A ComfyUI workflow that tests a LoRA at five strengths and three prompts in one queue, and saves a single grid image. Free, core nodes only: nothing to install.

• Rows are 3 prompts, columns are strength 0.0 / 0.4 / 0.7 / 1.0 / 1.3. Same seed in every cell, so the only thing that changes along a row is the strength.

• Keep one prompt that does NOT ask for what the LoRA teaches. That row shows whether it leaks into everything.

• Tested on ComfyUI 0.27.1 with an SDXL-family checkpoint: 15 images at 832x1216, 28 steps, about 7 minutes on an RTX 4060 with 8 GB.

How to use it

1. Load the JSON in ComfyUI.

2. Pick your checkpoint, and your LoRA in each of the five LoRA nodes (set one and copy it to the rest).

3. Type the trigger word in the TRIGGER node and edit the three prompts.

4. Queue. The last node saves the whole grid as one image.

Limits

• Strength 0.0 removes the LoRA weights but not the trigger word, which stays in the prompt. For a clean baseline, run once more with the trigger empty.

• Sampler settings (euler_ancestral, cfg 5, 28 steps) suit Illustrious-family checkpoints. Change them for other bases.

• Strength is applied to model and CLIP equally.

• 15 samplings per run. On a slower card, lower the steps or the size.

The example image is a watercolor style LoRA from Civitai, a third-party download, run at the five strengths.

If you use it, post your grid as an image on this model. I am curious which LoRAs hold up at 1.3 and which break at 0.7, and what you would change in the layout. Questions and bug reports are welcome in the comments.