|
|
--- |
|
|
license: apache-2.0 |
|
|
datasets: |
|
|
- FastVideo/Wan-Syn_77x448x832_600k |
|
|
base_model: |
|
|
- Wan-AI/Wan2.1-T2V-14B-Diffusers |
|
|
--- |
|
|
# FastVideo FastWan2.1-T2V-14B-480P-Diffusers |
|
|
<p align="center"> |
|
|
<img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logo.jpg" width="200"/> |
|
|
</p> |
|
|
<div> |
|
|
<div align="center"> |
|
|
<a href="https://github.com/hao-ai-lab/FastVideo" target="_blank">FastVideo Team</a>  |
|
|
</div> |
|
|
|
|
|
<div align="center"> |
|
|
<a href="https://arxiv.org/pdf/2505.13389">Paper</a> | |
|
|
<a href="https://github.com/hao-ai-lab/FastVideo">Github</a> |
|
|
</div> |
|
|
</div> |
|
|
|
|
|
|
|
|
|
|
|
## Introduction |
|
|
|
|
|
This model is jointly finetuned with [DMD](https://arxiv.org/pdf/2405.14867) and [VSA](https://arxiv.org/pdf/2505.13389), based on [Wan-AI/Wan2.1-T2V-1.3B-Diffusers](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers). It supports efficient 3-step inference and generates high-quality videos at **61×448×832** resolution. We adopt the [FastVideo 480P Synthetic Wan dataset](https://huggingface.co/datasets/FastVideo/Wan-Syn_77x448x832_600k), consisting of 600k synthetic latents. |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Overview |
|
|
|
|
|
- 3-step inference is supported and achieves up to **50x speed up** on a single **H100** GPU. |
|
|
- Supports generating videos with resolution **61×448×832**. |
|
|
- Finetuning and inference scripts are available in the [FastVideo](https://github.com/hao-ai-lab/FastVideo) repository: |
|
|
- [Finetuning script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/distill/v1_distill_dmd_wan_VSA.sh) |
|
|
- [Inference script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/inference/v1_inference_wan_dmd.sh) |
|
|
- Try it out on **FastVideo** — we support a wide range of GPUs from **H100** to **4090**, and also support **Mac** users! |
|
|
|
|
|
### Training Infrastructure |
|
|
|
|
|
Training was conducted on **8 nodes with 64 H200 GPUs** in total, using a `global batch size = 64`. |
|
|
We enable `gradient checkpointing`, set `HSDP_shard_dim = 8`, `sequence_parallel_size = 4`, and use `learning rate = 1e-5`. |
|
|
We set **VSA attention sparsity** to 0.9, and training runs for **3000 steps (~52 hours)** |
|
|
The detailed **training example script** is available [here](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/distill/Wan-Syn-480P/distill_dmd_VSA_t2v_14B_480P.slurm). |
|
|
|
|
|
|
|
|
|
|
|
If you use FastWan2.1-T2V-14B-480P-Diffusers model for your research, please cite our paper: |
|
|
``` |
|
|
@article{zhang2025vsa, |
|
|
title={VSA: Faster Video Diffusion with Trainable Sparse Attention}, |
|
|
author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao}, |
|
|
journal={arXiv preprint arXiv:2505.13389}, |
|
|
year={2025} |
|
|
} |
|
|
@article{zhang2025fast, |
|
|
title={Fast video generation with sliding tile attention}, |
|
|
author={Zhang, Peiyuan and Chen, Yongqi and Su, Runlong and Ding, Hangliang and Stoica, Ion and Liu, Zhengzhong and Zhang, Hao}, |
|
|
journal={arXiv preprint arXiv:2502.04507}, |
|
|
year={2025} |
|
|
} |
|
|
``` |
