LoRA Test — Identity Loader

LoRATestIdentityLoader is the plain-jane member of the LoRA-Lens family, and that's the point. Where the "LoRA Lab · Identity / Baseline Loader" exists for the dashboard to inject into runs it builds, this one is for the hand-built comparison graph - the old-school way of doing what the LoRA Lens dashboard automates. You wire one of these per candidate LoRA, connect them to your sampler, and run the same prompts and seeds through each. It's a one-job node: take a model, apply one identity LoRA, hand it back.

How it works is essentially identical to its Lab sibling. Pick None in lora_name (or set strength_model to 0) and the model passes through untouched - that's your no-LoRA baseline, the control row every honest test needs. Pick any actual file from the dropdown and it loads the .safetensors from ComfyUI/models/loras, caching by path plus modification time so repeated runs don't reload it, and patches only the model - load_lora_for_models(model, None, ...) - never the CLIP. Given modern identity LoRAs don't train the text encoder anyway (a post-Flux shift documented in the KB's lora-training.md), "model-only" is correct behavior, not a shortcut.

Inputs that matter:

• lora_name - dropdown of everything in ComfyUI/models/loras. The "None" entry is the baseline.
• strength_model - default 1.0, −4 to 4. Negative strengths invert the LoRA, handy for seeing what it actually contributes.
• model - your checkpoint's model.

Output: model, ready for the sampler.

The real question is which loader to reach for. If you're running the current LoRA Lens dashboard, use that. If you're building a small graph by hand - a couple of LoRAs, one prompt, one seed - this node plus its cousins (LoRAPromptQueueControllerV4, LoRATestGridCollector) is the manual path. The principle is the same either way: fix every variable except the candidate LoRA, and keep a baseline row so you know what the base model would have produced alone.

Gotchas. The dropdown is snapshotted when the node is instantiated - drop a new LoRA into models/loras and it won't appear until you recreate the node or restart ComfyUI. The classic "LoRA not found" FileNotFoundError means the file isn't in models/loras or the name in the node no longer matches a file on disk. And remember this node is a model patch, not a magic bullet: with strength at 0 it does nothing at all, which is the correct way to run your baseline.

Installation is the shared pack story: ComfyUI Manager (search "ComfyUI LoRA Lens") or git clone https://github.com/astropuzzo/ComfyUI-LoRA-Lens.git into custom_nodes, install requirements.txt with ComfyUI's own Python (insightface, onnxruntime-gpu, and friends - the CUDA-sensitive ones), then restart. First startup kicks off a background download of the face-recognition weights. Install into system Python instead of ComfyUI's venv or python_embeded and you'll be debugging import errors before you ever touch a LoRA.