Shishir1996/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview

Veda-MiniMax-H3-T2VA (Preview)

Paper · Project Page · Code · Turbo LoRA · Deployment Guide

Introduction

Veda is a learned sparse-attention method for video diffusion models. Attention dominates the inference cost of video diffusion, yet only a small fraction of it contributes meaningfully to the output. Veda trains a lightweight predictor, distilled from the full model, to identify the most important 10% of attention and skip the rest.

This repository provides the Veda predictor for MiniMax-H3 text-to-audio-video generation.

Highlights

• Faster inference. Up to 6.8× attention speedup and 3.1× end-to-end speedup, with larger gains on longer videos.
• Preserved quality. Visual and audio quality on par with full attention while skipping 90% of it.
• Plug-and-play. Veda is not a LoRA: it leaves the backbone weights and style untouched and only changes how attention is computed. It is compatible with any LoRA and with fine-tuned MiniMax-H3 variants.

This checkpoint is trained and evaluated with MiniMax-H3-Turbo-Lora for 8-step generation.

Samples

Comparison between full attention (left) and Veda (right) under the same prompt, seed and Turbo LoRA. All videos are 16:9, 14.4 s, generated on a single RTX PRO 6000 Blackwell GPU. Generation time is shown in each title bar.

Raw files are available under media/.

Performance

Speedup per denoising step relative to full attention (MiniMax-H3 + Turbo LoRA, 10% attention kept).

GPU	Clip	Attention speedup	End-to-end speedup
RTX PRO 6000 Blackwell	16:9 · 14.4 s	6.79×	3.12×
RTX 4090	16:9 · 5.17 s	4.75×	1.77×
RTX 4090	16:9 · 10.1 s	6.22×	2.42×
RTX 4090	16:9 · 14.4 s	6.31×	2.82×

Usage

Requires the Miowtion codebase. See AGENTS.md for hardware requirements and per-architecture notes.

git clone https://github.com/veda-sparse/Miowtion.git && cd Miowtion
pip install -e '.[gpu,encode]'

hf download MiniMaxAI/MiniMax-H3 --local-dir weights/MiniMax-H3
hf download Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview \
  --local-dir weights/veda/h3-t2va-8nfe-600

Encode the prompt, then generate:

echo '{"id": "demo", "task": "t2va", "prompt": "<structured T2VA prompt>"}' > prompts.jsonl
python scripts/encode_samples.py --root weights/MiniMax-H3 \
  --manifest prompts.jsonl --out artifacts/samples/demo

python scripts/generate.py \
  --root weights/MiniMax-H3 --variant FL2VA \
  --schedule turbo --num-steps 8 \
  --adapter weights/turbo_lora/<8-step-lora>.safetensors \
  --sample-cache artifacts/samples/demo --sample-id demo \
  --geometry 16:9@37 --attention veda \
  --predictor weights/veda/h3-t2va-8nfe-600/minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors \
  --out-dir artifacts/generate/demo

Use --attention dense veda to render both modes with a side-by-side video. On 24 GB GPUs, add --offload-blocks 50 --mlp-chunk-rows 4096.

Option	Description
--geometry	<aspect>@<latent_t>, one of the 12 packed plans
--keep-ratio	Override the default 0.1 (trained at 0.1)
--dense-steps	Denoising steps to keep dense
--offload-blocks	Number of transformer blocks streamed from host memory
--mlp-chunk-rows	MLP chunk size; reduce if out of memory

Loading the predictor directly:

from miowtion.veda import bundle

loaded = bundle.load(
    'minimax_h3_t2va_veda_8nfe_600step_preview_fp8.safetensors', device='cuda')
loaded.predictor          # TileScorePredictor
loaded.plans.select(geo)  # tile plan for a geometry
loaded.keep_ratio         # 0.1

License

This model inherits the MiniMax H3 Community License from its base model.

Citation

@inproceedings{han2026veda,
  title={Veda: Scalable Video Diffusion via Distilled Sparse Attention},
  author={Han, Shihao and Yang, Hao and Hu, Xinting and Mei, Xiaofeng
          and Jiang, Yi and Qi, Xiaojuan},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}