TaoLiveAIGC/TaoMate-HD-preview

TaoMate-HD Preview

Progressive-Resolution Autoregressive Audio-Video Generation

TaoMate-HD generates speech and high-resolution video together in continuous chunks. Two low-resolution denoising evaluations establish coarse structure; one high-resolution evaluation refines spatial detail. Stage-local audiovisual history connects consecutive chunks and prompt windows.

This preview provides the inference LoRA for the MiniMax H3 FL2VA base model. The same adapter supports serial and accelerated inference on eight GPUs.

Developed by the Alibaba TaoLive AIGC Team. Powered by MiniMax H3.

• Paper
• Project page
• Inference code and installation
• MiniMax H3 base model

News

• 2026-11-03: First release of the inference code and preview LoRA weights. Read the paper on arXiv.

Download

pip install -U huggingface_hub

hf download MiniMaxAI/MiniMax-H3 \
  --include "model_index.json" "FL2VA/*" \
  --local-dir models/MiniMax-H3

hf download TaoLiveAIGC/TaoMate-HD-preview \
  --include "config.json" "adapter_model.safetensors" \
  --local-dir models/TaoMate-HD

The adapter consists of:

File	Contents
config.json	LoRA rank and alpha
adapter_model.safetensors	Inference LoRA weights

Inference

After installing the inference repository, run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python inference.py \
  --model models/MiniMax-H3 \
  --adapter models/TaoMate-HD \
  --prompts examples/serum.json \
  --duration 10 --resolution 768x1376 \
  --mode serial --gpus 8 --output outputs/serum

Use --mode accelerated for the eight-GPU pipeline. See the code repository for dependencies, parameters, and prompt formatting.

Citation

@misc{gan2026taomatehdprogressiveresolutionautoregressiveaudiovideo,
  title={TaoMate-HD: Progressive-Resolution Autoregressive Audio-Video Generation},
  author={Qijun Gan and Chenwei Zhang and Xinyang Li and Meiguang Jin and Junfeng Ma and Shen Qiu},
  year={2026},
  eprint={2611.03394},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2611.03394}
}

License and acknowledgments

This release follows the MiniMax H3 Community License Agreement. See NOTICE for attribution. The upstream license terms apply to the base model and this adapter.

We thank MiniMax for H3 and the open research community for work on autoregressive generation, distribution-matching distillation, and efficient attention.