Introduction

September’s TaoMate-H3 already taught MiniMax H3 to stream synced speech and picture in chunks. On Nov 3, Alibaba’s TaoLive AIGC team followed up with TaoMate-HD — progressive-resolution autoregressive audio-video generation: coarse 480p structure first, then a 768p refine, with audio riding the same trajectory.

This is a new preview LoRA + inference repo + paper, not a rehash of the Sept streaming note. If you already bookmarked TaoMate-H3, treat HD as the high-res streaming sibling.

What shipped

  • Paper: arXiv:2611.03394 — TaoMate-HD abstract frames high-res streaming digital humans and the speech-interrupt problem when text keeps arriving.
  • Weights: TaoLiveAIGC/TaoMate-HD-preview — adapter_model.safetensors (~2.48 GB) + tiny config.json LoRA rank/alpha. News line on the card: 2026-11-03 first release of inference code and preview LoRA.
  • Code: TaoLiveAIGC/TaoMate-HD (created Nov 2; ★1 at scout) — install docs, serial vs accelerated multi-GPU path.
  • Base: still MiniMax H3 FL2VA (MiniMaxAI/MiniMax-H3); you download the foundation separately, then drop the HD adapter on top.
  • Project page: taoliveaigc.github.io/TaoMate-HD

How the progressive pass works

Per short chunk:

  • Two low-resolution denoising evaluations (~480p) lock coarse video structure.
  • One high-resolution evaluation (~768p) refines spatial detail.
  • Audio is generated along the same denoising trajectory.
  • Each stage keeps its own clean audiovisual predictions as stage-local history for the next chunk — so long streams can continue across prompt windows without a full reset.

Optional accelerated mode overlaps LOW0 / LOW16 / HIGH / decode across eight GPUs. Serial and accelerated share the same adapter and prompt format.

Numbers that stuck

On the team’s 100 long sequences (1–10 minutes), the README table (no markdown tables in our CMS — bullets instead) highlights:

  • TaoMate-HD serial: temporal 0.913, visual realism 5.00, audio aesthetics 4.407, CER 19.57%, ~4.68 FPS on 8× H20.
  • TaoMate-HD accelerated: same temporal 0.913, CER 21.28%, ~8.99 FPS.
  • Progressive resolution alone: about 1.60× vs direct 768p under matched BF16/SDPA; optimized deploy another ~1.92× over serial Progressive.
  • Continuation training cut CER from 30.75% → 19.57% (~11.18 pp).

Older single-GPU TaoMate still wins raw FPS in their comparison row — HD trades some speed for the progressive quality / CER jump. Pick the sibling that matches your latency vs fidelity brief.

How to try it today

Open weights + code, but this is a lab GPU story (multi-H20 class), not a one-click Comfy node yet:

conda create -n taomate-hd python=3.10 -y && conda activate taomate-hd
git clone https://github.com/TaoLiveAIGC/TaoMate-HD.git
cd TaoMate-HD
pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128
pip install -e .
sudo apt-get install ffmpeg   # Debian/Ubuntu

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

Then run their inference.py with eight visible devices for serial or --mode accelerated after installing SageAttention / FlashAttention-3 / the documented W8A8 stack.

Original Source

https://huggingface.co/TaoLiveAIGC/TaoMate-HD-preview

Conclusion

If September’s TaoMate got you hooked on chunked H3 speech+picture, TaoMate-HD is the Nov 3 upgrade path: progressive 480→768, a real preview adapter, and receipts for long-form CER. Bring a serious GPU rack — then stream.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=taomate-hd-progressive-minimax-h3-av-streams

—Aurelia ♡