| --- |
| license: apache-2.0 |
| library_name: pytorch |
| pipeline_tag: image-to-video |
| --- |
| # HandWorld |
|
|
| Weights for **HandWorld: Hand-Centric Unified Video Action Generation**, CVPR 2026. |
|
|
| [Paper](https://openaccess.thecvf.com/content/CVPR2026/html/Sun_HandWorld_Hand-Centric_Unified_Video_Action_Generation_CVPR_2026_paper.html) |
|
|
| - `stage2.safetensors`: complete action-conditioned video model. |
| - `stage3.safetensors`: complete joint video/action prediction model. |
| - `preprocess/detector.pt`: optional left/right hand detector for RGB preprocessing. |
|
|
| Place the desired stage file in the inference repository's `weights/` directory. Download `models_t5_umt5-xxl-enc-bf16.pth`, `Wan2.2_VAE.pth`, and `google/umt5-xxl/` separately from the official [Wan-AI/Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B/tree/main) repository into `weights/Wan2.2-TI2V-5B/`, preserving those relative paths. No earlier-stage checkpoint or original Wan video-DiT shards are required. HaMeR and licensed MANO assets are obtained separately for preprocessing. |
|
|
