Alibaba's TaoMate-HD sharpens a stream from 480p to 768p

Alibaba's TaoLive AIGC team has released TaoMate-HD, an audio-video model that generates a talking digital human and its speech in one continuous stream. The team posted the paper on arXiv on November 2, 2026, then opened the inference code and preview weights on November 3.

TaoMate-HD builds each short chunk of video on a coarse grid and then refines it. Two denoising steps work at 480p to set the coarse structure and motion, and a final step sharpens the frame to 768p, with audio generated along the same path. The project page calls the method progressive resolution. Each stage keeps its own clean audio-video prediction as history for the next chunk, so the stream holds a character's face and voice across dialogue updates.

TaoMate-HD is a LoRA adapter for MiniMax H3, specifically the FL2VA base model. The paper lists six authors from Alibaba Group and Nanjing University: Qijun Gan, Chenwei Zhang, Xinyang Li, Meiguang Jin, Junfeng Ma and Shen Qiu.

The GitHub repository reports a 1.60x speedup over direct 768p generation at matched settings, and a further 1.92x from a stage-parallel deployment across eight GPUs. Continuation training cut character error rate from 30.75% to 19.57%, an 11.18 percentage-point gain. The model supports hour-long generation, and the project page shows four 60-minute streams at 1376x768.

The preview weights are open on Hugging Face as a 2.48 GB adapter_model.safetensors file and a config.json holding the LoRA rank and alpha. The release follows the MiniMax H3 Community License Agreement. The bundled example runs a 10-second clip at 768x1376, 24 FPS and 32 kHz stereo audio, with --mode serial --gpus 8; --mode accelerated overlaps the pipeline across eight GPUs.

For creators, a model that holds a digital anchor together for an hour is the piece live shopping and always-on channels have been missing, if the output holds up outside the team's own clips. The preview is an adapter, so running it means bringing the MiniMax H3 base model and eight GPUs.

Download the preview LoRA from Hugging Face and clone the inference repo at github.com/TaoLiveAIGC/TaoMate-HD to run it; the MiniMax H3 base model is a separate download.