Nougan Lora Loader (Multi-Model) 📁

If you've ever built a graph that runs the same style LoRA stack across several checkpoints - an A/B comparison, a turbo-versus-base test, a style applied to two different bases - you've hit the wall of duplicating LoRA loaders until the graph is unreadable. NouganLoraLoaderMulti is the multi-model twin of the suite's NouganLoraLoader: one shared LoRA stack, applied to up to five models at once, all patched models coming out of one node.

How it works

It's the same comfy.sd.load_lora_for_models machinery as the single loader, with a fan-out. The lora_data JSON - built by the same themed stack editor with favourites, folder filters, toggles, and the randomizer - is applied to the primary model and to whatever you wire into model_2 through model_5. Each patched model exits through its own output. Secondary models are patched without a CLIP connection (the apply call passes None for clip on the extras), which is the right call: the text encoder only needs patching once, and the primary CLIP output covers that.

You also get a lora_stack output of type LORA_STACK on the primary, so stack-aware nodes downstream can still read what was applied.

The inputs that matter

• model - the primary model; the stack always applies and its result is output 1.
• lora_data - the shared stack, edited once.
• clip - optional, for the primary model's text encoder.
• model_2 … model_5 - up to four more models, same stack.

Outputs: MODEL, CLIP, lora_stack, plus MODEL 2–MODEL 5.

Installing

Search Nougan in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/Nougan_Nodes
# restart ComfyUI

No pip deps - everything is the pack plus ComfyUI's own stack.

Where people get burned

VRAM is the honest constraint. Five models, each holding the patched weights, is a lot of resident memory - on a 24GB card you'll realistically use two or three of the extra slots, and the node won't warn you. Also remember the stack is shared, which is the feature: there's no per-model tuning here, so if one model needs a different strength than another, this node isn't the tool (use separate single loaders). And the same ±10 strength clamp and randomizer behavior as the single loader apply - check the auto-roll toggle if repeats aren't re-rolling. For "same style, many bases" workflows it turns five loaders into one node, which is exactly why it exists.