| --- |
| license: mit |
| language: |
| - en |
| base_model: |
| - microsoft/TRELLIS-image-large |
| pipeline_tag: image-to-3d |
| tags: |
| - animatable |
| - rigging |
| - 3D |
| - Tripo |
| - VAST |
| --- |
| # AniGen_Weights |
| |
| Pretrained checkpoints for [AniGen](https://github.com/VAST-AI-Research/AniGen), a unified framework for generating animatable 3D assets from a single image. |
| |
| <p align="center"> |
| <a href="https://arxiv.org/pdf/2604.08746"><img src="https://img.shields.io/badge/arXiv-Paper-red?logo=arxiv&logoColor=white" alt="arXiv"></a> |
| <a href="https://yihua7.github.io/AniGen_web/"><img src="https://img.shields.io/badge/Project_Page-Website-green?logo=googlechrome&logoColor=white" alt="Project Page"></a> |
| <a href="https://www.tripo3d.ai"><img src="https://img.shields.io/badge/Tripo-AI_3D_Workspace-orange" alt="Tripo"></a> |
| <a href="https://huggingface.co/spaces/VAST-AI/AniGen"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Live_Demo-blue" alt="Hugging Face Demo"></a> |
| <a href="https://github.com/VAST-AI-Research/AniGen"><img src="https://img.shields.io/badge/GitHub-Repository-black?logo=github&logoColor=white" alt="GitHub"></a> |
| </p> |
| |
| This repository stores the contents of the `ckpts/` directory used by the AniGen codebase, including: |
| |
| - AniGen stage checkpoints |
| - DINOv2 vision encoder weights |
| - DSINE normal estimation weights |
| - VGG backbone weights |
| |
| ## What Is Included |
| |
| The repository is organized exactly like the `ckpts/` folder expected by the main AniGen repo: |
| |
| ```text |
| ckpts/ |
| ├── anigen/ |
| │ ├── ss_dae/ |
| │ ├── slat_dae/ |
| │ ├── ss_flow_duet/ |
| │ ├── ss_flow_epic/ |
| │ ├── ss_flow_solo/ |
| │ ├── slat_flow_auto/ |
| │ ├── slat_flow_control/ |
| │ └── slat_flow_gsn_auto/ |
| ├── dinov2/ |
| ├── dsine/ |
| └── vgg/ |
| ``` |
| |
| Approximate total size: about 23 GB. |
| |
| ## Recommended Checkpoints |
| |
| For most users, we recommend: |
| |
| - `ss_flow_duet` for sparse structure generation |
| - `slat_flow_auto` for structured latent generation |
| |
| This combination is also the default setup used by the AniGen inference example. |
| |
| ## Checkpoint Overview |
| |
| ### Core AniGen checkpoints |
| |
| | Folder | Purpose | |
| | --- | --- | |
| | `ckpts/anigen/ss_dae` | Sparse Structure autoencoder | |
| | `ckpts/anigen/slat_dae` | Structured Latent autoencoder | |
| | `ckpts/anigen/ss_flow_duet` | SS flow model with stronger skeleton detail | |
| | `ckpts/anigen/ss_flow_epic` | SS flow model balancing geometry and skeleton quality | |
| | `ckpts/anigen/ss_flow_solo` | SS flow model with stronger geometry generalization | |
| | `ckpts/anigen/slat_flow_auto` | SLAT flow model with automatic joint-count prediction | |
| | `ckpts/anigen/slat_flow_control` | SLAT flow model with controllable joint density | |
| | `ckpts/anigen/slat_flow_gsn_auto` | Additional SLAT variant included in the release | |
| |
| ### Dependency checkpoints |
| |
| | Folder | Purpose | |
| | --- | --- | |
| | `ckpts/dinov2` | DINOv2 encoder files and pretrained ViT-L/14 weights | |
| | `ckpts/dsine` | DSINE normal estimation weights | |
| | `ckpts/vgg` | VGG weights used by the pipeline | |
| |
| ## How To Use |
| |
| Clone the main AniGen repository first: |
| |
| ```bash |
| git clone --recurse-submodules https://github.com/VAST-AI-Research/AniGen.git |
| cd AniGen |
| ``` |
| |
| Then download this weights repository so that the folder structure is preserved under the project root. |
| |
| ### Option 1: Download with `huggingface_hub` |
| |
| ```bash |
| python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='VAST-AI/AniGen_Weights', repo_type='model', local_dir='.', local_dir_use_symlinks=False)" |
| ``` |
| |
| After download, you should have paths like: |
| |
| ```text |
| ckpts/anigen/ss_flow_duet/ckpts/denoiser.pt |
| ckpts/anigen/slat_flow_auto/ckpts/denoiser.pt |
| ckpts/dsine/dsine.pt |
| ckpts/vgg/vgg16-397923af.pth |
| ``` |
| |
| ### Option 2: Download from the web UI |
| |
| You can also download this repository from the Hugging Face file browser and place the entire `ckpts/` folder at the root of the AniGen project. |
| |
| ## Run AniGen With These Weights |
| |
| Once the `ckpts/` folder is in place, you can run: |
| |
| ```bash |
| python example.py --image_path assets/cond_images/trex.png |
| ``` |
| |
| Or launch the Gradio demo: |
| |
| ```bash |
| python app.py |
| ``` |
| |
| ## Notes |
| |
| - Keep the directory names unchanged. The AniGen code expects the exact `ckpts/...` layout shown above. |
| - The code repository may automatically fetch missing files in some setups, but this weights repository is the recommended way to pre-download and manage checkpoints explicitly. |
| - `slat_flow_control` supports joint density control, while `slat_flow_auto` is the best default for general use. |
| |
| ## Related Links |
| |
| - Best AI 3D studio -- Tripo: https://www.tripo3d.ai |
| - Main code repository: https://github.com/VAST-AI-Research/AniGen |
| - Project page: https://yihua7.github.io/AniGen-web/ |
| - Demo: https://huggingface.co/spaces/VAST-AI/AniGen |
| - Paper: https://arxiv.org/pdf/2604.08746 |
| |
| ## Citation |
| |
| ```bibtex |
| @article{huang2026anigen, |
| title = {AniGen: Unified $S^3$ Fields for Animatable 3D Asset Generation}, |
| author = {Huang, Yi-Hua and Zhou, Zi-Xin and He, Yuting and Chang, Chirui |
| and Pu, Cheng-Feng and Yang, Ziyi and Guo, Yuan-Chen |
| and Cao, Yan-Pei and Qi, Xiaojuan}, |
| journal = {ACM SIGGRAPH}, |
| year = {2026} |
| } |
| ``` |
