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.
Before you start
- You need a working ComfyUI workflow for the H3 FL2VA base checkpoint first. The LoRA is an add-on to that workflow, not a standalone model. The official README's model table links to the base model.
- Do not load the official
adapter_model.safetensorsdirectly. It is published in PEFT format with key names ComfyUI loaders do not expect. Community converters have remapped the keys. - What you get in ComfyUI is 3-step acceleration. The streaming runtime described on the 3-step LoRA page is a separate Python program.
Pick a converted file
| Hugging Face repo | File | Size | Notes from the card |
|---|---|---|---|
| CZMartin22/ | TaoMate-H3-3step-ComfyUI.safetensors | BF16, about 1.24 GB | Alpha tensors embedded; reports 208 patches attaching without missing-key warnings |
| Asirus/ | taomate_h3_3step_MAXQUALITY_bf16.safetensorstaomate_h3_3step_LIGHT_fp16.safetensors | BF16 or FP16, about 1.2 GB each | Author recommends MAXQUALITY, especially at 10 GB VRAM and above |
| Robert1212star/ | taomate_h3_3step_comfy.safetensors | About 2.48 GB | Key-layout conversion only; no training, merging or pruning |
| drbaph's turbo LoRA collection | The taomate_fl2va_3step_ema file in its experimental/ folder | About 2.48 GB | Marked experimental; the card lists manual sigma values |
Install the LoRA
- Download one converted
.safetensorsfile from the table above. - Place it in
ComfyUI/models/loras/. - Restart ComfyUI if the file does not show up in your LoRA loader.
- In your H3 FL2VA workflow, add a
LoraLoaderModelOnly(orLoraLoader) node between the model loader and the sampler, and select the file. - Set the sampler to 3 steps and CFG 1.0, using the values in the next section.
- Run a short, low-resolution test before committing to a long render.
From a terminal, the BF16 conversion can be fetched straight into the right folder:
huggingface-cli download CZMartin22/TaoMate-H3-3step-ComfyUI \
TaoMate-H3-3step-ComfyUI.safetensors \
--local-dir ComfyUI/models/loras/Recommended settings
| Parameter | Value | Where it comes from |
|---|---|---|
| Steps | 3 | All conversion cards |
| CFG | 1.0 | CZMartin22 and Asirus cards; higher values burn the output |
| LoRA strength | 1.0 (0.9 to 1.0) | CZMartin22 and Asirus cards |
| Sampler | euler, or er_sde | CZMartin22 and Asirus cards |
| Scheduler | simple, or linear FlowMatch | CZMartin22 and Asirus cards |
| Video shift / audio shift | 8 to 12 / 3 | Asirus card |
| Manual sigmas, 3 steps | 1.0, 0.961165, 0.853333, 0.0 | drbaph card |
| Manual sigmas, 4 steps | 1.0, 0.970874, 0.907249, 0.640000, 0.0 | drbaph card |
The manual sigma lists are for workflows that feed a custom sigma schedule into the sampler instead of using a named scheduler. If your workflow uses the simple scheduler, start there and only switch to manual sigmas if the result looks off.
Speed and VRAM reports
The Asirus card includes timings for a 5-second clip at about one megapixel, using a pruned int8 FL2VA base checkpoint. They are one person's measurements on one prompt, so read them as a rough guide.
| VRAM | TaoMate 3-step (BF16) | TaoMate 3-step (FP16) | Turbo 4-step v1.2 |
|---|---|---|---|
| 6 GB | 373.22 s | 380.59 s | 474.30 s |
| 16 GB | 102.81 s | 104.84 s | 130.51 s |
Troubleshooting
- The LoRA is not in the dropdown. Confirm the file is in
ComfyUI/models/loras/and restart ComfyUI. - Missing-key warnings, or the LoRA has no visible effect. Key layouts differ between custom nodes and loaders. Try a different conversion from the table. With the CZMartin22 file, the console should report 208 patches attached.
- Burned or harsh output. Set CFG back to 1.0. If it is only over-sharpened, lower the LoRA strength slightly.
- Still slow. The LoRA cuts the number of steps, not the memory the base model needs. On small cards the base checkpoint, not the LoRA, decides whether a run fits.
If you would rather not run it locally
The Taomate H3 studio on this site makes short clips in the browser from a text prompt or an opening image. It is an independent hosted service with fixed settings, so it is not a way to run this ComfyUI workflow in the cloud, and its results will not match a local 3-step run. For the official files and runtime requirements, see the download page.