| --- |
| license: apache-2.0 |
| base_model: Wan-AI/Wan2.2-S2V-14B |
| pipeline_tag: image-to-video |
| --- |
| |
| # Wan2.2-S2V-14B (Diffusers layout) |
|
|
| This is [Wan-AI/Wan2.2-S2V-14B](https://huggingface.co/Wan-AI/Wan2.2-S2V-14B) |
| — a model that animates a picture of a person in sync with a voice recording — |
| repacked into the folder layout that |
| [FastVideo](https://github.com/hao-ai-lab/FastVideo) loads. |
|
|
| **No weights were changed.** The video model tensors are byte-identical copies |
| of the official release. The text encoder, VAE, tokenizer and scheduler come |
| from Wan-AI's own Diffusers release of Wan2.1 (the same weights S2V uses), and |
| the wav2vec2 speech encoder is the one bundled inside the official S2V repo. |
|
|
| Made with |
| [`scripts/checkpoint_conversion/wan_s2v_to_diffusers.py`](https://github.com/hao-ai-lab/FastVideo/pull/1683) |
| from FastVideo PR #1683. |
|
|
| ## Verification status |
|
|
| - Checked: all 1260 video-model tensors load into FastVideo's model with the |
| correct names and shapes (the conversion script verifies this automatically |
| and refuses to produce output otherwise). |
| - Not yet checked: end-to-end video generation on a GPU. If you run it and |
| find problems, please comment on the PR above. |
|
|
| ## Usage (with FastVideo, once PR #1683 is merged) |
|
|
| ```python |
| from fastvideo import VideoGenerator |
| |
| gen = VideoGenerator.from_pretrained("SuhaanCoding/Wan2.2-S2V-14B-Diffusers") |
| gen.generate_video( |
| prompt="...", |
| image_path="face.jpg", |
| audio_path="speech.wav", |
| output_path="out.mp4", |
| save_video=True, |
| ) |
| ``` |
|
|
