FastVideo-Minimax-FastH3-Preview-v0.2

A few-step (4-step) distillation preview of MiniMax-H3, the 33B dual-modality (video + audio) diffusion transformer — distilled with data-free DMD2 by the FastVideo team.

The base model samples with 50 denoising steps; this student walks a 4-step grid on the release's shift-12 rectified-flow schedule (12.5× fewer transformer evaluations), generating synchronized video and audio in one pipeline call.

Preview status (v0.2): step 2900 of a 4000-step run — the same run that produced v0.1 (step 1400), carried 1500 steps further. Sample quality is still maturing, most visibly on high-motion detail.

What's new since v0.1

• 1500 more distillation steps (1400 → 2900) on the same data-free DMD2 run: sharper still detail and steadier audio/video sync.
• Corrected sampling contract in the card. v0.1 documented num_inference_steps=4, which makes the scheduler build its own 4-point sigma grid — 3 forwards on native spacing, not the 4 trained jump points. Sample with the explicit trained ladder instead (see Usage). The same off-operating-point mismatch affected this run's in-training validation renders, so judge the student by fresh samples on the ladder below, not by earlier validation clips.
• Fixed the repo id in the usage snippet (v0.1's card had it doubled).

What's in the repo

Diffusers-format (modular pipeline) layout. Only the transformer/ weights differ from the base release — the distilled student, in bf16. All other components (Qwen3-VL text encoder, video/audio VAEs, tokenizer, processor, schedulers) are unmodified copies of the base release, included so the repo is self-contained.

The student was trained with block-sparse video attention (VSA, 64-token tiles, 90% sparsity) and carries its trained sparse-gate parameters (attn.to_gate_compress); it can be run dense (default) or with VSA for additional inference speedup.

Usage (FastVideo)

Sample on the trained ladder — [999, 749, 500, 250] on the shared 1000-step grid, one forward per entry, each scheduler applying its own shift:

from fastvideo import VideoGenerator

gen = VideoGenerator.from_pretrained(
    "FastVideo/FastVideo-Minimax-FastH3-Preview-v0.2",
    num_gpus=1,
    dmd_denoising_steps=[999, 749, 500, 250],
)
video = gen.generate_video(
    prompt="<your H3-format multimodal prompt>",
    guidance_scale=1.0,   # the base model is guidance-distilled
)

The ladder can also be set without touching code:

export FASTVIDEO_DMD_DENOISING_STEPS=999,749,500,250

To run the student under the sparse attention it was trained with, select the VSA-H3 backend and match both knobs of the training contract — sparsity alone, at the default 256-token tile, is a different operating point:

export FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3
# generator args: VSA_sparsity=0.9, VSA_tile_size=64

Prompts follow the MiniMax-H3 multimodal prompt format (integrated_multimodal_description: ... overall_soundscape: ...); see the base model card for the prompting guide.

Training summary

• Method: data-free DMD2 (distribution matching distillation) — student / frozen teacher / trained fake-score critic, backward-simulation rollout (the student walks its own 4-step sampling grid during training), x0-space critic regression, shifted score-time sampling matched to the dual video/audio noise clocks (shifts 12 / 3).
• Student grid: [999, 749, 500, 250] — 4 steps on the release sampler's shift-12 schedule.
• Attention: student trained with VSA block-sparse attention (64-token tiles, 90% video-tile sparsity); teacher and critic dense.
• Data: text prompts only (data-free) — ~258k prompts (VidProM-H3 + synthetic t2va prompt set); no video data used.
• Resolution: 768×1344, 124 frames (5s) with synchronized audio.
• Optimization: global batch 64, lr 1e-6 (student and critic), fp32 master weights, bf16 compute, 2900 of 4000 steps.
• Hardware: 32× NVIDIA GB200.

Limitations

• Preview checkpoint — quality below the base model's 50-step sampling, especially on fine motion and audio detail; improves with training.
• Inherits all content limitations and usage restrictions of the base model.
• The 4-step ladder is what the student was trained for; other step counts and other timestep grids are off-distribution.

License

Distributed under the MiniMax H3 Community License (see LICENSE), inherited from the base model. Review the license (including its territory and acceptable-use terms) before use or redistribution.

Notes

• The transformer_ref component (reference-conditioning variant) is not packaged here; its entry in modular_model_index.json points at the base MiniMaxAI/MiniMax-H3 repo and is fetched from there if used. This preview distills the text-to-video+audio path only.

Acknowledgements

We thank Nuva Lab 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.

We thank the NVIDIA FastGen team for the DMD2 framework and H3 reference experiment that helped us align the score clock, modality shifts, and backward simulation.

We also thank MiniMax for releasing H3-Base, and the vLLM project, NVIDIA, and MBZUAI for their continued sponsorship and support of FastVideo.