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.
The short answer
TaoMate is an open-source project for generating video and audio together, in real time and over long durations. It is built by the TaoLive AIGC team at Alibaba's Taobao & Tmall Group. Two releases share the name:
- TaoMate (July 2026) is a real-time digital-human model for long-form audio-video generation.
- TaoMate-H3 (September 2026) is a low-latency streaming runtime with a 3-step LoRA, built on the open H3 video model.
If you arrived here after searching for a LoRA, a ComfyUI workflow or "3 step", you are looking for TaoMate-H3. The 3-step LoRA page covers it in detail.
TaoMate vs TaoMate-H3 at a glance
| TaoMate | TaoMate-H3 | |
|---|---|---|
| What it is | Real-time digital-human model for long-form audio-video generation | Low-latency streaming audio-video runtime with a 3-step LoRA |
| First public | Paper posted to arXiv on 27 July 2026 | Announced in the project news on 8 September 2026 |
| Base model | LTX-2.3 (22B dev checkpoint) with the Gemma 3 12B IT text encoder | The FL2VA checkpoint of the open H3 video model |
| What you download | A full model.pt checkpoint | A rank-128 LoRA adapter, adapter_model.safetensors |
| Validated hardware | 72 GB GPUs: one or two for batch inference, four for the interactive demo | One node with 8 × NVIDIA H20 96 GB; the runtime accepts 4 or 8 GPUs |
| Output | 512×768 reference configuration; up to 12 prompt segments, about one minute per case | 480p, 768p and aligned 1080p; duration in multiples of five seconds |
| Licence on the model card | apache-2.0 | The base model's Community License Agreement |
What the original TaoMate does
The paper behind the project is titled TaoMate: Anchor-Guided Memory Bridging Evolving and Reference States for Real-Time Audio-Video Digital Human Generation. The problem it tackles is specific to long videos of one person: the face, the voice and the lip sync have to stay consistent across many consecutive segments. A short cache keeps recent motion but forgets what the subject looked like earlier, while attending to the whole history gets slow and lets errors pile up.
TaoMate's answer is a persistent memory. It keeps one fixed visual anchor of the subject, compresses the finished video and audio blocks into a memory of constant size, and reads from that memory while generating the next block. Because this memory is separate from the denoising of each block, blocks can also be processed in parallel stages, which the authors use to speed up generation.
The repository ships the inference runtime, multi-GPU launchers and an interactive browser demo. The demo uses four GPUs: two for the generator, one for text conditioning and decoding, and one for a local dialogue model with speech recognition. In the README's own measurement, a 60-second 512×768 case peaked at 66.8 GiB of memory on a single GPU.
What changed with TaoMate-H3
In September 2026 the team moved the streaming approach onto the open H3 video model and published it as a separate repository. Their release note says the new version improves both audio and video quality.
TaoMate-H3 is not a full retrained checkpoint. It is runtime code plus a LoRA adapter distilled so that each small chunk needs three denoising steps. At the time of our check, the text-to-audio-video (T2AV) weights were public. The README lists a streaming FL2AV model as planned before 15 October 2026 and a Ref2AV model without a date.
Which one do you need?
- A talking digital human or the interactive avatar demo: the original TaoMate repository.
- Faster few-step generation in ComfyUI: a converted TaoMate-H3 LoRA. See the ComfyUI setup guide.
- The official files and install steps: the download and source links.
- A short clip without any hardware: the hosted Taomate H3 studio on this site, which is a separate service described below.
Where this website fits
Taomate H3 at taomateh3.pro is an independent, hosted video studio. You write a scene or upload an opening frame, review the credit quote, and get a short clip back in the browser. It is not the TaoMate project, it is not operated by the TaoLive AIGC team, and it does not expose LoRA, step or sampler settings. If you want to study or modify the open-source models, use the official repositories linked under Sources.