FastVideo/FastWan2.1-T2V-1.3B-Diffusers

1---2base_model:3- Wan-AI/Wan2.1-T2V-1.3B-Diffusers4datasets:5- FastVideo/Wan-Syn_77x448x832_600k6license: apache-2.07pipeline_tag: text-to-video8library_name: diffusers9---10 11# FastVideo FastWan2.1-T2V-1.3B-Diffusers Model12<p align="center">13  <img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logo.png" width="200"/>14</p>15<div>16  <div align="center">17    <a href="https://github.com/hao-ai-lab/FastVideo" target="_blank">FastVideo Team</a>&emsp;18  </div>19 20  <div align="center">21    <a href="https://arxiv.org/pdf/2505.13389">Paper</a> | 22    <a href="https://github.com/hao-ai-lab/FastVideo">Github</a> |23    <a href="https://hao-ai-lab.github.io/FastVideo">Project Page</a>24  </div>25</div>26 27## Online Demo28You can try our models [here](https://fastwan.fastvideo.org/)！29 30## Introduction31We're excited to introduce the **FastWan2.1 series**—a new line of models finetuned with our novel **Sparse-distill** strategy. This approach jointly integrates DMD and VSA in a single training process, combining the benefits of both **distillation** to shorten diffusion steps and **sparse attention** to reduce attention computations, enabling even faster video generation.32 33FastWan2.1-T2V-1.3B-Diffusers is built upon Wan-AI/Wan2.1-T2V-1.3B-Diffusers. It supports efficient **3-step inference** and produces high-quality videos at 61×448×832 resolution. For training, we use the FastVideo 480P Synthetic Wan dataset, which contains 600k synthetic latents.34 35---36 37## Model Overview38 39- 3-step inference is supported and achieves up to **16 FPS** on a single **H100** GPU.40- Our model is trained on **61×448×832** resolution, but it supports generating videos with **any resolution**.(quality may degrade)41- Finetuning and inference scripts are available in the [FastVideo](https://github.com/hao-ai-lab/FastVideo) repository:  42  - [1 Node/GPU debugging finetuning script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/distill/v1_distill_dmd_wan_VSA.sh)43  - [Slurm training example script](https://github.com/hao-ai-lab/FastVideo/blob/main/examples/distill/Wan2.1-T2V/Wan-Syn-Data-480P/distill_dmd_VSA_t2v_1.3B.slurm)  44  - [Inference script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/inference/v1_inference_wan_dmd.sh)45```python46# install FastVideo and VSA first47git clone https://github.com/hao-ai-lab/FastVideo48pip install -e .49cd csrc/attn50git submodule update --init --recursive51python setup_vsa.py install52 53num_gpus=154export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN55export MODEL_BASE=FastVideo/FastWan2.1-T2V-1.3B-Diffusers56# export MODEL_BASE=hunyuanvideo-community/HunyuanVideo57# You can either use --prompt or --prompt-txt, but not both.58fastvideo generate \59    --model-path $MODEL_BASE \60    --sp-size $num_gpus \61    --tp-size 1 \62    --num-gpus $num_gpus \63    --height 480 \64    --width 848 \65    --num-frames 81 \66    --num-inference-steps 3 \67    --fps 16 \68    --prompt-txt assets/prompt.txt \69    --negative-prompt "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" \70    --seed 1024 \71    --output-path outputs_video_dmd/ \72    --VSA-sparsity 0.8 \73    --dmd-denoising-steps "1000,757,522"74```75- Try it out on **FastVideo** — we support a wide range of GPUs from **H100** to **4090**, and also support **Mac** users!76 77### Training Infrastructure78 79Training was conducted on **4 nodes with 32 H200 GPUs** in total, using a `global batch size = 64`.  80We enable `gradient checkpointing`, set `gradient_accumulation_steps=2`, and use `learning rate = 1e-5`.  81We set **VSA attention sparsity** to 0.8, and training runs for **4000 steps (~12 hours)**   82 83If you use the FastWan2.1-T2V-1.3B-Diffusers model for your research, please cite our paper:84```85@article{zhang2025vsa,86  title={VSA: Faster Video Diffusion with Trainable Sparse Attention},87  author={Zhang, Peiyuan and Huang, Haofeng and Chen, Yongqi and Lin, Will and Liu, Zhengzhong and Stoica, Ion and Xing, Eric and Zhang, Hao},88  journal={arXiv preprint arXiv:2505.13389},89  year={2025}90}91@article{zhang2025fast,92  title={Fast video generation with sliding tile attention},93  author={Zhang, Peiyuan and Chen, Yongqi and Su, Runlong and Ding, Hangliang and Stoica, Ion and Liu, Zhengzhong and Zhang, Hao},94  journal={arXiv preprint arXiv:2502.04507},95  year={2025}96}97```