README.md

---
license: apache-2.0
language:
  - en
pipeline_tag: text-to-video
base_model:
  - Wan-AI/Wan2.1-T2V-1.3B
  - gdhe17/Self-Forcing
tags:
  - video-generation
  - autoregressive
  - diffusion
  - adversarial-gradient-forcing
  - arxiv:2610.14553
---

<h1 align="center" style="font-family: Georgia, 'Times New Roman', serif; font-size: 2.15em; font-weight: 500; line-height: 1.2; margin: 12px 0 8px;">Adversarial <span style="color: #658bc9;">Gradient Forcing</span></h1>
<p align="center" style="font-size: 1.02em; line-height: 1.5; margin: 0 0 14px; opacity: 0.85;">Robust Few-Step Autoregressive Video Generation</p>
<p align="center" style="margin: 8px 0 14px;">
  <sub>Zhejiang University &nbsp;·&nbsp; Alibaba Group &nbsp;·&nbsp; Nanjing University</sub>
</p>
<div align="center" style="display: flex; flex-wrap: wrap; justify-content: center; align-items: center; gap: 8px; margin: 12px 0 22px;"> <a href="https://github.com/wuli-art/Adversarial-Gradient-Forcing"><img src="assets/github.svg" height="44" style="display: inline-block; vertical-align: middle; margin: 4px 3px;" alt="GitHub code"></a> <a href="https://wuli-art.github.io/Adversarial-Gradient-Forcing/"><img src="assets/project.svg" height="44" style="display: inline-block; vertical-align: middle; margin: 4px 3px;" alt="Project page"></a> <a href="https://arxiv.org/abs/2610.14553"><img src="assets/arxiv.svg" height="44" style="display: inline-block; vertical-align: middle; margin: 4px 3px;" alt="arXiv: 2610.14553"></a>
</div>

**AGF** improves long-video generation by training models to withstand **adversarial perturbations in their generated history**. Trained on five-second rollouts, it supports generation up to four minutes with the same four-step sampler.

<p align="center"> <a href="https://wuli-art.github.io/Adversarial-Gradient-Forcing/#gallery"><img src="https://raw.githubusercontent.com/wuli-art/Adversarial-Gradient-Forcing/main/project-page/assets/showcase/agf-scenes.gif" width="960" alt="Selected AGF video excerpts: pirate ships, a robot, felt creatures, a creature by candlelight, a watercolor dance, and a yawning child."></a>
  <br>
  <sub>Generated with AGF · Selected video excerpts</sub>
</p>
<p align="center"><a href="https://wuli-art.github.io/Adversarial-Gradient-Forcing/#gallery">Full videos and comparisons ↗</a></p>

<table style="display: table; width: 100%; table-layout: fixed;">
  <tr>
    <td align="center" width="33%" style="text-align: center; padding: 12px 8px;"><strong>1.3B</strong><br><sub>Wan2.1 backbone</sub></td>
    <td align="center" width="33%" style="text-align: center; padding: 12px 8px;"><strong>4 steps</strong><br><sub>Denoising per block</sub></td>
    <td align="center" width="33%" style="text-align: center; padding: 12px 8px;"><strong>Up to 4 min</strong><br><sub>Video generation</sub></td>
  </tr>
</table>

## 🧩 Model details

AGF selects adversarial history perturbations using distribution-matching gradients from later video blocks, then trains jointly on clean and perturbed histories. A first-step prediction consistency regularizer helps preserve scene structure during the initial training stage.

| Property | Paper checkpoint |
| --- | --- |
| Architecture | Wan2.1 1.3B autoregressive video generator |
| Initialization | Self-Forcing ODE initialization |
| Training | 200 regularized AGF + 200 pure AGF outer iterations |
| Selected weights | Global step 400, `generator_ema` |
| Training rollouts | 5 seconds; effective batch size 64 |
| Training prompts | Filtered, LLM-extended VidProM prompts used by Self-Forcing |
| Inference | 832 × 480, 16 fps, four denoising steps per block |

Iterations count outer training loops; the full schedule contains **80 generator updates**.

## 🚀 Download and inference

The paper checkpoint is available as [`model.pt`](model.pt) (17.03 GB). Download it with the Hugging Face CLI:

```bash
hf download Wuli-art/Adversarial-Gradient-Forcing model.pt \
  --local-dir checkpoints/AGF
```

Follow the [installation and inference guide](https://github.com/wuli-art/Adversarial-Gradient-Forcing/blob/main/docs/inference.md) for the Python/CUDA environment and Wan 1.3B assets. Place the downloaded checkpoint at `checkpoints/AGF/model.pt` and run from the code repository root:

```bash
AGF_RUN_ID=preview bash scripts/infer.sh \
  --checkpoint checkpoints/AGF/model.pt \
  --prompts prompts/examples.txt --latent-frames 21 --gpus 0
```

The script selects **`generator_ema` by default**, matching the paper. Use `--latent-frames 21`, `240`, or `960` for approximately **5 seconds**, **1 minute**, or **4 minutes**. The default policy uses 12 latent frames of local attention, including 3 persistent sink frames, with relative sink positions enabled. See the [training guide](https://github.com/wuli-art/Adversarial-Gradient-Forcing/blob/main/docs/training.md) for the full training and checkpoint-resumption workflow.

## 📊 Results

Paper results under the shared evaluation protocol (higher is better). The 5-second and 60-second evaluations use 944 and 240 VBench prompts, respectively; the 240-second evaluation uses 120 MovieGenVideoBench extended prompts.

| Model | 5 s · VBench Total | 60 s · VBench Total | 240 s · Quality 5 |
| --- | ---: | ---: | ---: |
| **AGF** | **82.03** | **81.47** | **84.82** |

Quality 5 (Q5) is the arithmetic mean of subject consistency, background consistency, motion smoothness, aesthetic quality, and imaging quality. It excludes dynamic degree and is not an official VBench Total score. See the [evaluation protocol](https://github.com/wuli-art/Adversarial-Gradient-Forcing/blob/main/docs/evaluation.md) for metric definitions and reproducibility details.

## Intended use and limitations

AGF is intended for text-to-video generation and research on autoregressive video models. Long sequences may exhibit visual artifacts, motion inconsistencies, or prompt drift.

## 📝 Citation

```bibtex
@misc{xia2026adversarial,
  title         = {Adversarial Gradient Forcing for Robust Few-Step Autoregressive Video Generation},
  author        = {Chenxi Xia and Yixuan Xu and Shan Liu and Jiapeng Wang
                   and Maohua Li and Hanlin Tang and Kan Liu and Tao Lan
                   and Weiyi Lu and Xinggao Liu},
  year          = {2026},
  eprint        = {2610.14553},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.14553}
}
```

License: Apache 2.0. Third-party models, datasets, and other assets retain their respective licenses.