The “3-step” name can suggest that loading TaoMate H3 makes the entire workflow run in three steps and always makes it faster. The video demonstrates a different use: generate with the original model first, then use a LoRA-enabled second sampling stage to enhance lighting and detail.
This guide accompanies the September 19 video and covers the settings, results and trade-offs of that staged approach.
Prepare the model, LoRA and matching workflow
You need a working H3 ComfyUI workflow and a TaoMate H3 LoRA compatible with its loader.
The video mentions a version extracted by Kijai and also uses a larger ComfyUI conversion. Similar filenames do not establish compatibility between different conversions and every H3 node.
The conversion author’s file is taomate_h3_3step_comfy.safetensors. Place it in ComfyUI/models/loras/ and select it in the LoRA loader. The download and node-project links are at the end.
1. Understand the two stages
The video uses this sequence:
Original H3 model
↓ First 4 steps: establish the image
Model with the TaoMate H3 LoRA
↓ Last 3 steps: enhance lighting and detail
Output
The total sampling process in this example is therefore 7 steps. It is a different configuration from the conversion’s description of generating the entire result in three steps.
The video also mentions a 9-step approach discussed by the community. To reproduce this particular example, start with the 4-step + 3-step configuration rather than combining numbers from different posts.
2. Configure the two samplers
The video uses Clownshark sampler-related nodes. Distinguish the total step count from the number of steps actually executed by each stage:
| Location | Setting in the video | Purpose |
|---|---|---|
| Total steps in both samplers | 7 in each | Keep the overall sampling schedule consistent |
| Steps executed by the first stage | 4 | Use the original model without the acceleration LoRA |
| Steps executed by the second stage | 3 | Refine with the TaoMate-enabled model |
| LoRA strength | Compare from the mentioned 0.7 | Check enhancement and changes in prompt adherence |
eta | The video mentions 0.5 | Check it together with the demonstrated sampler configuration |
The key is that the first stage does not load this acceleration LoRA; only the second stage uses it. Applying it to the entire model chain from the start is not the staged configuration described here.
There is no upscaling in the final three steps of the demonstrated run. You can add another upscaling stage to an existing workflow, but first check the version without it so you can tell which change produced which effect.
3. Text-to-video and multi-reference video
The video demonstrates both text-to-video and multi-reference video, and the author’s multi-reference test also runs. They use a similar staged approach, but their conditioning inputs differ. Do not insert reference material into a model branch that does not support it.
First verify that your base H3 workflow generates correctly, then replace the model and sampling settings in the second stage. This makes it easier to distinguish base-model or reference-input problems from LoRA compatibility problems.
4. Visual gains come with trade-offs
Lighting and detail
This is the aspect the author likes most. When used as a second-stage refiner, the LoRA produces stronger lighting, visual effects and detail in the demonstrated results.
Prompt adherence
Some enhanced results also follow the prompt less closely. The author tries lowering LoRA strength and making the prompts more specific.
Compare more than visual attractiveness: check whether the subject, action and camera requirements changed. Higher strength is not always better.
Blur during fast motion
Fast-moving scenes in the video still tend to blur. The author does not particularly recommend relying on this configuration alone for them. For intense action, compare moving detail with the same action prompt before deciding whether to keep the enhancement.
Runtime
The demonstrated RunningHub workflow takes about 12–13 minutes, roughly two minutes longer than the ordinary acceleration-LoRA second-pass workflow used for comparison.
It is therefore more useful to treat TaoMate here as a visual-quality trade-off. It improves the look in that test without making the full workflow faster. Recompare when changing the machine, resolution or duration.
Make a useful comparison
Keep a copy of the working original workflow and create another with staged sampling. Use the same subject, dimensions and duration in both. Compare lighting/detail, prompt adherence, motion blur and total runtime.
Decide whether those changes suit your goal before increasing LoRA strength or adding upscaling.