File size: 2,833 Bytes
3739ed5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90e2d46
 
3739ed5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
---
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

<p align="center">
  <a href="https://arxiv.org/abs/2611.25553"><img src="https://img.shields.io/badge/arXiv-2611.25553-b31b1b.svg" alt="Paper: arXiv 2611.25553"></a>
  <a href="https://github.com/PlantPotatoOnMoon/TwoPass"><img src="https://img.shields.io/badge/GitHub-Code-181717?logo=github" alt="Code: GitHub"></a>
  <a href="#citation"><img src="https://img.shields.io/badge/Citation-BibTeX-2563eb" alt="Citation / BibTeX"></a>
</p>

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.