Veda-MiniMax-H3 R2VA (Preview)

Demo · Project Page · Code · Paper · Deployment Guide

Introduction

Veda is a learned sparse-attention method for video diffusion models. This checkpoint brings Veda to MiniMax-H3 reference-to-audio-video generation. It selects visual reference and generated video tiles independently; text, VLM conditioning, and audio attention remain dense. The predictor is not a LoRA and leaves the backbone weights unchanged.

This preview was fine-tuned for 600 updates from the T2VA predictor on 5.17-second R2VA clips. It uses the MiniMax-H3 Turbo LoRA for eight-step generation, with up to 32 visual-reference tiles and 32 current-video tiles per query tile and head. The predictor is stored in FP8 and dequantized to BF16 when loaded.

Highlights

• Learned sparse attention. Visual reference and generated-video attention are selected independently; text, VLM conditioning, and audio stay dense.
• Fixed visual budgets. The released preview uses 32 reference tiles and 32 current-video tiles per query tile and head.
• Plug-and-play predictor. The FP8 predictor is separate from the H3 backbone and Turbo LoRA weights.

Samples

The examples below use the same seed, prompt, geometry, and Turbo LoRA for Dense and Veda. Each generated output is 1344×768, 124 frames, and about 5.17 seconds. Dense (left) and Veda (right) are stitched into one synchronized comparison video, with attention and denoising speedups in its top title bar. Each comparison uses the original Dense audio track.

Performance

Measured with eight-step Turbo LoRA inference on one RTX PRO 6000 Blackwell GPU. Veda uses fixed budgets of 32 reference and 32 current-video tiles.

GPU	Clip	Attention speedup	End-to-end speedup
RTX PRO 6000 Blackwell	16:9 · 5.17 s	2.42×	1.69×
RTX PRO 6000 Blackwell	9:16 · 5.17 s	2.65×	1.67×
RTX PRO 6000 Blackwell	1:1 · 5.17 s	2.06×	1.37×
RTX PRO 6000 Blackwell	4:3 · 5.17 s	2.19×	1.54×
RTX PRO 6000 Blackwell	16:9 · 10.1 s	2.84×	2.05×
RTX PRO 6000 Blackwell	16:9 · 14.4 s	2.90×	2.20×

Usage

See AGENTS.md for installation and GPU-specific deployment notes. The commands below use the R2VA code on the main branch.

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

hf download MiniMaxAI/MiniMax-H3 --include "model_index.json" "Ref2VA/*" \
  --local-dir weights/MiniMax-H3
hf download larryvrh/MiniMax-H3-Turbo-Lora \
  minimax_h3_turbo_v4_step600_ema.safetensors --local-dir weights/turbo_lora
hf download Veda-Sparse/Minimax-H3-R2VA-Veda-Preview \
  --local-dir weights/veda/h3-r2va-preview

Prepare the reference files, then encode and generate:

python scripts/prepare_official_r2va_demo.py \
  --out artifacts/examples/minimax_h3_ref2va
python scripts/encode_samples.py --root weights/MiniMax-H3 \
  --manifest artifacts/examples/minimax_h3_ref2va/manifests/prompts.jsonl \
  --out artifacts/samples/demo
python scripts/generate.py --config configs/infer_r2va_preview.yaml \
  --predictor weights/veda/h3-r2va-preview/minimax_h3_r2va_veda_preview_fp8.safetensors \
  --attention dense veda

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}
}