Add PC-Talk checkpoint download instructions and usage terms

README.md

@@ -1,24 +1,97 @@
-# Checkpoints
+---
+license: other
+license_name: checkpoint-specific-upstream-terms
+license_link: https://github.com/BQ-Wang0511/PC-Talk/blob/main/THIRD_PARTY_NOTICES.md
+library_name: pytorch
+pipeline_tag: image-to-video
+tags:
+  - talking-head-generation
+  - audio-driven-face-animation
+  - liveportrait
+  - cvpr-2026
+  - arxiv:2503.14295
+---
 
-The standalone source repository expects the following runtime weights. These
-files are distributed separately and are intentionally excluded from Git:
+# PC-Talk — CVPR 2026 Checkpoints
+
+Runtime checkpoints for **PC-Talk: Precise Facial Animation Control for
+Audio-Driven Talking Face Generation**.
+
+- [Paper](https://arxiv.org/abs/2503.14295)
+- [Project page](https://bq-wang0511.github.io/PC-Talk/)
+- [Source code and full inference instructions](https://github.com/BQ-Wang0511/PC-Talk)
+
+## Download
+
+Clone the source code and download the checkpoints directly into their runtime
+location:
+
+```bash
+git clone https://github.com/BQ-Wang0511/PC-Talk.git
+cd PC-Talk
+pip install -U huggingface_hub
+hf download doubi-killer/PC-Talk --local-dir pctalk/checkpoints
+```
+
+The repository contains 12 model files (about 920 MB), preserving the layout
+expected by PC-Talk:
 
 ```text
-checkpoints/
-├── lac.pth
-├── emc.pth
-├── style_encoder.pth
-├── audio_encoder.pth
-├── liveportrait/
-└── insightface/
+audio_encoder.pth
+lac.pth
+emc.pth
+style_encoder.pth
+liveportrait/
+├── landmark.onnx
+├── base_models/
+│   ├── appearance_feature_extractor.pth
+│   ├── motion_extractor.pth
+│   ├── spade_generator.pth
+│   └── warping_module.pth
+└── retargeting_models/
+    └── stitching_retargeting_module.pth
+insightface/models/buffalo_l/
+├── 2d106det.onnx
+└── det_10g.onnx
 ```
 
-`lac.pth` contains both the LAC `model` and the refinement `model_ft`, so it is
-used for both stages by default. EMC is loaded only when emotion control is
-requested, and the style encoder is loaded only for a reference style.
+`lac.pth` includes both the LAC model and its refinement model. EMC is used for
+emotion control; the style encoder is used for reference-style input. The
+LivePortrait and face-analysis weights are included for the inference backend.
 
-These files were selected from research experiment outputs, not from a
-documented public release. Their SHA-256 hashes are recorded in `SHA256SUMS`.
-Before publishing them, document training configuration and provenance, and
-confirm redistribution rights for the underlying datasets and every derived
-checkpoint.
+The original `SHA256SUMS` manifest uses paths relative to the **source-code
+repository root**. After downloading, verify the files from that root:
+
+```bash
+sha256sum -c pctalk/checkpoints/SHA256SUMS
+```
+
+These are downloadable inference assets, not a Transformers-format model. Use
+the PC-Talk source code to run inference. See its README for installation,
+image/video examples, emotion controls and articulation settings.
+
+## Usage terms and safety
+
+The source-code license does **not** automatically license every model weight.
+Third-party components retain their upstream terms:
+
+- [LivePortrait](https://github.com/KwaiVGI/LivePortrait)
+- [InsightFace](https://github.com/deepinsight/insightface)
+- [Wav2Lip](https://github.com/Rudrabha/Wav2Lip)
+
+In particular, the InsightFace model assets are restricted to non-commercial
+research use and must be replaced for commercial deployment. Consult the
+upstream licenses and the source repository's third-party notices before use.
+Only load PyTorch checkpoints from trusted sources.
+
+## Citation
+
+```bibtex
+@inproceedings{wang2026pc,
+  title={PC-Talk: Precise Facial Animation Control for Audio-Driven Talking Face Generation},
+  author={Wang, Baiqin and Zhu, Xiangyu and Shen, Fan and Xu, Hao and Lei, Zhen},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  pages={25153--25162},
+  year={2026}
+}
+```