><td>247</td><td>67.2</td><td>73.8</td><td>85.2</td><td>86.5</td><td>64.1</td><td>54.2</td><td>0.39</td><td></td></tr><tr><td>DocMark-8B</td><td>252</td><td>73.4</td><td>75.4</td><td>85.5</td><td>87.1</td><td>68.5</td><td>59.6</td><td>0.64</td><td></td></tr></table>
+
+Table 1. Performance comparison with expert models on specialized tasks. The best results are marked in bold.
+
+document images. Specifically, the input image is split into several sub-images of resolution  $448 \times 448$ , with a maximum of 12 sub-images used in our training settings.
+
+Training Settings Our model is trained through a two-stage process. In the first stage, we utilize DocMark-Pile to empower the model with markup language parsing abilities. Additionally, we incorporate the image captioning dataset LLaVA-558K [28] to ensure alignment between the visual input and text tokens. We pre-train the model for one epoch with a learning rate of 2e-5 and a total batch size of 128. In the second stage, we fine-tune our model on the proposed CoT Reasoning dataset DocMark-Instruct. As described in Sec. 3.2.2, we create additional CoT annotations for the training sets of TextVQA [46], DocVQA [36], InfographicsVQA [35], ChartQA [34], DVQA [16], XFUND [56], and Geo170K [10]. Apart from DocMark-Instruct, we also integrate general instruction tuning data from ShareGPT4V [5] and text-only multi-turn conversation datasets [20, 65]. This integration preserves the model's capacity for general understanding and conversational interaction. We fine-tune our model on these datasets for one epoch with a learning rate of 1e-5 and a total batch size of 128.
+
+# 4.2. Evaluation on Markup Language Generation Abilities
+
+Evaluation Settings The ability to generate accurate markup languages is crucial for our model to understand complex document formats and provide adaptive context for facilitating downstream task reasoning. We comprehensively evaluate our model on specialized tasks and compare it with domain-expert models. We utilize the text recognition benchmark in [31], which covers a broad range of scene texts. We also assess our model's accuracy in expressing formulas present in images using the LaTeX format, utilizing the test set of HME100K [61]. Moreover, we adopt the ChartQA-SE benchmark proposed in [4] and report AP@strict, AP@slight, and AP@high metrics. We also evaluate on FUNSD and SROIE for key information extraction task. These metrics measure the accuracy of the
+
+structural extraction results compared to the ground-truth values. For TiKZ code generation, we use the DaTikZ test set proposed in [2]. We report the average of image similarity and code similarity to assess the code consistency. By thoroughly evaluating our model across these diverse tasks, we aim to demonstrate its capability to accurately generate markup languages, thereby facilitating effective visual document understanding and reasoning.
+
+Results. We present the comparison results of our model against various expert models in Tab. 1. As can be seen, our DocMark models outperform the expert models by a significant margin on the text recognition and HME100K [61] benchmarks. Moreover, our model demonstrates impressive capabilities in structural information extraction, particularly on the ChartQA-SE [4] and KIE tasks. On DaTiKZ-test [2], DocMark-8B achieves performance nearly on par with DTCL $_{7B}$ , without additional search costs for TiKZ code generation. These results showcase the model's effectiveness in markup language conversion, enhancing visual perception and supporting contextually grounded reasoning.
+
+# 4.3. Evaluation on Downstream Understanding tasks
+
+Evaluation Settings To validate the effectiveness of adaptive markup language generation, we evaluate the model's performance on a range of downstream tasks for document understanding. We adopt diverse evaluation benchmarks that encompass a wide variety of multi-modal scenarios. We report accuracy on TextVQA [46] to assess the model's ability to answer questions related to natural scene text. Additionally, we evaluate performance on DocVQA [36], InfoVQA [35], AI2D [19], and ChartQA [34] for comprehension of text-dense documents and charts. To assess the model's general OCR abilities, we utilize OCRBench [31], a comprehensive benchmark containing various text-related images. For webpage understanding, we report the WebQA accuracy as proposed in VisualWebBench [30]. Furthermore, we evaluate the model's math reasoning performance using MathVision [53] dataset to validate its ability to interpret complex mathematical diagrams.
+
+<table><tr><td>Method</td><td>LLM</td><td>TextVQVal</td><td>DocVQVal</td><td>InfoVQVal</td><td>ChartQA</td><td>AI2D</td><td>OCRBench</td><td>WebQA</td><td>MathVision</td></tr><tr><td colspan="10">Lightweight Specialist Models</td></tr><tr><td>Donut</td><td>-</td><td>43.5</td><td>67.5</td><td>11.6</td><td>41.8</td><td>30.8</td><td>-</td><td>-</td><td></td></tr><tr><td>Pix2Struct-Base</td><td>-</td><td>-</td><td>72.1</td><td>38.2</td><td>56.0</td><td>40.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Pix2Struct-Large</td><td>-</td><td>-</td><td>76.6</td><td>40.0</td><td>58.6</td><td>42.1</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan="10">Multimodal LMMs</td></tr><tr><td>Vary-toy</td><td>1.8B</td><td>-</td><td>65.6</td><td>-</td><td>59.1</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MiniCPM-V-2</td><td>2.8B</td><td>74.1</td><td>69.7</td><td>37.9</td><td>59.8</td><td>62.6</td><td>594</td><td>-</td><td>15.1</td></tr><tr><td>InternVL2</td><td>2B</td><td>73.2</td><td>86.0</td><td>57.6</td><td>76.4</td><td>74.2</td><td>783</td><td>66.5</td><td>15.7</td></tr><tr><td>DocMark (Ours)</td><td>2B</td><td>74.8</td><td>87.8</td><td>61.2</td><td>79.8</td><td>82.5</td><td>813</td><td>70.1</td><td>18.8</td></tr><tr><td>Ureader</td><td>7B</td><td>57.6</td><td>80.7</td><td>46.4</td><td>70.0</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DocPedia</td><td>7B</td><td>60.2</td><td>47.1</td><td>15.2</td><td>46.9</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Vary</td><td>7B</td><td>-</td><td>76.3</td><td>-</td><td>66.1</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Monkey</td><td>9B</td><td>67.6</td><td>66.5</td><td>36.1</td><td>65.1</td><td>62.6</td><td>514</td><td>-</td><td>-</td></tr><tr><td>TextMonkey</td><td>9B</td><td>64.3</td><td>66.7</td><td>28.6</td><td>59.9</td><td>-</td><td>558</td><td>-</td><td>-</td></tr><tr><td>CogAgent</td><td>17B</td><td>76.1</td><td>81.6</td><td>44.5</td><td>68.4</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DocOwl2</td><td>8B</td><td>66.7</td><td>80.7</td><td>46.4</td><td>70.0</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>InternVL2</td><td>8B</td><td>77.4</td><td>90.8</td><td>72.8</td><td>82.2</td><td>83.7</td><td>794</td><td>77.5</td><td>20.2</td></tr><tr><td>DocMark (Ours)</td><td>8B</td><td>78.0</td><td>89.8</td><td>68.3</td><td>84.2</td><td>86.2</td><td>823</td><td>78.9</td><td>21.1</td></tr></table>
+
+Table 2. Performance comparisons on downstream document understanding tasks. The best results are marked in bold.
+
+![](images/159ed93e65e209d0d27be2d91a8495032c8b46ee27698a73dc7513c38c384c62.jpg)  
+Figure 5. Ablation study on employing different training dataset strategies.
+
+Results We present the comparison results on downstream understanding tasks in Tab. 2. Our DocMark models consistently outperform state-of-the-art multimodal large language models across nearly all benchmarks, both for the 2B and 7B parameter sizes. Notably, DocMark significantly enhances performance on common document understanding tasks, while also achieving remarkable results on MathVision, demonstrating its effectiveness in addressing complex multimodal math problems. Furthermore, our models show significant improvements in general OCR and webpage understanding, as highlighted by the results on OCRBench and WebQA. When compared to state-of-the-art models like the InternVL2 series, DocMark achieves superior accuracy with substantially less training data. These results illustrate that the proposed framework greatly enhances document understanding and reasoning capabilities.
+
+# 4.4. Ablation Study
+
+To investigate the impact of proposed datasets, we examine various training dataset strategies using DocMark-2B. We train the model without incorporating DocMark-Pile and DocMark-Instruct, relying solely on image-caption datasets and general question-answering datasets for fine-tuning. We report the accuracy of the model on four document-related QA datasets for analysis. Subsequently, we integrate DocMark-Pile for model pretraining and employ DocMark-Instruct for fine-tuning. As illustrated in Fig. 5, the absence of DocMark-Pile significantly degrades model performance, particularly for text-dense documents and structured data. Furthermore, incorporating DocMark-Instruct substantially enhances performance when built upon DocMark-Pile pretraining, demonstrating its effectiveness in improving reasoning capabilities. We provide more performance analysis in our supplementary.
+
+# 5. Conclusion
+
+In this paper, we have introduced an innovative framework that utilizes adaptive markup language generation to enhance visual document understanding. Our approach addresses the core challenge of efficiently representing and perceiving visual content and layout information in images by incorporating specialized pretraining tasks and an adaptive generation pipeline. The proposed method not only improves the versatility of MLLMs in handling complex document formats, but also enables them to perform chain-of-thought reasoning to produce contextually grounded answers. Extensive experiments demonstrated that our models outperform existing state-of-the-art approaches on several challenging visual document understanding tasks.
+
+# Acknowledgments
+
+This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24, by the National Natural Science Foundation of China (No.62206272). Hongsheng Li is a PI of CPII under the InnoHK.
+
+# References
+
+[1] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In The Twelfth International Conference on Learning Representations, 2024. 5  
+[2] Jonas Belouadi, Simone Paolo Ponzetto, and Steffen Eger. DeTikZify: Synthesizing graphics programs for scientific figures and sketches with TikZ. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 5, 7  
+[3] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. 7  
+[4] Jinyue Chen, Lingyu Kong, Haoran Wei, Chenglong Liu, Zheng Ge, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Onechart: Purify the chart structural extraction via one auxiliary token. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 147-155, 2024. 7  
+[5] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 7  
+[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 1, 3, 6  
+[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 6  
+[8] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729, 2019. 4  
+[9] Hao Feng, Qi Liu, Hao Liu, Wengang Zhou, Houqiang Li, and Can Huang. Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding. arXiv preprint arXiv:2311.11810, 2023. 3  
+[10] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu,
+
+Zhenguo Li, and Lingpeng Kong. G-llava: Solving geometric problem with multi-modal large language model, 2023. 7  
+[11] Philippe Gervais, Asya Fadeeva, and Andrii Maksai. Math-writing: A dataset for handwritten mathematical expression recognition. arXiv preprint arXiv:2404.10690, 2024. 4  
+[12] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 3  
+[13] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1, 3, 4  
+[14] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3  
+[15] Qing Jiang, Jiapeng Wang, Dezhi Peng, Chongyu Liu, and Lianwen Jin. Revisiting scene text recognition: A data perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20543-20554, 2023. 3  
+[16] Kushal Kafle, Scott Cohen, Brian Price, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In CVPR, 2018. 7  
+[17] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazan Almazan, and Lluis Pere De Las Heras. Icdar 2013 robust reading competition. In 2013 12th international conference on document analysis and recognition, pages 1484-1493. IEEE, 2013. 3  
+[18] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 3  
+[19] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 7  
+[20] Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stanley, Richard Nagyfi, et al. Openassistant conversations-democratizing large language model alignment. Advances in Neural Information Processing Systems, 36, 2024. 7  
+[21] Jianfeng Kuang, Wei Hua, Dingkang Liang, Mingkun Yang, Deqiang Jiang, Bo Ren, and Xiang Bai. Visual information extraction in the wild: practical dataset and end-to-end solu
+
+tion. In International Conference on Document Analysis and Recognition, pages 36-53. Springer, 2023. 5  
+[22] Hugo Laurençon, Léo Tronchon, and Victor Sanh. Unlocking the conversion of web screenshots into html code with the websight dataset. arXiv preprint arXiv:2403.09029, 2024. 4  
+[23] Bohan Li, Ye Yuan, Dingkang Liang, Xiao Liu, Zhilong Ji, Jinfeng Bai, Wenyu Liu, and Xiang Bai. When counting meets hmer: counting-aware network for handwritten mathematical expression recognition. In European conference on computer vision, pages 197-214. Springer, 2022. 7  
+[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3  
+[25] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 1, 3  
+[26] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 1, 3  
+[27] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 3  
+[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 7  
+[29] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3  
+[30] Junpeng Liu, Yifan Song, Bill Yuchen Lin, Wai Lam, Graham Neubig, Yuanzhi Li, and Xiang Yue. Visualwebbench: How far have multimodal llms evolved in web page understanding and grounding?, 2024. 7  
+[31] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 7  
+[32] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 1  
+[33] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1049-1059, 2022. 3  
+[34] Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL
+
+2022, pages 2263-2279, Dublin, Ireland, 2022. Association for Computational Linguistics. 7  
+[35] Minesh Mathew, Viraj Bagal, Ruben Pérez Tito, Dimosthenis Karatzas, Ernest Valveny, and C. V Jawahar. Infographicvqa, 2021. 7  
+[36] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 7  
+[37] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 5  
+[38] Ahmed Nassar, Nikolaos Livathinos, Maksym Lysak, and Peter Staar. Tableformer: Table structure understanding with transformers. arXiv preprint arXiv:2203.01017, 2022. 4  
+[39] OpenAI. Chatgpt. https://chat.openai.com, 2023.3  
+[40] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023.3  
+[41] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 3  
+[42] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. Cord: a consolidated receipt dataset for post-ocr parsing. In Workshop on Document Intelligence at NeurIPS 2019, 2019. 5  
+[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3  
+[44] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv e-prints, pages arXiv-2403, 2024. 3  
+[45] Baoguang Shi, Cong Yao, Minghui Liao, Mingkun Yang, Pei Xu, Linyan Cui, Serge Belongie, Shijian Lu, and Xiang Bai. Icdar2017 competition on reading chinese text in the wild (rctw-17). In 2017 14th iapr international conference on document analysis and recognition (ICDAR), pages 1429-1434. IEEE, 2017. 3  
+[46] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 7  
+[47] Amanpreet Singh, Guan Pang, Mandy Toh, Jing Huang, Wojciech Galuba, and Tal Hassner. Textocr: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8802-8812, 2021. 3  
+[48] Hongbin Sun, Zhanghui Kuang, Xiaoyu Yue, Chenhao Lin, and Wayne Zhang. Spatial dual-modality graph reasoning for key information extraction. arXiv preprint arXiv:2103.14470, 2021.5
+
+[49] Yipeng Sun, Zihan Ni, Chee-Kheng Chng, Yuliang Liu, Canjie Luo, Chun Chet Ng, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, et al. Icdar 2019 competition on large-scale street view text with partial labeling-rrc-Isvt. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1557–1562. IEEE, 2019. 3  
+[50] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023. 6  
+[51] 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. 3  
+[52] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 3  
+[53] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset, 2024. 7  
+[54] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language model. In European Conference on Computer Vision, pages 408-424. Springer, 2025. 1  
+[55] Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, and Xiang Bai. Toward understanding wordart: Corner-guided transformer for scene text recognition. In European conference on computer vision, pages 303-321. Springer, 2022. 4  
+[56] Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, and Furu Wei. Xfund: a benchmark dataset for multilingual visually rich form understanding. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3214-3224, 2022. 5, 7  
+[57] Wenyuan Xue, Baosheng Yu, Wen Wang, Dacheng Tao, and Qingyong Li. Tgrnet: A table graph reconstruction network for table structure recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1295-1304, 2021. 4  
+[58] Fan Yang, Lei Hu, Xinwu Liu, Shuangping Huang, and Zhenghui Gu. A large-scale dataset for end-to-end table recognition in the wild. Scientific Data, 10(1):110, 2023. 4  
+[59] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3  
+[60] Wenwen Yu, Chengquan Zhang, Haoyu Cao, Wei Hua, Bohan Li, Huang Chen, Mingyu Liu, Mingrui Chen, Jianfeng Kuang, Mengjun Cheng, et al. Icdar 2023 competition on structured text extraction from visually-rich document images. In International Conference on Document Analysis and Recognition, pages 536-552. Springer, 2023. 5  
+[61] Ye Yuan, Xiao Liu, Wondimu Dikubab, Hui Liu, Zhilong Ji, Zhongqin Wu, and Xiang Bai. Syntax-aware network for handwritten mathematical expression recognition. In Pro-
+
+ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4553-4562, 2022. 4, 7  
+[62] Liu Yuliang, Jin Lianwen, Zhang Shuaiqiao, and Zhang Sheng. Detecting curve text in the wild: New dataset and new solution. arXiv preprint arXiv:1712.02170, 2017. 3  
+[63] Ruiyi Zhang, Yufan Zhou, Jian Chen, Jiuming Gu, Changyou Chen, and Tong Sun. Llava-read: Enhancing reading ability of multimodal language models. arXiv preprint arXiv:2407.19185, 2024. 3  
+[64] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In European conference on computer vision, pages 564-580. Springer, 2020. 4  
+[65] Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36, 2024. 7

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc7ddf07b976ff789237ff0b2fafcc1a1e996c715ac3aafd81f7c8fa355ec2fe
+size 419896

CVPR/2025/Adaptive Markup Language Generation for Contextually-Grounded Visual Document Understanding/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44204779d7a6addb184f8088cfbc0151e30f8c46da832ad77d70d046bc8ddeca
+size 313153

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/22adf80a-55fe-4236-9782-243fa367cc68_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:364d76119922a44f9dcbdcb7ace1549f923053548f263fed5a433dfedd7fb32d
+size 74417

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/22adf80a-55fe-4236-9782-243fa367cc68_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac7734140e99a29d4a4c509f4fc9c688625f3337646cad88dd4eef90ae2679f2
+size 89021

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/22adf80a-55fe-4236-9782-243fa367cc68_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d95ea1ff7d4a9787a086fee560dccfac8a8e4c48363e3099d84322ef24e3daad
+size 2264874

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/full.md

@@ -0,0 +1,346 @@
+# Adaptive Non-uniform Timestep Sampling for Accelerating Diffusion Model Training
+
+Myunsoo Kim $^{1*}$  Donghyeon Ki $^{1*}$  Seong-Woong Shim $^{1}$  Byung-Jun Lee $^{1,2}$ $^{1}$ Korea University  $^{2}$ Gauss Labs Inc.
+
+{m970326, peopleln, ssw030830, byungjunlee}@korea.ac.kr
+
+# Abstract
+
+As a highly expressive generative model, diffusion models have demonstrated exceptional success across various domains, including image generation, natural language processing, and combinatorial optimization. However, as data distributions grow more complex, training these models to convergence becomes increasingly computationally intensive. While diffusion models are typically trained using uniform timestep sampling, our research shows that the variance in stochastic gradients varies significantly across timesteps, with high-variance timesteps becoming bottlenecks that hinder faster convergence. To address this issue, we introduce a non-uniform timestep sampling method that prioritizes these more critical timesteps. Our method tracks the impact of gradient updates on the objective for each timestep, adaptively selecting those most likely to minimize the objective effectively. Experimental results demonstrate that this approach not only accelerates the training process, but also leads to improved performance at convergence. Furthermore, our method shows robust performance across various datasets, scheduling strategies, and diffusion architectures, outperforming previously proposed timestep sampling and weighting heuristics that lack this degree of robustness.
+
+# 1. Introduction
+
+In recent years, diffusion models have demonstrated their high expressive power and achieved significant success across a wide range of domains, including image [9, 23, 24], video [18, 35], text [4, 12, 22], audio generation [26], combinatorial optimization [25, 30], reinforcement learning [1, 15, 32], and more. However, with the increasing complexity of target data distributions, achieving convergence in diffusion model training has become increasingly computationally demanding. For example, training images with Stable-Diffusion-2.0 [24] requires 24,000 A100 GPU hours, while Open-Sora [35] requires 48,000 H800 GPU hours.
+
+![](images/6c2a13a96a5c9fe4ad8b858487429a32a7bf816f8a718676e1974069a759a755.jpg)  
+Figure 1. Comparison of FID scores of our approach and other acceleration methods against relative wall clock time, where 1D represents the time it takes for the baseline to converge. Although our learning method is initially slower than the heuristics due to its learning-based nature, it converges to a point with better optimality within 1D, and achieves a significantly lower FID score by 1.2D.
+
+hours for video training. These high computational costs not only pose significant barriers to advancing generative AI applications but also have negative environmental consequences.
+
+While numerous studies have explored ways to accelerate diffusion model training, our research specifically targets modifications to the uniform training process of diffusion models across diffusion timesteps. It is well recognized that non-uniform training schemes over timesteps can enhance training speed. For example, several heuristics have been proposed [7, 11, 17, 31].
+
+However, the underlying reasons for the success of these methods remain largely unexplored, and they are often regarded as heuristics that may have potential limitations in robustness across diverse experimental conditions.
+
+Through a series of experiments, we observed significant variation in the stochastic gradient variance across different diffusion training timesteps.
+
+We hypothesize that this variability plays a key role in the imbalanced training observed with uniform sampling
+
+and may explain the effectiveness of previously proposed heuristics in improving diffusion model training. However, due to the strong interdependence of gradients across different timesteps, we found that this insight alone does not directly lead to optimal acceleration in diffusion training.
+
+To this end, we propose a more direct approach by developing an efficient algorithm designed to sample timesteps that are estimated to yield the greatest reduction in the objective. This algorithm approximates the impact of gradient updates for each timestep on the variational lower bound and increases the sampling frequency of specific timesteps that require further optimization.
+
+This enables the diffusion training process to converge significantly faster than previous heuristics across various settings. To summarize, our three main contributions are as follows:
+
+1. Through an analysis of gradient variance in diffusion model training, we offer an explanation for why a nonuniform training process across timesteps can lead to faster convergence.  
+2. Unlike previous heuristic-based diffusion acceleration methods, we propose a learning-based approach that adaptively samples timesteps by approximately minimizing the variational bound on the negative log-likelihood.  
+3. We conduct experiments across various image datasets, noise scheduling strategies and diffusion architectures to demonstrate the robustness of our method.
+
+# 2. Related Works
+
+Non-uniform training of diffusion models Several studies have explored timestep-wise non-uniform training of diffusion models to accelerate training. These approaches can be categorized into two main groups:
+
+- 1) Weighting methods: These methods apply hand-designed weightings to the loss at each timestep. For example, P2 [7] assigns greater weights to timesteps containing perceptually rich content and lesser weights to those involving imperceptible details. Similarly, MinSNR [11] adjusts loss weights based on clamped signal-to-noise ratios.  
+- 2) Sampling methods: These approaches use hand-designed samplers to select timesteps in a non-uniform manner. For instance, Log-Normal [17] samples timesteps from a log-normal distribution. SpeeD [31] modifies sampling probabilities to downweight timesteps with lower losses and prioritize those with rapidly changing process increments.
+
+While these heuristics have proven effective in accelerating diffusion model training, their direct relationship to the underlying objective remains unclear, raising concerns about their applicability in new, potentially more effective settings that may emerge in the future.
+
+Learning to optimize (L2O) L2O is a field in machine learning focused on learning optimization algorithm. It typically involves a meta-training set composed of multiple training and validation datasets. Using this meta-training set, L2O learns to update parameters based on gradients to minimize validation set errors, either through supervised learning [3, 33] or reinforcement learning [6, 20]. Although our objective similarly focuses on learning to optimize, we address a more challenging task by not assuming metatraining set for transfer learning. Instead, we focus on learning to optimize in an online manner, without the benefit of prior training data.
+
+# 3. Background: Diffusion Models
+
+Diffusion model [14, 27] consists of two key processes: a predefined forward noising process and a reverse denoising process, which is learned through variational inference to match the forward process.
+
+Forward process Starting with data  $x_0 \sim q(x_0)$ , the forward process operates as a Markov-Gaussian process that incrementally introduces noise, generating a sequence of noisy latent variables  $\{x_1, x_2, \ldots, x_T\}$ :
+
+$$
+q \left(x _ {t} \mid x _ {t - 1}\right) := \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I\right), \tag {1}
+$$
+
+where  $\beta_{t}$  is a noise schedule. Latent variables can be equivalently sampled by  $x_{t} \coloneqq \sqrt{\bar{\alpha}_{t}} x_{0} + \sqrt{1 - \bar{\alpha}_{t}} \epsilon$ ,  $\epsilon \sim \mathcal{N}(0, I)$ , where  $\bar{\alpha}_{t} \coloneqq \prod_{s=1}^{t} (1 - \alpha_{s})$  and  $\alpha_{t} \coloneqq 1 - \beta_{t}$ .
+
+Reverse process Diffusion models generate data  $x_0$  from  $x_{T}$  through the learned reverse process  $p_{\theta}(x_{t - 1}|x_t)$ , gradually denoising the latent variables from a Gaussian noise:
+
+$$
+p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) := \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \sigma_ {t} ^ {2} I\right), \tag {2}
+$$
+
+$$
+\mu_ {\theta} \left(x _ {t}, t\right) := \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(x _ {t}, t\right)\right) \tag {3}
+$$
+
+where  $\sigma_t^2$  is a variance of the denoising process, and  $\mu_{\theta}$  is parameterized as proposed by Ho et al. [14] for further simplification.  $\epsilon_{\theta}$  is a deep neural network with parameter  $\theta$  that predicts the noise vector  $\epsilon$  based on  $x_{t}$  and  $t$ .
+
+Training The training of a diffusion model is achieved by optimizing the variational lower bound on the log-likelihood, which is expressed as a sum of KL-divergence terms comparing Gaussian distributions and can be reformulated in terms of the normalized noise:
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {V L B}} (\theta) := \mathbb {E} _ {x _ {0}} \left[ \sum_ {t} D _ {\mathrm {K L}} \left(q \left(x _ {t - 1} \mid x _ {t}, x _ {0}\right) \| p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right)\right) \right] \\ = \mathbb {E} _ {x _ {0}, \epsilon , t} \left[ \frac {\beta_ {t} ^ {2}}{2 \sigma_ {t} ^ {2} \alpha_ {t} (1 - \bar {\alpha} _ {t})} \| \epsilon - \epsilon_ {\theta} (x _ {t}, t) \| ^ {2} \right], \tag {4} \\ \end{array}
+$$
+
+![](images/1e5ab18fe8599423585678385759bfefa30ad1d345da7caa10590fccb331d67f.jpg)
+
+![](images/10972fd68b16344fb1ec67d22ce41493a59457368b2b1740688594f15fa04e61.jpg)
+
+![](images/78334cd5ba00bc2bbaa2134e3086dd94f14fc77820d63606095032e128e7754d.jpg)
+
+![](images/bbdf93126d79f24d916b627f6c9011562423a1bd9b0d2e49c55461c0d64e7c88.jpg)  
+Figure 2. Comparison of gradient variance (top) and diffusion loss (bottom) over timesteps for three different noise schedules: linear, cosine, and quadratic. The color bar on the right indicates the progression of epochs, with red representing the early epochs and blue representing the later epochs. The results are based on training DDPM [14] with  $\mathcal{L}_{\mathrm{DDPM}}$  for 20 epochs.
+
+![](images/7e9cb76a40b26b1c20ed9bfb1e74f1150dd5dde874605bce0d26c276b69c89a7.jpg)
+
+![](images/6ad845b55cdef6e4f25acf3d9ed28e0b605848cb2cede07136409e7ba5db7787.jpg)
+
+where  $t \sim \mathcal{U}\{1, T\}$ . A standard way to train a diffusion model is to ignore the constant terms [14], resulting in an objective of:
+
+$$
+\mathcal {L} _ {\mathrm {D D P M}} (\theta) := \mathbb {E} _ {x _ {0}, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \| ^ {2} \right]. \tag {5}
+$$
+
+# 4. Demystifying Non-Uniform Diffusion Model Training Across Timesteps
+
+In this section, we offer a hypothesis explaining why nonuniform training strategies may offer advantages over the uniform training approach used by Ho et al. [14]. For simplicity, we begin by assuming an independent parameterization across timesteps, expressed as:
+
+$$
+\epsilon_ {\theta} \left(x _ {t}, t\right) = \epsilon_ {\theta_ {t}} \left(x _ {t}\right), \quad \forall t, \tag {6}
+$$
+
+with a independent set of parameters  $\theta = \{\theta_t\}_{t=1}^T$  for each timestep. This parameterization allows us to have independent optimization objective per timestep, i.e.,  $\mathcal{L}_{\mathrm{VLB}}(\theta) = \sum_{t=1}^{T} \mathbb{E}_{x_0, \epsilon}[\mathcal{L}_t(\theta)]$  with
+
+$$
+\mathcal {L} _ {t} (\theta) := c _ {t} \| \epsilon - \epsilon_ {\theta_ {t}} (x _ {t}) \| ^ {2}, \quad c _ {t} := \frac {\beta_ {t} ^ {2}}{2 \sigma_ {t} ^ {2} \alpha_ {t} (1 - \bar {\alpha} _ {t})}. \tag {7}
+$$
+
+We now have divided the problem into  $T$  independent subproblems, where the training with uniform timestep sampling corresponds to uniformly selecting one of these subproblems  $\mathcal{L}_t$  and performing a stochastic gradient step on its corresponding parameter  $\theta_t$ . Informally, if certain subproblems are particularly more challenging and require more iterations to optimize, there is a potential for improvement through non-uniform training strategies.
+
+In this sense, we assess the difficulty of these subproblems by measuring a variance of stochastic gradients,  $\operatorname{Var}_{x_0,\epsilon}[\nabla_{\theta_t}\mathcal{L}_t(\theta)]$ . It is well known that the convergence of stochastic gradient descent is only guaranteed with the learning rates that scale with the variance of the gradient:
+
+Theorem 1. [Garrigos and Gower [10], informal] Given a few reasonable assumptions, for every  $\varepsilon > 0$ , we can guarantee that  $\mathbb{E}\left[\mathcal{L}_t(\theta_K) - \inf \mathcal{L}_t\right] \leq \varepsilon$  provided that
+
+$$
+\gamma = \mathcal {O} \left(\frac {1}{\sqrt {K}} \left(\operatorname {V a r} _ {x _ {0}, \epsilon} \left[ \nabla_ {\theta_ {t}} \mathcal {L} _ {t} \left(\theta_ {*}\right) \right]\right) ^ {- \frac {1}{2}}\right), \tag {8}
+$$
+
+$$
+K \geq \mathcal {O} \left(\frac {1}{\varepsilon^ {2}} \operatorname {V a r} _ {x _ {0}, \epsilon} \left[ \nabla_ {\theta_ {t}} \mathcal {L} _ {t} \left(\theta_ {*}\right) \right]\right). \tag {9}
+$$
+
+Here,  $\gamma$  is the learning rate of Stochastic Gradient Descent (SGD),  $\theta_{K}$  is the parameter after  $K$  steps of gradient updates, and  $\theta_{*}$  is the optimal parameter.  $K$  is the total number of iterations required to achieve the accuracy  $\varepsilon$ .
+
+In short, Theorem 1 suggests that the learning rate should be scaled inversely proportional to the standard deviation of stochastic gradient. With an appropriately adjusted learning rate, the number of iterations needed for SGD to converge increases in proportion to the gradient's variance. Since modern optimizers like Adam [19] are designed to ensure convergence by inversely scaling the learning rate with the exponential moving standard deviation, we can infer that the iterations required for a subproblem  $\mathcal{L}_t$  to converge will approximately grow in proportion to the variance of the stochastic gradient at the optimal parameter.
+
+![](images/e12850ae528d0330879b9983d90285e58859a1d3cd3cbfe8f09dc8e21adc4bb9.jpg)  
+Figure 3. Adaptive Non-Uniform Timestep Sampling. Visualization of the overall architecture of our algorithm. Following an update from  $\theta_{k}$  to  $\theta_{k + 1}$ , we compute the set  $\{\delta_{k,\tau}^{t}\}_{\tau = 1}^{T}$  based on  $\theta_{k}$  and  $\theta_{k + 1}$ , and add it to a queue. A feature selection method is then applied to identify  $|S|$  timesteps that best explains  $\Delta_k^t$  from this queue. Then a timestep sampler  $\pi_{\phi}$  is trained to maximize  $\tilde{\Delta}_k^t$ , which samples timestep  $s$ $t$  that is expected to achieve the largest reduction in the diffusion loss. This process continues iteratively to minimize the diffusion loss at the chosen timesteps.
+
+Gradient variance of diffusion model Considering that the gradient variance at each timestep reflects the difficulty of the optimization problem at that specific timestep, we present the measured gradient variance during diffusion model training in the top row of Fig. 2. The results clearly indicate an imbalance in gradient variance across timesteps. Notably, for all three schedules, the gradient variances at early timesteps are significantly higher than those at later timesteps, and this trend persists without fading quickly. Since the variance at optimal parameters serves as a lower bound on the number of gradient iterations required for convergence (Theorem 1), we can anticipate that the early timesteps might pose a bottleneck to training efficiency when timesteps are sampled uniformly.
+
+The bottom row of Fig. 2 shows the actual timestepwise diffusion loss, and as expected, it shows that the loss near later timesteps decreases substantially faster than the loss near early timesteps. These results suggest that oversampling early timesteps and undersampling later timesteps could enhance overall training efficiency by effectively addressing the bottleneck presented by the subproblems at early timesteps.
+
+Interdependence of  $\mathcal{L}_t$  We have motivated the nonuniform sampling of timesteps by showing that we have varying gradient variance across timesteps, and this leads to a conclusion of non-uniform sampler that samples proportional to the gradient variance. However, without the assumption of independence across subproblems  $\mathcal{L}_t$  based on a specific parameterization, the optimality of such a sampler is not guaranteed.
+
+Fig. 4 illustrates the extent of interdependence among subproblems during diffusion model training. The figure reveals that the loss increase in untrained timesteps is
+
+significantly greater than the loss reduction in the trained timesteps, highlighting the strong interdependence between subproblems. Given that gradient variance can differ by up to a hundredfold (Fig. 2), it is evident that sampling in proportion to the gradient variance alone would likely be insufficient to effectively reduce the loss at later timesteps. Later through experiments, we confirmed that sampling timesteps in proportion to gradient variance yields poor performance (Tab. 4).
+
+# 5. Adaptive Non-Uniform Timestep Sampling for Diffusion Model Training
+
+Building on the need for non-uniform timestep sampling scheme discussed in the previous section, this section introduces an adaptive timestep sampler  $\pi$ , designed to enhance the training efficiency of diffusion models. We present a method for jointly training this sampler with the diffusion model in an online manner. This approach accelerates the diffusion training process with  $\pi$ , without requiring prior training data for its optimization.
+
+# 5.1. Non-uniform Timestep Sampler
+
+We propose a non-uniform timestep sampling strategy for diffusion model training, utilizing a timestep sampler  $\pi$  to improve training efficiency:
+
+$$
+\theta \leftarrow \theta - \gamma \nabla_ {\theta} \mathcal {L} _ {t}, t \sim \pi (\cdot | x _ {0}), \mathcal {L} _ {t} = c _ {t} \| \epsilon - \epsilon_ {\theta} (x _ {t}, t) \| ^ {2}, \tag {10}
+$$
+
+where  $c_{t}$  is a schedule-dependent constant as defined in Eq. (7). Note that we no longer use independent parameterization across timesteps.
+
+To leverage the fact that diffusion train loss is similar across closely related timesteps, we parameterize a neural
+
+![](images/eaea7cd9bbb7f6bbc0da176dedbd61cab75b8c52bbf719a9b8e002716953d9a0.jpg)  
+Figure 4. Difference in diffusion loss observed (after and before model update) after training a diffusion model by sampling only timesteps within the range [0, 200). A significant increase in loss is notably observed for the unsampled timesteps.
+
+network that generates two positive real scalars, which are then used to sample  $t$  from a Beta distribution:
+
+$$
+a, b \sim \pi_ {\phi} (\cdot | x _ {0}), \quad a > 0, b > 0, \quad t \sim \operatorname {B e t a} (a, b). \tag {11}
+$$
+
+This ensures our timestep sampler accounts for the high correlation between similar timesteps. For simplicity, we will continue to denote this process as  $t \sim \pi_{\phi}(\cdot | x_0)$ .
+
+Our goal is to train the timestep sampler  $\pi_{\phi}$  to minimize the original objective, the variational lower bound  $\mathcal{L}_{\mathrm{VLB}}$  as effectively as possible within a finite number of training steps. To achieve this, we seek to optimize the sampler  $\pi_{\phi_k}$  to sample timesteps  $t$  that yield the greatest reduction of  $\mathcal{L}_{\mathrm{VLB}}$  at each iteration  $k$ . In other words, the sampler  $\pi_{\phi_k}$  is trained to maximize the difference in  $\mathcal{L}_{\mathrm{VLB}}$  before and after updating the diffusion model at gradient step  $k$  using the sampled timestep  $t$ :
+
+$$
+\phi_ {k} = \arg \max  _ {\phi_ {k}} \mathbb {E} _ {x _ {0}, \epsilon , t \sim \pi_ {\phi_ {k}} (\cdot | x _ {0})} \left[ \Delta_ {k} ^ {t} \right], \quad \text {w h e r e} \tag {12}
+$$
+
+$$
+\Delta_ {k} ^ {t} = \frac {1}{T} \sum_ {\tau = 1} ^ {T} \delta_ {k, \tau} ^ {t} = \frac {1}{T} \sum_ {\tau = 1} ^ {T} \mathcal {L} _ {\tau} (\theta_ {k}) - \mathcal {L} _ {\tau} (\theta_ {k} - \gamma \nabla_ {\theta_ {k}} \mathcal {L} _ {t}).
+$$
+
+With such an optimized  $\phi_{k}$ , the diffusion model utilizing the sampler  $\pi_{\phi_k}$  would be able to adaptively select the best timesteps throughout the training process, minimizing  $\mathcal{L}_{\mathrm{VLB}}$  after the update  $\theta_{k}$ . In practice, we also include entropy regularization for the sampler  $\pi_{\phi}$  in  $\Delta_k^t$  to avoid premature convergence.
+
+Look-ahead challenge Evaluating  $\mathcal{L}_{\tau}(\theta_k - \gamma \nabla_{\theta_k}\mathcal{L}_t)$  requires computing how  $\theta$  changes after the update, which inherently involves a look-ahead process across different  $ts$ , making it computationally infeasible. To develop a practical algorithm, we opt for a single gradient update of  $\phi_k$ , using the last update of  $\theta$  as a reference. By applying the likelihood ratio gradient, we use the following update for  $\phi_k$ :
+
+$$
+\phi_ {k + 1} = \phi_ {k} + \gamma \cdot \Delta_ {k} ^ {t} \cdot \nabla_ {\phi_ {k}} \log \pi_ {\phi_ {k}} (\cdot | x _ {0}), \tag {13}
+$$
+
+Algorithm 1 Training DM with Timestep Sampler  
+Parameter:  $\epsilon_{\theta},\pi_{\phi}$    
+1: for gradient step  $k$  do   
+2:  $x_0\sim q(x_0),\epsilon \sim \mathcal{N}(0,I)$    
+3:  $a,b\sim \pi_{\phi_k}(\cdot |x_0),t\sim \mathrm{Beta}(a,b)$    
+4:  $x_{t} = \sqrt{\bar{\alpha}_{t}} x_{0} + \sqrt{1 - \bar{\alpha}_{t}}\epsilon$    
+5:  $\theta_{k + 1} = \theta_k - \gamma \cdot \nabla_{\theta_k}\| \epsilon -\epsilon_{\theta_k}(x_t,t)\| ^2$    
+6: if  $k$  mod  $f_{S} = 0$  then   
+7:  $\tilde{\Delta}_k^t =$  approximate  $\Delta_k^t$  with Algorithm 2   
+8:  $\phi_{k + 1} = \phi_k + \gamma \cdot \tilde{\Delta}_k^t\nabla_{\phi_k}\log \pi_{\phi_k}(a,b|x_0)$    
+9: end if   
+10: end for
+
+Algorithm 2 Approximation of  $\Delta_k^t$  
+Input:  $\theta_{k},\theta_{k + 1}$    
+Initialize: fixed-size queue  $Q$    
+1: Sample one  $x_0$  randomly from dataset   
+2: Calculate  $\{\delta_{k,\tau}^{t}\}_{\tau = 1}^{T}$  for the sampled  $x_0$    
+3: Push  $\{\delta_{k,\tau}^{t}\}_{\tau = 1}^{T}$  into queue  $Q$    
+4: if  $|Q| > 1$  then   
+5: Determine the set of timesteps  $S$  using a feature selec tion method with queue  $Q$    
+6: end if   
+7: return  $\tilde{\Delta}_k^t = \frac{1}{|S|}\sum_{\tau \in S}\delta_{k,\tau}^t$  for  $x_0s$  in current mini-batch
+
+where we use the identity  $\nabla_{\phi}\pi_{\phi} = \pi_{\phi}\cdot \nabla_{\phi}\log \pi_{\phi}$ . This is a stochastic approximation using sampled  $\epsilon$  and the Beta distribution parameters  $a$  and  $b$ . Algorithm 1 outlines the training procedure based on this gradient update for  $\phi_{k}$ . Note that  $c_{t}$  is not explicitly included in the algorithm, as it can be effectively subsumed into the sampler  $\pi_{\phi}$  if needed.
+
+# 5.2. Approximation of  $\Delta_k^t$
+
+Even with the reduced computation offered by the surrogate algorithm described above, evaluating  $\Delta_k^t$  still requires calculating the exact objective  $\mathcal{L}_{\mathrm{VLB}}$  before and after updating  $\theta_{k}$ , i.e.,  $\Delta_k^t = \mathcal{L}_{\mathrm{VLB}}(\theta_k) - \mathcal{L}_{\mathrm{VLB}}(\theta_{k + 1})$ . Fortunately, by leveraging the correlation between the losses of nearby timesteps, we can accurately approximate  $\Delta_k^t$  using a small set of timesteps  $S\subset \{t\}_{t = 1}^{T}$ , i.e.,
+
+$$
+\Delta_ {k} ^ {t} \approx \tilde {\Delta} _ {k} ^ {t} = \frac {1}{| S |} \sum_ {\tau \in S} \delta_ {k, \tau} ^ {t} = \frac {1}{| S |} \sum_ {\tau \in S} \mathcal {L} _ {\tau} \left(\theta_ {k}\right) - \mathcal {L} _ {\tau} \left(\theta_ {k + 1}\right). \tag {14}
+$$
+
+To determine  $S$ , for each  $k$ , we calculate  $\{\delta_{k,\tau}^t\}_{\tau = 1}^T$  for a single  $x_0$  and store these values in a queue. Using any off-the-shelf feature selection algorithm, we then identify the top  $|S|$  timesteps that best explain  $\Delta_k^t$  based on the data from the queue, with  $|S|$  being a hyperparameter. We mainly used a very small subset  $|S| = 3$  to approximate  $\Delta_k^t$ . Later through experiments, we demonstrate that this
+
+<table><tr><td></td><td colspan="5">Linear</td><td colspan="3">Cosine</td><td colspan="4">Quadratic</td></tr><tr><td></td><td>0.2M</td><td>0.4M</td><td>0.6M</td><td>0.8M</td><td>1.0M</td><td>0.2M</td><td>0.4M</td><td>0.6M</td><td>0.2M</td><td>0.4M</td><td>0.6M</td><td>0.8M</td></tr><tr><td>DDPM [14]</td><td>8.16</td><td>6.03</td><td>5.37</td><td>4.28</td><td>3.19</td><td>3.74</td><td>3.14</td><td>3.14</td><td>4.04</td><td>4.32</td><td>3.01</td><td>3.01</td></tr><tr><td>P2 [7]</td><td>6.45</td><td>5.21</td><td>4.83</td><td>4.70</td><td>4.70</td><td>5.64</td><td>5.21</td><td>5.04</td><td>4.93</td><td>4.06</td><td>3.74</td><td>3.66</td></tr><tr><td>Min-SNR [11]</td><td>4.40</td><td>3.51</td><td>3.28</td><td>3.16</td><td>3.16</td><td>3.51</td><td>3.07</td><td>3.07</td><td>4.76</td><td>4.17</td><td>4.17</td><td>4.17</td></tr><tr><td>Log-normal [17]</td><td>5.38</td><td>4.36</td><td>3.92</td><td>3.60</td><td>3.60</td><td>4.55</td><td>3.89</td><td>3.78</td><td>4.78</td><td>3.81</td><td>3.61</td><td>3.51</td></tr><tr><td>SpeeD [31]</td><td>4.08</td><td>3.66</td><td>3.57</td><td>3.57</td><td>3.57</td><td>4.12</td><td>3.57</td><td>3.57</td><td>3.99</td><td>3.54</td><td>3.52</td><td>3.52</td></tr><tr><td>Ours</td><td>3.99</td><td>3.21</td><td>3.10</td><td>2.94</td><td>2.94</td><td>3.87</td><td>3.06</td><td>3.00</td><td>3.96</td><td>3.07</td><td>2.95</td><td>2.95</td></tr></table>
+
+Table 1. Performance (FID) comparison on the CIFAR-10 dataset across different acceleration methods utilizing Linear, Cosine, and Quadratic schedules. We reported the lowest FID score achieved for each training duration. The algorithm achieving the lowest FID score is highlighted in bold.  
+
+<table><tr><td></td><td>100 epochs</td><td>300 epochs</td></tr><tr><td>LDM</td><td>12.87</td><td>8.72</td></tr><tr><td>P2</td><td>10.69</td><td>6.88</td></tr><tr><td>Min-SNR</td><td>10.87</td><td>6.83</td></tr><tr><td>Log-normal</td><td>12.81</td><td>7.28</td></tr><tr><td>SpeeD</td><td>14.40</td><td>7.98</td></tr><tr><td>Ours</td><td>10.94</td><td>6.62</td></tr></table>
+
+Table 2. Performance (FID) comparison on the CelebA-HQ  $256 \times$  256 dataset across different acceleration methods. The algorithm achieving the lowest FID score is highlighted in bold.
+
+small subset is sufficient for achieving strong performance with respect to wall-clock time (Fig. 7).
+
+The process of approximating  $\Delta_k^t$  using this feature selection approach is described in Algorithm 2. Note in practice, we further reduce the computational burden by updating  $\phi$  only once per  $f_{S}$  gradient iterations of the diffusion model. The overall structure of our algorithms is in Fig. 3. The computational overhead introduced by the proposed timestep sampling mechanism is in Appendix Sec. 9.
+
+# 6. Experiments
+
+In all experiments, we trained each diffusion architecture with its official implementation. Simultaneously, the timestep sampler  $\pi_{\phi}$  was trained using  $\Delta_k^t$ , which approximates the change in  $\mathcal{L}_{\mathrm{VLB}}$  before and after updating the diffusion model. We begin by evaluating the effectiveness of our learning method in comparison to heuristic approaches in Sec. 6.1. Then we assess the compatibility of our model with heuristic approaches in Sec. 6.2. Finally, we present additional analyses to further support our method in Sec. 6.3.
+
+Experimental Setup Frechet Inception Distance (FID) [13] is used to evaluate both the fidelity and coverage of
+
+generated images. We mainly used a machine equipped with NVIDIA RTX 4090 GPUs to train the models. Detailed hyperparameter values and implementation details can be found in Appendix Sec. 10 and Sec. 11.
+
+# 6.1. Comparison to Heuristic Acceleration Methods
+
+Comparison across various noise schedules To compare our model and other acceleration techniques across various noise schedules, we used the CIFAR-10  $32 \times 32$  dataset [2], utilizing DDPM [14] as the backbone architecture.
+
+Tab. 1 demonstrates that our method, despite its learning-based nature, generally achieves faster convergence and better performance at each iteration compared to heuristic approaches [7, 11, 17, 31]. Moreover, our method demonstrates consistent robustness across different noise schedules. In contrast, while Min-SNR performs the best among heuristic approaches with linear and cosine noise schedules, its performance declines significantly with a quadratic noise schedule. As illustrated in Fig. 5, our method achieves robust performance by adaptively sampling timesteps throughout the training process, whereas heuristic approaches rely on fixed sampling strategies. Furthermore, Fig. 1 shows that our method also achieves faster convergence in terms of wall-clock time compared to heuristic approaches.
+
+Comparison on high resolution images To evaluate the acceleration methods on high-resolution images, we used the CelebA-HQ  $256\times 256$  dataset [16], utilizing LDM [24] as the backbone architecture. Tab. 2 demonstrates that although heuristic methods initially achieve better performance, our learning-based approach surpasses them over time. We also found that P2 performs comparably well in this setting relative to the results in Tab. 1, suggesting that the performance of heuristics can vary across different settings. Visualizations of the generated CelebA-HQ images from our method and baselines are in Appendix Sec. 12.
+
+![](images/cc6ef70acc6871a7dad6abc3028d2647b2ad6a7c3bacaa4baf8b14a300150a98.jpg)  
+Figure 5. Visualization of timestep sampling schemes of our method, Min-SNR, P2, Log-Normal, and SpeeD. For weighting methods (Min-SNR, P2), weights at each timestep are converted to probabilities. The color bar on the right represents the progression of epochs, providing a view of how each scheduling method samples timesteps across the training process.
+
+# 6.2. Compatibility with Heuristic Approaches
+
+In Sec. 6.1, we demonstrated that while our learning-based method initially accelerates more slowly than the heuristic approaches due to its learning-based nature, it ultimately converges to a point with better optimality in later iterations. However, we found that heuristic weighting schemes can be incorporated into our model to achieve faster convergence from the beginning as well. Indeed, heuristic weightings can be regarded as part of the schedule-dependent constant  $c_{t}$ , and can be effectively subsumed into our sampler  $\pi_{\phi}$  (Sec. 5.1). Thus, our sampler can effectively account for the additional weight incorporated into the diffusion loss through heuristic approaches.
+
+To assess the compatibility of our model with heuristic approaches, we trained our model along with Min-SNR weighting on the ImageNet  $256 \times 256$  benchmark [8]. We used the encoder from LDM [24] to convert images of dimensions  $256 \times 256 \times 3$  into latent representations of size  $32 \times 32 \times 4$ . We used U-ViT [5] as our diffusion backbone. Tab. 3 shows that our method with heuristic weighting can achieve faster convergence from the beginning as well, ultimately converging to a lower FID score compared to the baseline.
+
+# 6.3. Analysis
+
+Analysis of non-uniform timestep sampling Fig. 6 (right column) illustrates the timestep distribution sampled by our method when applied to DDPM. Fig. 6 (left, center column) illustrates the variations in gradient variance and diffusion loss observed when training DDPM using  $\mathcal{L}_{\mathrm{VLB}}$  rather than  $\mathcal{L}_{\mathrm{DDPM}}$ , without incorporating our proposed method. Training with  $\mathcal{L}_{\mathrm{VLB}}$  differs from the conventional diffusion training with  $\mathcal{L}_{\mathrm{DDPM}}$  (Fig. 2) in that  $\mathcal{L}_{\mathrm{VLB}}$  weights the losses across timesteps differently using
+
+<table><tr><td></td><td>0.2M</td><td>0.4M</td><td>0.6M</td><td>0.8M</td><td>1.0M</td></tr><tr><td>Min-SNR</td><td>11.5</td><td>6.73</td><td>5.28</td><td>4.80</td><td>4.54</td></tr><tr><td>Min-SNR +Ours</td><td>11.3</td><td>6.59</td><td>5.19</td><td>4.66</td><td>4.33</td></tr></table>
+
+Table 3. Performance (FID) comparison on the ImageNet  $256 \times$  256 dataset between Min-SNR and our method with Min-SNR. The algorithm achieving the lowest FID score is highlighted in bold.  
+
+<table><tr><td></td><td>0.2M</td><td>0.4M</td><td>0.6M</td><td>0.8M</td><td>1.0M</td></tr><tr><td>Variance</td><td>21.3</td><td>16.0</td><td>13.5</td><td>12.1</td><td>11.3</td></tr><tr><td>Ours</td><td>3.99</td><td>3.21</td><td>3.10</td><td>2.94</td><td>2.94</td></tr></table>
+
+Table 4. Performance (FID) comparison on the CIFAR-10 dataset between our method and the approach that samples timesteps proportional to gradient variance. We reported the lowest FID score achieved for each training duration. The algorithm achieving the lowest FID score is highlighted in bold.
+
+$c_{t}$ . Since  $c_{t}$  has larger values at later timesteps, it leads to increased loss and gradient variance at these timesteps, especially with the cosine schedule as shown in Fig. 6.
+
+In Sec. 4, we argued that the number of iterations required for convergence would be proportional to the gradient variance at convergence, assuming independence across timesteps. The timestep sampling distributions shown in the plots support this intuition, with the linear schedule heavily sampling low timesteps and the cosine schedule sampling predominantly from both low and high timesteps. This suggests that the direction that most effectively reduces the original objective,  $\mathcal{L}_{\mathrm{VLB}}$ , somewhat aligns with sampling timesteps that exhibit higher gradient variance. On the other
+
+![](images/d87f1f67d475a349a8be8debe711032c2d19b863e4afb0f328ec2c58c8f95a58.jpg)
+
+![](images/92d89d1bbcd4333ff72411fd5aa7879cec8fd8f71e82eb0d8d44368efbf744bb.jpg)
+
+![](images/14dbc871b3c3469cc75438787daf11ad20de222e40e7514cf89fb4ae766ec2b2.jpg)
+
+![](images/99733aab40dc4d47eb354e4a6ea984f984de37fb839d8150b45ea58438a2f2f1.jpg)  
+Figure 6. (Left, Center) Comparison of gradient variance (top) and diffusion loss (bottom) over timesteps for linear and cosine schedules. The color bar on the right indicates the progression of epochs. Unlike the results presented in Fig. 2, these results are based on training DDPM [14] with  $\mathcal{L}_{\mathrm{VLB}}$  for 20 epochs. (Right) Visualization of our model's timestep distributions over training, with a linear schedule (top) and a cosine schedule (bottom).
+
+![](images/692c07629cfd794e0524019f7e63effa18e59c77749d64d821ba63499a3e7c0e.jpg)
+
+![](images/644b24e135693d85e075de8711f6ba5a2b15775d1ae07593b877cf1fe715eb70.jpg)
+
+![](images/f3fa4342f0fd2dcbe02a8505e2d55152aac6179928b4a362574e589b92777894.jpg)  
+Figure 7. Comparison of FID scores of baseline and our method against relative wall clock time, where 1D represents the time it takes for the baseline to converge.  $|S|$  denotes the number of timesteps used in our model to approximate  $\Delta_k^t$ .
+
+hand, Tab. 4 demonstrates that sampling timesteps in proportion to gradient variance results in poor performance, likely due to the interdependencies among the subproblems  $\mathcal{L}_t$ .
+
+Ablation study on  $|S|$  Fig. 7 illustrates the effect of varying the number of selected timesteps  $|S|$  on the performance of our proposed model. All choices of  $|S|$  show acceleration over the baseline, suggesting that we can approximate  $\Delta_k^t$  using only a small subset of timesteps. We observed that  $|S| = 3$  achieves the best performance in terms of wall clock time. Thus, we primarily used  $|S| = 3$  to approximate  $\Delta_k^t$  throughout our experiments.
+
+# 7. Conclusion
+
+In this paper, we discovered the variations in stochastic gradient variance across different diffusion training timesteps and argued that such variation can potentially lead to slow convergence. However, the strong interdependence of gradients across different timesteps suggests that this insight, on its own, does not directly result in optimal acceleration of diffusion training. To this end, we propose a more direct approach that estimates the impact of gradient updates on the variational lower bound for each timestep and adjusts the sampling frequency to prioritize timesteps that require further optimization. Experimental results demonstrate that our method significantly accelerates the convergence of the diffusion training process compared to previous heuristics, while also exhibiting robust performance across various datasets, scheduling strategies, and diffusion architectures.
+
+Limitations and future work Our learning-based method incurs higher computational costs compared to heuristic approaches, with resource demands increasing significantly in larger problem domains. Furthermore, we have not yet explored the application of our method within the score-based diffusion framework, which combines score-based generative modeling with diffusion probabilistic modeling and shows notable results in image generation tasks [17, 29, 34]. Addressing these limitations in terms of computational efficiency and extending the method's applicability within score-based diffusion models remains a promising direction for future research.
+
+# 8. Acknowledgements
+
+This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220311, Development of Goal-Oriented Reinforcement Learning Techniques for Contact-Rich Robotic Manipulation of Everyday Objects, No. RS-2024-00457882, AI Research Hub Project, and No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University)), the IITP(Institute of Information & Coucommunications Technology Planning & Evaluation)-ITRC(Information Technology Research Center)(IITP-2024-RS-2024-00436857) grant funded by the Korea government (Ministry of Science and ICT), the NRF (RS-2024-00451162) funded by the Ministry of Science and ICT, Korea, BK21 Four project of the National Research Foundation of Korea, and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(RS-2025-00560367).
+
+# References
+
+[1] Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022. 1  
+[2] Krizhevsky Alex. Learning multiple layers of features from tiny images. https://www.cs.toronto.edu/kriz/learning-features-2009-TR.pdf, 2009. 6  
+[3] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoffman, David Pfau, Tom Schaul, Brendan Shillingford, and Nando De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems, 29, 2016. 2  
+[4] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981-17993, 2021. 1  
+[5] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 7  
+[6] Irwan Bello, Barret Zoph, Vijay Vasudevan, and Quoc V Le. Neural optimizer search with reinforcement learning. In International Conference on Machine Learning, pages 459-468. PMLR, 2017. 2  
+[7] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 1, 2, 6  
+[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7
+
+[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1  
+[10] Guillaume Garrigos and Robert M Gower. Handbook of convergence theorems for (stochastic) gradient methods. arXiv preprint arXiv:2301.11235, 2023. 3  
+[11] Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffusion training via min-snr weighting strategy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7441-7451, 2023. 1, 2, 6  
+[12] Zhengfu He, Tianxiang Sun, Kuanning Wang, Xuanjing Huang, and Xipeng Qiu. Diffusionbert: Improving generative masked language models with diffusion models. arXiv preprint arXiv:2211.15029, 2022. 1  
+[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6  
+[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 6, 8, 1  
+[15] Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. arXiv preprint arXiv:2205.09991, 2022. 1  
+[16] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arxiv 2017. arXiv preprint arXiv:1710.10196, pages 1-26, 2018. 6  
+[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 2, 6, 8  
+[18] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 1  
+[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3  
+[20] Ke Li and Jitendra Malik. Learning to optimize. arXiv preprint arXiv:1606.01885, 2016. 2  
+[21] Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, et al. Pytorch distributed: Experiences on accelerating data parallel training. arXiv preprint arXiv:2006.15704, 2020. 1  
+[22] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 1  
+[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation
+
+and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1  
+[24] 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. 1, 6, 7  
+[25] Sebastian Sanokowski, Sepp Hochreiter, and Sebastian Lehner. A diffusion model framework for unsupervised neural combinatorial optimization. arXiv preprint arXiv:2406.01661, 2024. 1  
+[26] Flavio Schneider. Archisound: Audio generation with diffusion. arXiv preprint arXiv:2301.13267, 2023. 1  
+[27] 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  
+[28] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1  
+[29] 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. 8  
+[30] Zhiqing Sun and Yiming Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization. Advances in Neural Information Processing Systems, 36:3706-3731, 2023. 1  
+[31] Kai Wang, Yukun Zhou, Mingjia Shi, Zhihang Yuan, Yuzhang Shang, Xiaojiang Peng, Hanwang Zhang, and Yang You. A closer look at time steps is worthy of triple speed-up for diffusion model training. arXiv preprint arXiv:2405.17403, 2024. 1, 2, 6  
+[32] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning. arXiv preprint arXiv:2208.06193, 2022. 1  
+[33] Olga Wichrowska, Niru Maheswaranathan, Matthew W Hoffman, Sergio Gomez Colmenarejo, Misha Denil, Nando Freitas, and Jascha Sohl-Dickstein. Learned optimizers that scale and generalize. In International conference on machine learning, pages 3751-3760. PMLR, 2017. 2  
+[34] Zike Wu, Pan Zhou, Kenji Kawaguchi, and Hanwang Zhang. Fast diffusion model. arXiv preprint arXiv:2306.06991, 2023.8  
+[35] Zangwei Zheng, Xiangyu Peng, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac7485b83be62dba59a18f8142f2b2a1c62990e13737590a197520de53b98fea
+size 604243

CVPR/2025/Adaptive Non-Uniform Timestep Sampling for Accelerating Diffusion Model Training/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9de0f841d3274dd4213969c60c2b21f5fc28718878d6cf50a1095f19061e6726
+size 435557

CVPR/2025/Adaptive Parameter Selection for Tuning Vision-Language Models/1c740f1f-79f0-4120-87dd-dfad61af0ea1_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d5b5022ea9d667182390c46dcf0bf8290d952672d525504d7b1020489463d70
+size 90178

CVPR/2025/Adaptive Parameter Selection for Tuning Vision-Language Models/1c740f1f-79f0-4120-87dd-dfad61af0ea1_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10a6387c753d93bcc84e1184a8acdc234cfb5f1156eb01a584c88de4a2ff8bd
+size 109943