README.md
3.4 KB · 85 lines · markdown Raw
1 ---
2 license: other
3 license_name: minimax-h3-community
4 license_link: LICENSE
5 base_model: MiniMaxAI/MiniMax-H3
6 library_name: diffusers
7 pipeline_tag: text-to-video
8 tags:
9 - text-to-video
10 - video
11 - audio
12 - text-to-audio-video
13 - distillation
14 - dmd2
15 - few-step
16 - minimax-h3
17 - fastvideo
18 - fasth3
19 - preview
20 ---
21
22 <p align="center">
23 <a href="https://github.com/hao-ai-lab/FastVideo"><img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logos/logo.svg" width="320" alt="FastVideo"></a>
24 </p>
25
26 # FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
27
28 The recommended FastH3 Preview v1 checkpoint from
29 [FastVideo](https://github.com/hao-ai-lab/FastVideo). It generates synchronized
30 video and audio from text with four transformer forwards. This step-1300 model
31 was trained with data-free DMD2 and VSA-H3 at 90% sparsity.
32
33 [Blog](https://haoailab.com/blogs/fasth3-preview/) ·
34 [Matching LoRA](https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA/tree/main/vsa-datafree) ·
35 [FastH3 collection](https://huggingface.co/collections/FastVideo/fastvideo-fasth3)
36
37 > This checkpoint requires FastVideo's VSA-H3 attention backend. Use the
38 > matching LoRA above if you prefer to download only the distilled adapter.
39
40 ## Run with FastVideo
41
42 Install [uv](https://docs.astral.sh/uv/getting-started/installation/), then use
43 the CUDA 13 / Blackwell path below. It selects FastVideo's published CUDA
44 kernel wheel instead of compiling the kernel locally. See the
45 [installation guide](https://hao-ai-lab.github.io/FastVideo/getting_started/installation/)
46 for other platforms.
47
48 ```bash
49 git clone https://github.com/hao-ai-lab/FastVideo.git
50 cd FastVideo
51 uv venv --python 3.12 --seed
52 source .venv/bin/activate
53 UV_TORCH_BACKEND=cu130 uv pip install \
54 --no-sources-package fastvideo-kernel \
55 -e ".[fasth3]"
56 ```
57
58 ```bash
59 python examples/inference/basic/basic_fasth3.py \
60 --model-path FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree \
61 --prompt "your prompt" \
62 --no-warmup \
63 --repeats 1
64 ```
65
66 The tested defaults use four B200 GPUs and the trained four-forward schedule.
67 On other multi-GPU CUDA systems, follow the installation guide and add
68 `--no-replicated-dit --vsa-kernel triton --no-fa4`. The GPU count must divide
69 H3's 56 attention heads.
70
71 ## Scope
72
73 This preview supports text-to-audio-video generation. FL2VA and Ref2VA were
74 not distilled. Difficult motion, fine detail, and some audio may remain below
75 the base MiniMax H3 model. This checkpoint inherits the
76 [MiniMax H3 Community License](LICENSE).
77
78 ## Acknowledgements
79
80 We thank [Nuva Lab](https://nuvalab.ai/) for bringing production grounding to FastH3 through its experience with real-world creative video-agent workloads. Its production-aligned post-training insights help bridge open-source research to practical data-assisted distillation for commercial video workflows, with Omni Ref as the next focus.
81
82 We thank the [NVIDIA FastGen](https://github.com/NVlabs/FastGen) team for the [DMD2](https://arxiv.org/abs/2405.14867) framework and H3 reference experiment that helped us align the score clock, modality shifts, and backward simulation.
83
84 We also thank [MiniMax](https://huggingface.co/MiniMaxAI/MiniMax-H3) for releasing H3-Base, and the [vLLM project](https://vllm.ai/), [NVIDIA](https://www.nvidia.com/en-us/), and [MBZUAI](https://mbzuai.ac.ae/) for their continued sponsorship and support of [FastVideo](https://github.com/hao-ai-lab/FastVideo).
85