Where is the official TaoMate-H3 GitHub repository?

It is github.com/TaoLiveAIGC/TaoMate-H3, published by the Alibaba TaoLive AIGC team. It contains the streaming runtime, the README, the licence and example prompt files.

Is TaoMate-H3 on Hugging Face?

Yes. The official 3-step LoRA weights are at huggingface.co/TaoLiveAIGC/TaoMate-H3, and the README also links a ModelScope mirror. Other Hugging Face repositories with TaoMate-H3 in the name are community conversions for ComfyUI.

How large is the TaoMate-H3 download?

The official adapter_model.safetensors is about 2.48 GB, plus two small config files. The base model is a separate and much larger download that the README links to.

Can I run the official TaoMate-H3 runtime on one GPU?

Not according to the README. The --gpus option accepts 4 or 8, and the validated configuration is eight NVIDIA H20 96 GB GPUs on Linux. Single-GPU users typically load a converted LoRA in ComfyUI.

Has the TaoMate-H3 FL2AV model been released?

Not when we checked on October 4, 2026. The README lists TaoMate-H3 FL2AV as planned before 15 October 2026 and Ref2AV without a date. Only the T2AV LoRA weights were public.

Does taomateh3.pro host TaoMate-H3 downloads?

No. This site is an independent hosted video studio and does not distribute model weights. Use the official GitHub, Hugging Face or ModelScope pages.

TaoMate-H3 Download: GitHub, Hugging Face, ModelScope

About this page. Taomate H3 (this website) is an independent video-creation service. It is not run by, affiliated with or endorsed by the TaoLive AIGC team or any model vendor, and it does not host or distribute the TaoMate-H3 weights. The notes below summarise the public upstream documentation and link to it.

Official sources

Where	Link	What is there
GitHub	TaoLiveAIGC/TaoMate-H3	Streaming runtime code, README, licence, example prompt files
Hugging Face	TaoLiveAIGC/TaoMate-H3	The 3-step LoRA adapter weights
ModelScope	TaoLiveAIGC/TaoMate-H3	Mirror of the model, linked from the README
Parent project	TaoLiveAIGC/TaoMate	The original TaoMate digital-human model; see What is TaoMate?

Anything else carrying the TaoMate-H3 name, including this website, is not an official distribution point.

What is in the Hugging Face repository

File	Size	Purpose
`adapter_model.safetensors`	About 2.48 GB	The 3-step LoRA weights (rank 128, alpha 128, step-3000 generator EMA)
`adapter_config.json`	Under 1 KB	Adapter configuration
`config.json`	Under 1 KB	Configuration file shipped with the adapter
`LICENSE`, `NOTICE`, `README.md`	Small text files	Licence terms, attribution notice and documentation

The repository holds the adapter only. The base model, the open H3 video model, is downloaded separately; the README's model table links to it and its install section downloads the FL2VA folder.

Requirements for the official runtime

Linux
Python 3.10 or 3.11
NVIDIA Hopper/SM90 GPUs; 8 × H20 96 GB is the validated configuration
CUDA 12.8 and PyTorch 2.8
FFmpeg with H.264 and AAC support

If that is not the hardware you have, the practical route is a converted LoRA in ComfyUI. The ComfyUI guide lists the community files.

Install and download

After creating the Python environment described in the README (PyTorch 2.8 with CUDA 12.8, Triton, vLLM and FlashAttention), install the package:

git clone https://github.com/TaoLiveAIGC/TaoMate-H3.git
cd TaoMate-H3
pip install -e .

The runtime downloads the adapter to models/TaoMate-H3 on first use. To fetch it ahead of time:

hf download TaoLiveAIGC/TaoMate-H3 \
  --include "config.json" "adapter_config.json" "adapter_model.safetensors" \
  --local-dir models/TaoMate-H3

Running the official runtime

One command runs the whole pipeline and writes video.mp4 into the output directory. It starts its own local workers, so no separate launcher is needed.

python -m taomate_h3 \
  --model-root /path/to/base-model \
  --prompt "your prompt" \
  --duration 10 \
  --resolution 768x1376 \
  --gpus 8 \
  --output outputs/demo_10s

`--prompt` and `--prompt-json` cannot be combined.
Parameter	Meaning	Default
`--model-root`	Base model directory that contains `FL2VA/`	Required
`--adapter`	Local LoRA directory; omit to use the official adapter	`models/TaoMate-H3`
`--prompt`	One prompt reused for every five-second block	None
`--prompt-json`	JSON file with one prompt per five-second block	None
`--duration`	Total seconds; must be a multiple of 5	5, or inferred from the JSON
`--resolution`	Width × height; short edge 480, 768 or 1088; both edges divisible by 32	`768x1376`
`--gpus`	GPU count, 4 or 8	8
`--seed`	Request seed	8301
`--output`	New or empty output directory	Required

Common resolutions

For an exact 1080-pixel edge, the README suggests cropping the 1088-pixel edge after inference.
Format	Portrait	Landscape
480p	`480x864`	`864x480`
768p	`768x1376`	`1376x768`
1080p	`1088x1920`	`1920x1088`

Community ComfyUI conversions

These Hugging Face repositories repackage the official adapter for ComfyUI loaders. They are maintained by community members, not by the TaoLive AIGC team:

CZMartin22/TaoMate-H3-3step-ComfyUI: BF16, about 1.24 GB
Asirus/TaoMate_H3_3_Step_LoRA: BF16 and FP16 files, about 1.2 GB each
Robert1212star/TaoMate-H3-3Step-ComfyUI: full-size, about 2.48 GB

Settings and troubleshooting for these files are in the ComfyUI guide. For what the adapter does and how fast it is, see the 3-step LoRA page.

Licence

TaoMate-H3 is released under the H3 base model's Community License Agreement, and the full text is the LICENSE file in both official repositories. It includes conditions on territory, commercial use and attribution. Read it before redistributing the weights or building a product on them.

No suitable hardware?

If you only want a short video and not the model files, the Taomate H3 studio on this site generates clips in the browser from text or an opening image. It is an independent hosted service with fixed settings, not a download mirror and not the official runtime.

Where to download TaoMate-H3: GitHub, Hugging Face and ModelScope