🦊 RS LoRA Tester

The classic LoRA-testing workflow is an act of self-harm: wire one LoRA, queue, check the image, disconnect, wire the next, queue again - 15 minutes of clicking for 20 candidates. RS LoRA Tester automates the part that was never interesting. Load a list of LoRAs into this node, arm batch mode, and one queue press spawns one generation per LoRA, each with exactly that LoRA applied. The results come out and you compare. Same idea as rgthree's Power Lora Loader for the stacking case, but this one's built specifically for "which of these should I even use?"

How it works

The node keeps its roster in lora_data (a JSON string the UI manages - you never hand-edit it) and applies LoRAs to the model input and, optionally, the clip input. The inputs are familiar territory:

• model - your base model.
• use_clip - the CLIP ON/OFF toggle. Flick it off and the node runs "model only," which is the right call for LoRAs that don't train the text encoder.
• lora_data - the managed JSON list.
• clip - optional CLIP input.

Outputs are MODEL and CLIP, wired to your sampler exactly like any LoraLoader.

The magic is in the interface. Add LoRAs from a searchable folder tree, reorder with the drag handles, and toggle any row off without deleting it (disabled rows are ignored by the workflow). Strengths get + / - buttons with ±0.05 precision or direct typing. Then hit ⚙️ LoRA Batch and press ComfyUI's Queue - one session per LoRA is created automatically.

The two ways to use it

Batch of different LoRAs - the headline mode. Select ten candidates, arm the batch, queue, and walk away. Each session applies exactly one LoRA, so every output is a clean A/B test against the same base prompt.

Same LoRA at different strengths - arguably more useful once you've narrowed it down. Add the same LoRA twice (duplicates are allowed), set one at 0.5 and one at 0.8, queue, and see the strength curve in one batch. This is how you find out that your character LoRA is best at 0.6, not the 1.0 the download page suggested.

A couple of quality-of-life details: the list persists per node across sessions, the ✔️ Update LoRA list button picks up new files in loras/ without a restart, and the batch button shows a running state while sessions are queued, then auto-resets when the queue empties.

Common issues

• LoRA produces nothing visible - base-model mismatch or a missing trigger word, the same two failure modes that hit any LoRA loader. Test with the exact base model the LoRA was trained for and include the trigger word from its model page.
• Batch count looks wrong - disabled rows are skipped, which is a feature, but it means "arm the batch" reflects only the enabled rows. Check your toggles.
• Anime LoRAs look washed out - many Illustrious/Pony LoRAs want CLIP skip 2; that's a graph setting, not this node's fault.

Pair it with RS Save Image LoRA and the output filenames will name themselves - which is exactly what you want when twenty results come out of one batch.

Installing it

Part of the RaykoStudio pack. ComfyUI Manager → search ComfyUI_RaykoStudio → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Raykosan/ComfyUI_RaykoStudio

Restart ComfyUI. Dependencies (pycairo, opencv-python, scipy, Pillow>=10) come via Manager; manual clones need pip install -r requirements.txt. Apache 2.0, tested on ComfyUI 0.15+, Python 3.10+, Torch 2.8+.