One-to-All Animation: Alignment-Free
Character Animation and Image Pose Transfer

Shijun Shi^1*, Jing Xu^2*, Zhihang Li³, Chunli Peng⁴, Xiaoda Yang⁵, Lijing Lu³,
Kai Hu^1†, Jiangning Zhang^5†

¹Jiangnan University   ²University of Science and Technology of China   ³Chinese Academy of Sciences
⁴Beijing University of Posts and Telecommunications   ⁵Zhejiang University

^*Equal contribution   ^†Corresponding authors

   

🌟 Highlights

We provide a complete and reproducible training and evaluation pipeline:

• ✅ Full Training Code: Three-stage progressive training from scratch
• ✅ Complete Benchmarks: Reproduction code and pre-trained checkpoints
• ✅ Flexible Training Codebase: Multi-resolution, multi-aspect-ratio, and multi-frame training codebase
• ✅ Datasets: Pre-processed open-source datasets + self-collected cartoon data

🔥 Update

• [2026.02.24] 🎉🎉🎉 One-to-All Animation has been accepted by CVPR 2026!
• [2025.12.22] 🔥🔥🔥 GPU-poor? Run One-to-All1.3B + ComfyUI for free on Kaggle’s 16 GB T4. We’ve released a zero-setup guide that runs the One-to-All ComfyUI workflow on Kaggle’s 16 GB T4—completely free😊. It takes 11 minutes to generate a 10-second 832×480 video. Tutorial: https://ncn0ojsozocg.feishu.cn/wiki/J9Ohwmtudin0vtkuyPccIZkhnZz
• [2025.12] kijai's ComfyUI WanVideoWrapper now integrates One‑to‑All Animation 14B! Huge thanks to kijai for the amazing work!!! Note: Our model supports both retargeted pose and direct pose (with reference preprocessing) from the original video. In addition, using lighter colors for the facial skeleton and landmarks helps achieve better identity consistency.
• [2025.11] Paper reproduction and evaluation code released.
• [2025.11] Sample training data and Benchmark on HuggingFace released.
• [2025.11] Inference and Training codes are released.
• [2025.11] 1.3B-v1, 1.3B-v2 and 14B checkpoints are released.

🎭 Showcase

Our model can adapt a single reference image to various motion patterns, demonstrating flexible motion control capabilities.

14B Model

Reference	Motion 1	Motion 2	Motion 3

1.3B Model

The 1.3 B model also delivers strong performance (from 1.3b_2 ckpt).

Reference	Motion 1	Motion 2	Motion 3

Also support longer video & out-of-domain cases

    

🔧 Dependencies and Installation

1. Clone Repo

   git clone https://github.com/ssj9596/One-to-All-Animation.git
   cd One-to-All-Animation

2. Create Conda Environment and Install Dependencies

   # create new conda env
   conda create -n one-to-all python=3.12
   conda activate one-to-all
   
   # install pytorch
   pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124
   # or
   pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 -i https://mirrors.aliyun.com/pypi/simple/
   
   # install python dependencies
   pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/
   
   
   # (Recommended) install flash attention 3 (or 2) from source:
   # https://github.com/Dao-AILab/flash-attention

3. Download Models

   • Download pretrained models

    cd ./pretrained_models 
    python download_pretrained_models.py

   • Download checkpoints

   cd ./checkpoints
   python download_checkpoints.py

   💡 Tip: Edit the script and uncomment the specific models you want to download.

   • 1.3B_1: Best performance on video benchmark among 1.3B models (paper results).
   • 1.3B_2: Further trained on v1 with large camera movement data and increased image ratio. Better for dynamic video generation. Best on image benchmark (paper results).
   • 14B: Best overall performance among 14B models (paper results).

☕️ Quick Inference

We provide several examples in the examples folder. Run the following commands to try it out:

# Step 1: Prepare model input
cd video-generation
python infer_preprocess.py

# Step 2: Run inference with your preferred model
python inference_1.3b.py  # For 1.3B model
# or
python inference_14b.py   # For 14B model

You can enter the script to modify the input path.

🎬 Training from scratch

💡 Data Collection Required: We find current open-source datasets are not sufficient for training from scratch. We strongly recommend collecting at least 3,000 additional high-quality video samples for better results.

We divide the training process into several steps to help you train from scratch (using 1.3B as an example).

1. Download Pretrained Models

   Download the base model from HuggingFace: Wan-AI/Wan2.1-T2V-1.3B-Diffusers

2. Download Training Datasets and Pose Pool

   cd datasets
   bash setup_datasets.sh

   This will download and prepare:

   • Training datasets (open-source + cartoon): datasets/opensource_dataset/
   • Pose pool for face enhancement: datasets/opensource_pose_pool/
   Manual Download Links

   • opensource_dataset
   • opensource_pose_pool

3. Training

   We provide three-stage training scripts:

   • Stage 1: Reference Extractor

   cd video-generation
   bash training_scripts/train1.3b_only_refextractor_2d.sh
   # Convert checkpoint to FP32
   cd outputs_wanx1.3b/train1.3b_only_refextractor_2d/checkpoint-xxx
   mkdir fp32_model_xxx
   python zero_to_fp32.py . fp32_model_xxx --safe_serialization
   # Run inference (update model path in inference_refextractor.py first)
   cd ../../../
   # Edit inference_refextractor.py and change ckpt_path to:
   # ./outputs_wanx1.3b/train1.3b_only_refextractor_2d/checkpoint-xxx/fp32_model_xxx
   python inference_refextractor.py

   • Stage 2: Pose Control

   bash training_scripts/train1.3b_posecontrol_prefix_2d.sh

   • Stage 3: Token Replace for Long video generation

   bash training_scripts/train1.3b_posecontrol_prefix_2d_tokenreplace.sh

   💡 Training Notes:

   • Each stage uses different training resolutions - check the scripts for specific resolution settings
   • Fine-tuning from our checkpoints: If you want to continue training from our pre-trained models, directly use the Stage 3 script and modify the checkpoint path

📊 Reproduce Paper Results

We provide scripts to reproduce the quantitative results reported in our paper.

1. Download Benchmark

   cd benchmark
   bash setup_datasets.sh

2. Prepare Model Input

   cd ../video-generation
   python reproduce/infer_preprocess.py 

3. Run Inference

   We provide inference scripts for different model sizes and datasets:

   # TikTok dataset
   python reproduce/inference_tiktok1.3b.py   # 1.3B model
   python reproduce/inference_tiktok14b.py    # 14B model
   
   # Cartoon dataset
   python reproduce/inference_cartoon1.3b.py  # 1.3B model
   python reproduce/inference_cartoon14b.py   # 14B model
   

4. Prepare gt/pred pairs for Judge

   cd ../benchmark
   # TikTok dataset
   python prepare_eval_frames_tiktok.py
   # Cartoon dataset
   python prepare_eval_frames_cartoon.py

5. Run judge

   # prepare DisCo environment and lpips fvd ckpt for judge
   cd DisCo
   # TikTok dataset
   bash eval_tiktok.sh
   python summary.py

Acknowledgments

Our project is based on opensora. Some codes are brought from StableAnimator and Wan-Animate. Thanks for their awesome works.

📝 Citation

If you find our work helpful or inspiring, please feel free to cite it.

@article{shi2025one,
  title={One-to-All Animation: Alignment-Free Character Animation and Image Pose Transfer},
  author={Shi, Shijun and Xu, Jing and Li, Zhihang and Peng, Chunli and Yang, Xiaoda and Lu, Lijing and Hu, Kai and Zhang, Jiangning},
  journal={arXiv preprint arXiv:2511.22940},
  year={2025}
}

📄 License

This repository is released under the Apache License 2.0.

📧 Contact

If you have any questions, please feel free to reach us at ssj180123@gmail.com