Simplify model card to project and checkpoint basics

README.md

@@ -9,94 +9,29 @@ tags:
   - flame
   - dinov3
   - arxiv:2608.09238
-datasets:
-  - LinzhouLi/RealDenseFace-Ava256
 ---
 
 # RealDenseFace
 
 **Real-time Monocular 3D Face Reconstruction from Dense UV-space Priors**
 
-Linzhou Li, Tianjia Shao, Kun Zhou · State Key Lab of CAD&CG, Zhejiang University
+Linzhou Li, Tianjia Shao, Kun Zhou  
+State Key Lab of CAD&CG, Zhejiang University
 
-[Project](https://gapszju.github.io/RealDenseFace/) · [Paper](https://arxiv.org/abs/2608.09238) · [Code](https://github.com/LinzhouLi/RealDenseFace) · [Ava-256 training data](https://huggingface.co/datasets/LinzhouLi/RealDenseFace-Ava256)
+[Project](https://gapszju.github.io/RealDenseFace/) · [Paper](https://arxiv.org/abs/2608.09238) · [Code](https://github.com/LinzhouLi/RealDenseFace)
 
-## Overview
+RealDenseFace combines dense UV-space correspondence and relative-depth predictions with a tailored CUDA Gauss-Newton solver for FLAME reconstruction. It supports single-image fitting, offline sequence reconstruction, online tracking, and calibrated multi-view fitting.
 
-RealDenseFace reconstructs a FLAME face from a monocular RGB image in two stages. A DINOv3-based network predicts dense UV-to-image correspondences, relative depth, and their log-variances. Predictions sampled at FLAME vertex UV coordinates provide targets and confidence weights for a tailored CUDA Gauss-Newton solver.
+## Models
 
-The code supports single-image fitting, offline monocular sequence reconstruction, online tracking, and calibrated multi-view fitting. This repository hosts the pretrained dense-prior networks and their matching configurations; the inference and fitting implementation is maintained on GitHub.
+| Model | Encoder | Input size | Checkpoint |
+| --- | --- | --- | --- |
+| ViT-S | DINOv3 ViT-S+ | 512 × 512 | [vits.pth](weights/realdenseface/vits.pth) |
+| ViT-B | DINOv3 ViT-B | 512 × 512 | [vitb.pth](weights/realdenseface/vitb.pth) |
 
-## Checkpoints
+ViT-S is the default model for online tracking; ViT-B is the default for single-image and offline reconstruction. Matching model and fitting configurations are included in `configs/`.
 
-| Variant | Encoder | Input resolution | Checkpoint | Typical use |
-| --- | --- | --- | --- | --- |
-| ViT-S | DINOv3 ViT-S+ | 512 x 512 | `weights/realdenseface/vits.pth` | Default online tracking model |
-| ViT-B | DINOv3 ViT-B | 512 x 512 | `weights/realdenseface/vitb.pth` | Default single-image and offline model |
-
-The `.pth` files are PyTorch state dictionaries consumed by the official code. They are not Transformers checkpoints. Both model configurations use `scale_coord: 1.0` and `scale_depth: 8.0`; the official inferencer handles normalization and output rescaling.
-
-## Download and run
-
-Clone the implementation and run these commands from its root:
-
-```bash
-git clone https://github.com/LinzhouLi/RealDenseFace.git
-cd RealDenseFace
-pip install -U huggingface_hub
-hf download LinzhouLi/RealDenseFace --local-dir .
-```
-
-For a private repository, authenticate first with `hf auth login` using an account with access.
-
-Set up Python 3.11, a CUDA-enabled PyTorch build, and a compatible NVIDIA GPU/CUDA toolkit. Install the remaining dependencies and compile the solver:
-
-```bash
-pip install -r requirements.txt
-pip install -e ./cuda_extensions/flame_solver --no-build-isolation
-```
-
-Before running, obtain the additional resources described in the [GitHub installation instructions](https://github.com/LinzhouLi/RealDenseFace#installation):
-
-```text
-weights/facebox/face_box.pth
-weights/flame/flame_assets.npz
-weights/flame/flame2023.pkl
-```
-
-These resources are not included in this model repository. Obtain the standard FLAME 2023 model separately from the [official FLAME website](https://flame.is.tue.mpg.de/), subject to its license.
-
-Single-image fitting with ViT-B:
-
-```bash
-python fit_single_image.py --input path/to/image.jpg --output_npz output/face.npz --output_image output/face.jpg
-```
-
-To use ViT-S, add `--model_config configs/model/vits.yaml --model_weights weights/realdenseface/vits.pth`.
-
-Online video tracking with ViT-S:
-
-```bash
-python track_video_online.py --input assets/demo_video.mp4 --output_mp4 output/tracking.mp4
-```
-
-Outputs include fitted FLAME parameters and optional reconstruction visualizations. See the GitHub README for offline sequence and multi-view examples.
-
-## Training data and evaluation
-
-The paper describes training on processed AVA-256, NPHM, and FaceScape data. The [RealDenseFace-Ava256 dataset](https://huggingface.co/datasets/LinzhouLi/RealDenseFace-Ava256) provides the processed Ava-256 images and registrations; it is one component of the training data.
-
-The paper reports approximately 0.18 seconds for single-image fitting with unknown intrinsics, 84 FPS online tracking, and approximately 24x faster offline sequence processing than VHAP on an RTX 4090. These are paper-reported results, not measurements performed during this upload. Runtime depends on hardware, compilation, model variant, and fitting settings.
-
-## Usage notes
-
-The output geometry is constrained by the FLAME model. Camera calibration, image cropping, and the quality of the predicted dense priors affect reconstruction. The official task configurations differ: the single-image and offline configurations use relative-depth weight 2.0, while the online and NeRSemble multi-view configurations use weight 0.0. Use the matching configuration when reproducing a task.
-
-The model and fitting code require the custom runtime described above; uploading the checkpoints does not by itself provide a hosted inference API or interactive demo.
-
-## License
-
-This repository uses the MIT license selected for this release; see [LICENSE](LICENSE). Third-party components, including DINOv3, FaceBox, FLAME, and the source datasets, retain their respective licenses. This repository does not redistribute the FLAME model or grant rights to those external resources. The released Ava-256 dataset has its own CC BY-NC 4.0 terms.
+For installation and usage, see the [GitHub repository](https://github.com/LinzhouLi/RealDenseFace). FLAME model files must be obtained separately from the [official website](https://flame.is.tue.mpg.de/) under its license.
 
 ## Citation
 
@@ -107,4 +42,4 @@ This repository uses the MIT license selected for this release; see [LICENSE](LI
   journal = {arXiv preprint arXiv:2608.09238},
   year    = {2026}
 }
-```
+```

checksums.sha256

@@ -5,6 +5,6 @@ e0433a66441abcad8229fb93c373b81707a24659ddf7a7d9c8d6e7226dbb98e4  configs/fittin
 8c1cced899f4409c71f3e50e174d312a161fd26982aad0894ddde4a30246a6ab  configs/model/vitb.yaml
 1f02b837a62dfa1959b6e09be23332441eacf6a74892d50c0e18d569eeec8686  configs/model/vits.yaml
 718974a99973b53f9cfce42dc1528368b6cbf6b7d3cc97981269c9696fb5b966  LICENSE
-502f3280e21d32bd80304799ce8b7da80936520305389037d8215c63fd16d6d1  README.md
+7ed3d97070800600c8fa59a8c31eb277bf819375fca538300099ed9b41175917  README.md
 3b3bbd026c277aae561aed0047f43bbae2833c000779904f6dd73e97e321c604  weights/realdenseface/vitb.pth
 9644fb9c7b7b8b5b6ca0d06fdebae418c154f1f4224a1ef5298777cc399ea15c  weights/realdenseface/vits.pth