Instant Reference LoRA Train

If Instant Reference LoRA is the combined train-and-apply node, this is just the train half. It takes your reference images, captions them, trains the short LoRA, and hands you a lora_stack to do something with yourself - it does not patch your model or clip. Same pipeline, same caching, no application step.

You'd reach for this over the combined node when you want to train once and apply several ways: use the LoRA on a different checkpoint, at a different strength, or stack it with other LoRAs before applying. It's the "train/apply split" pattern the pack's README shows a workflow example for - train here, apply downstream with Instant Reference LoRA Apply.

Inputs

• model and clip - from a checkpoint loader, same requirement as the combined node. The node needs the real checkpoint path to train against.
• images - your reference batch. 5–15 clean images, same guidance as before.
• profile - the sdxl / anima enum. anima adds a vae socket and expects the qwen3-based clip.
• tagging_options / train_options (optional) - plug in the two helper nodes if you want control over captioning or training knobs.

Notice there's no model_strength or clip_strength here. The stack entry it builds is fixed at 1.0/1.0; you control strength when you apply it.

Outputs

Two: lora_stack (the stack containing the trained LoRA path, for wiring into an apply node or further chaining) and tags (a string of the generated captions - handy for a quick sanity check on what the tagger saw).

Caching

It behaves exactly like the combined node: a hash of images, checkpoint, profile, captions, and options determines whether it retrains or reuses a cached result from models/loras/instant-reference-generated/. Change anything in that fingerprint and you get a fresh 50-step training run. Set force_retrain in a Reference Train Options node if you want to insist on retraining anyway.

Install

ComfyUI Manager → search "Instant Reference" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/cstria0106/comfyui-instant-reference

As with the main node, the first actual training run does the heavy lifting - cloning sd-scripts into the node's runtime dir, building its own venv (Windows wants Python 3.12), and downloading the WD14 tagger. Expect that first run to take minutes.

Gotchas

Same family as the combined node: 50 steps means a rough adaptation, not a polished character model; the model input must come from a checkpoint loader; and only the two shipped profiles exist, so don't feed a Flux or Z-Image checkpoint in. The pack's author calls it "fairly rough" and untested beyond simple workflows - treat the first training run as a trial run, not a promise.