--- license: apache-2.0 base_model: Wan-AI/Wan2.2-TI2V-5B library_name: pytorch pipeline_tag: image-to-video tags: - world-model - video-prediction - robotics - robot-manipulation - action-conditioned - few-step - wan2.2 datasets: - fangqi/IRASim --- # TwoPass: Draft-Guided Few-Step Video Prediction for Robotic Manipulation
Checkpoints of **TwoPass**, an action-conditioned video prediction model for robotic manipulation fine-tuned from [Wan2.2-TI2V-5B](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B). Given one observed frame and the future robot actions, it predicts the following frames in as few as four network calls. | File | Dataset | Size | |---|---|---| | [`bridge/twopass_bridge.pt`](bridge/twopass_bridge.pt) | Bridge | 20.2 GB | | [`languagetable/twopass_languagetable.pt`](languagetable/twopass_languagetable.pt) | Language Table | 20.2 GB | Each file is a PyTorch checkpoint holding the full model weights (`model`, fp32) and the training step (`step`). ## Usage Install the code from [GitHub](https://github.com/PlantPotatoOnMoon/TwoPass), then download the checkpoints together with the Wan2.2 base weights: ```bash hf download Junjun2333/TwoPass --local-dir ./checkpoints hf download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B ``` Evaluate on the Bridge test set with 4 network calls per clip (use `configs/languagetable.json` and the Language Table checkpoint for the other dataset; data preparation is described in the GitHub README): ```bash python -m twopass.evaluate --config configs/bridge.json --nfe 4 \ --checkpoint ./checkpoints/bridge/twopass_bridge.pt --wan_dir ./Wan2.2-TI2V-5B \ --data_root $DATA_ROOT --i3d_path /path/to/i3d_torchscript.pt --out_dir ./eval/bridge python -m twopass.aggregate --config configs/bridge.json \ --data_root $DATA_ROOT --eval_dir ./eval/bridge ``` ## Citation ```bibtex @article{liu2026twopass, title = {TwoPass: Draft-Guided Few-Step Video Prediction for Robotic Manipulation}, author = {Liu, Yijun and Ge, Shijia and Feng, Tongtong and Wang, Xin and Meng, Yuan and Liu, Yuwei and Wang, Zhi and Zhu, Wenwu}, journal = {arXiv preprint arXiv:2611.25553}, year = {2026}, eprint = {2611.25553}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2611.25553} } ``` ## License Apache License 2.0, following Wan2.2.