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.
What the 3-step LoRA is
TaoMate-H3 is described by its authors as a low-latency streaming audio-video generation runtime built on the open H3 video model with a 3-step LoRA. The LoRA is the piece that makes it fast: an adapter distilled so the base model can finish each small chunk in three denoising steps while generating picture and sound on one synchronised timeline.
It is developed by the Alibaba TaoLive AIGC team. The current public release is the step-3000 generator EMA adapter with rank 128 and alpha 128.
Key facts
| Developer | Alibaba TaoLive AIGC team |
|---|---|
| Type | LoRA adapter, rank 128, alpha 128 (step-3000 generator EMA) |
| Base model | The FL2VA checkpoint of the open H3 video model |
| Steps | Three denoising intervals per chunk |
| Official file | adapter_model.safetensors, about 2.48 GB |
| Resolutions in the official runtime | 480p, 768p and aligned 1080p, portrait or landscape |
| Duration in the official runtime | Multiples of five seconds, with an optional prompt per five-second block |
| Licence | The base model's Community License Agreement |
Released and planned variants
| Variant | Status when checked | README note |
|---|---|---|
| TaoMate-H3 T2AV | Released on Hugging Face | 3-step LoRA weights for T2AV |
| TaoMate-H3 FL2AV | Not released | Planned before 15 October 2026 |
| TaoMate-H3 Ref2AV | Not released | No date given |
A naming detail trips people up. Community files with fl2va in the name, such as the taomate_fl2va_3step_ema file in drbaph's turbo LoRA collection, refer to the H3 FL2VA base checkpoint that the adapter is applied to. The official install steps download that same FL2VA checkpoint. These files are not the unreleased streaming FL2AV model.
How much faster is it?
The README publishes one benchmark, measured on a single node with eight NVIDIA H20 96 GB GPUs, a 480×864 canvas and a 10-second output:
| Metric | TaoMate-H3 | Base model alone | Difference |
|---|---|---|---|
| Pure DiT time | 14.810 s | 169.572 s | 11.45× faster |
| First final chunk latent | 6.148 s | 170.052 s | 27.66× faster |
| First playable video | 17.287 s | 183.313 s | 10.60× faster |
| Peak DiT memory | 31.37 GiB | 32.03 GiB | About the same |
Two things to keep in mind. These are data-centre numbers from the authors, not consumer-GPU timings. And the speed-up does not come with a memory saving: peak memory is roughly unchanged.
For smaller cards, the only figures available are self-reported by one community converter (Asirus). With a pruned int8 base checkpoint, a 5-second clip at about one megapixel took 102.81 s with the TaoMate LoRA against 130.51 s with a 4-step turbo LoRA on 16 GB of VRAM, and 373.22 s against 474.30 s on 6 GB. That is about 21 percent less time than the 4-step option in that single test.
Limits worth knowing
- CFG stays at 1.0. The conversion cards set CFG to 1.0, and one warns that guidance above 1.0 burns the output of this distilled adapter.
- The LoRA is not the whole project. Chunked streaming, long-form continuity across prompt boundaries and multi-GPU parallelism are features of the official runtime code. A LoRA file loaded in ComfyUI gives you the 3-step acceleration only.
- The official runtime needs serious hardware. It targets Linux with NVIDIA Hopper-class GPUs and accepts 4 or 8 of them; 8 × H20 96 GB is the validated setup.
- 1080p is "aligned". The runtime renders a 1088-pixel edge, and the README suggests cropping to 1080 after inference.
Two ways to run it
The official streaming runtime
Install the Python package from GitHub, download the base model and let the runtime fetch the adapter. This is the only route to the streaming and long-form behaviour. The download page lists the links, files and requirements.
ComfyUI with a converted LoRA
Most single-GPU users take this route: a community-converted .safetensors file in ComfyUI/models/loras/, three sampling steps and CFG 1.0. The ComfyUI guide has the file choices and settings.
Licence
TaoMate-H3 is released under its base model's Community License Agreement, and the licence file ships in the repository. It sets conditions on territory, commercial use and attribution, and the community conversions state that they remain subject to the same terms. Read the full text before you redistribute the weights or ship a product on them. This page is a summary, not legal advice.
How this relates to the Taomate H3 studio
The studio on this site is a separate hosted service for people who want a short clip without a local setup. It offers text-to-video and image-to-video with fixed durations and resolutions, and it does not let you load the 3-step LoRA or change step counts. For background on the project names, see What is TaoMate?