README.md

---
license: mit
pipeline_tag: text-to-video
---

# UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation

UniTemp is a few-step video diffusion model distilled from [Wan2.1](https://github.com/Wan-Video/Wan2.1) that generates videos in any temporal order (forward, backward, or in-between) with a single set of weights.

- **Paper:** [UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation](https://arxiv.org/abs/2606.18702)
- **Project Page:** [https://lzhangbj.github.io/projects/unitemp/](https://lzhangbj.github.io/projects/unitemp/)
- **Repository:** [https://github.com/lzhangbj/UniTemp](https://github.com/lzhangbj/UniTemp)

## Introduction
Existing autoregressive video diffusion models are typically restricted to forward temporal generation. UniTemp bridges this gap by supporting generation in arbitrary temporal directions. It introduces **blockwise anchor latents** to address inter-block discontinuities during backward generation and utilizes a **bidirectional distillation framework** for efficient training.

## Usage

Please refer to the [official repository](https://github.com/lzhangbj/UniTemp) for environment setup and model checkpoint preparation. The model supports three inference modes:

### Forward Generation (past -> future)
```bash
bash scripts/inference_forward.sh
```

### Backward Generation (future -> past)
```bash
bash scripts/inference_backward.sh
```

### In-between Generation
```bash
# head clip + tail clip, then interpolate the transition
bash scripts/inference_inbetween.sh
```

## Citation

```bibtex
@article{zhang2026unitemp,
  title={UniTemp: Unlocking Video Generation in Any Temporal Order via Bidirectional Distillation},
  author={Zhang, Lin and Mo, Sicheng and Cai, Zefan and Lin, Jinhong and Lin, Zihao and Gu, Jiuxiang and Singh, Krishna Kumar and Li, Yuheng and Li, Yin},
  journal={arXiv preprint arXiv:2606.18702},
  year={2026}
}
```