Nodes/ComfyUI Instant Reference/Instant Reference LoRA Load
ComfyUI Node

Instant Reference LoRA Load

Reuse a trained reference LoRA without retraining

By cstria0106·Created 6 months ago·Updated 5 months ago· 8
Instant Reference LoRA Load
  • model
  • clip
  • model
  • clip
  • lora_path
  • lora_stack
lora_path
model_strength1.00
clip_strength1.00

The quiet one in the pack. Instant Reference LoRA and Instant Reference LoRA Train will happily retrain for you, but their generated files land in models/loras/instant-reference-generated/ - and this node is how you grab one of those later without firing up another training run. Give it a path, set strengths, get patched model and clip back, plus the LoRA as a stack for further chaining.

The README doesn't even list it among the headline nodes, which tells you it's a utility rather than a workflow centerpiece. But once you've trained something you like, this is the node that turns it into a normal, reusable asset instead of a one-off graph artifact.

Inputs

  • lora_path - a string. The absolute path to a .safetensors LoRA, or you can paste the lora_path output from a training node. That's the honest use case: save the path, come back next session, load it here.
  • model_strength / clip_strength - how hard the LoRA is applied. Defaults to 1.0 each, range ±10.
  • model / clip (optional) - if you supply them, the node applies the LoRA and returns patched versions. If you don't, it still returns the lora_path and lora_stack, and the model/clip outputs just pass through whatever was given.

Outputs

model, clip (patched when inputs were provided), lora_path (resolved string path), and lora_stack (a single-entry stack you can hand to Instant Reference LoRA Apply or chain further).

Install

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

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

Gotchas

Since this node skips training entirely, it avoids the pack's heavy first-run setup - no sd-scripts clone, no venv build. The one real catch is the path: it expects a resolvable filesystem path, not a ComfyUI-relative LoRA name, so a bare mystyle.safetensors won't resolve the way it would in the built-in LoraLoader. Feed it the full path and you're fine. If you point it at a LoRA from another trainer that uses an exotic format, the in-node loader may balk - anything this pack produces, though, loads without drama.

CategoryInstant Reference

Inputs (5)

NameTypeDefaultDescription
lora_pathSTRING
model_strengthFLOAT1.00-10–10
clip_strengthFLOAT1.00-10–10
modeloptMODEL
clipoptCLIP

Outputs (4)

NameTypeDescription
modelMODEL
clipCLIP
lora_pathSTRING
lora_stackLORA_STACK