Diffusers
Safetensors
WanDMDPipeline
Files changed (1) hide show
  1. README.md +10 -49
README.md CHANGED
@@ -7,7 +7,7 @@ base_model:
7
  ---
8
  # FastVideo FastWan2.1-T2V-14B-480P-Diffusers
9
  <p align="center">
10
- <img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logo.png" width="200"/>
11
  </p>
12
  <div>
13
  <div align="center">
@@ -24,59 +24,17 @@ base_model:
24
 
25
  ## Introduction
26
 
27
- We'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 1 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.
28
-
29
- FastWan2.1-T2V-14B-480P-Diffuserss is built upon Wan-AI/Wan2.1-T2V-14B-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.
30
 
 
31
 
32
  ## Model Overview
33
 
34
- - 3-step inference is supported and achieves up to **50x speed up** at 480P, **70x speed up** at 720P during denoising loop on a single **H100** GPU.
35
- - Our model is trained on **61×448×832** resolution, but it supports generating videos with **any resolution**.(**480P, 720P, quality may degrade**)
36
  - Finetuning and inference scripts are available in the [FastVideo](https://github.com/hao-ai-lab/FastVideo) repository:
37
- - [1 Node/GPU debugging finetuning script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/distill/v1_distill_dmd_wan_VSA.sh)
38
- - [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_14B.slurm)
39
- - Inference script in FastVideo:
40
- ```python
41
- #!/bin/bash
42
-
43
- # install FastVideo and VSA first
44
- git clone https://github.com/hao-ai-lab/FastVideo
45
- pip install -e .
46
- cd csrc/attn
47
- git submodule update --init --recursive
48
- python setup_vsa.py install
49
-
50
- num_gpus=1
51
- export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
52
- export MODEL_BASE=FastVideo/FastWan2.1-T2V-14B-480P-Diffusers
53
-
54
- # 720P 14B
55
- # Torch compile is enabled. Expect generating the first video to be slow.
56
- # Speed on H200 after warmup 3/3 [00:13<00:00, 4.45s/it]:
57
- num_gpus=1
58
- export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN
59
- export MODEL_BASE=FastVideo/FastWan2.1-T2V-14B-480P-Diffusers
60
- # export MODEL_BASE=hunyuanvideo-community/HunyuanVideo
61
- # You can either use --prompt or --prompt-txt, but not both.
62
- fastvideo generate \
63
- --model-path $MODEL_BASE \
64
- --sp-size $num_gpus \
65
- --tp-size 1 \
66
- --num-gpus $num_gpus \
67
- --height 720 \
68
- --width 1280 \
69
- --num-frames 81 \
70
- --num-inference-steps 3 \
71
- --fps 16 \
72
- --prompt-txt assets/prompt.txt \
73
- --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" \
74
- --seed 1024 \
75
- --output-path outputs_video_dmd_14B_720P/ \
76
- --VSA-sparsity 0.9 \
77
- --dmd-denoising-steps "1000,757,522" \
78
- --enable_torch_compile
79
- ```
80
  - Try it out on **FastVideo** — we support a wide range of GPUs from **H100** to **4090**, and also support **Mac** users!
81
 
82
  ### Training Infrastructure
@@ -84,6 +42,9 @@ fastvideo generate \
84
  Training was conducted on **8 nodes with 64 H200 GPUs** in total, using a `global batch size = 64`.
85
  We enable `gradient checkpointing`, set `HSDP_shard_dim = 8`, `sequence_parallel_size = 4`, and use `learning rate = 1e-5`.
86
  We set **VSA attention sparsity** to 0.9, and training runs for **3000 steps (~52 hours)**
 
 
 
87
 
88
  If you use FastWan2.1-T2V-14B-480P-Diffusers model for your research, please cite our paper:
89
  ```
 
7
  ---
8
  # FastVideo FastWan2.1-T2V-14B-480P-Diffusers
9
  <p align="center">
10
+ <img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logo.jpg" width="200"/>
11
  </p>
12
  <div>
13
  <div align="center">
 
24
 
25
  ## Introduction
26
 
27
+ 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.
 
 
28
 
29
+ ---
30
 
31
  ## Model Overview
32
 
33
+ - 3-step inference is supported and achieves up to **50x speed up** on a single **H100** GPU.
34
+ - Supports generating videos with resolution **61×448×832**.
35
  - Finetuning and inference scripts are available in the [FastVideo](https://github.com/hao-ai-lab/FastVideo) repository:
36
+ - [Finetuning script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/distill/v1_distill_dmd_wan_VSA.sh)
37
+ - [Inference script](https://github.com/hao-ai-lab/FastVideo/blob/main/scripts/inference/v1_inference_wan_dmd.sh)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38
  - Try it out on **FastVideo** — we support a wide range of GPUs from **H100** to **4090**, and also support **Mac** users!
39
 
40
  ### Training Infrastructure
 
42
  Training was conducted on **8 nodes with 64 H200 GPUs** in total, using a `global batch size = 64`.
43
  We enable `gradient checkpointing`, set `HSDP_shard_dim = 8`, `sequence_parallel_size = 4`, and use `learning rate = 1e-5`.
44
  We set **VSA attention sparsity** to 0.9, and training runs for **3000 steps (~52 hours)**
45
+ 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).
46
+
47
+
48
 
49
  If you use FastWan2.1-T2V-14B-480P-Diffusers model for your research, please cite our paper:
50
  ```