Load LoRA for Model Only (Any)

LoRAs are trained deltas that modify how a model behaves - usually both the diffusion model and the CLIP text encoder, since a LoRA can shift what "your prompt words" produce as well as how the model denoises. But plenty of LoRAs, especially motion LoRAs, acceleration/distillation LoRAs, and a fair share of pure style LoRAs, don't meaningfully touch the CLIP side. LoraLoaderModelOnly_Any patches the diffusion model and skips the CLIP patch entirely - faster, and one less thing to re-wire if you're chaining several LoRAs and don't want to keep threading CLIP through each one.

The inputs that matter

Three fields, two of which you'll actually touch:

• model - the diffusion model to patch. Required.
• lora_name - the LoRA file, from your models/loras folder.
• strength_model - how strongly to apply it. Default 1, range -100 to 100, step 0.01.

Community consensus, cutting across basically every LoRA discussion in the wild: the default of 1.0 is often too strong. Try 0.5–0.8 as your starting point rather than assuming 1.0 is correct - a LoRA at full strength is a common cause of the "why does this look artifacted/overcooked" question. Going negative is legal (down to -100) and occasionally useful for subtracting a LoRA's effect, but most people never leave the 0–1.5 range.

Output is a single MODEL - "the modified diffusion model," per the node's own tooltip - straight into your sampler, or into the next LoRA node if you're stacking.

The any input

Every node in this pack adds one optional input beyond its stock counterpart: any, which accepts literally anything and does nothing with the value it receives. It's purely there to create a dependency - ComfyUI's executor runs a node the moment its required inputs are satisfied, and a loader typically has nothing blocking it, so without this trick, every model in your graph tends to load at once regardless of when it's actually needed. Wire something into any and this LoRA patch waits for that source before it runs, letting you decide the actual loading order by hand.

Installing it

ComfyUI Manager: search ComfyUI-LoaderUtils, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/lrzjason/ComfyUI-LoaderUtils

No extra dependencies - it wraps ComfyUI's own LoraLoaderModelOnly class. LoRA files go in models/loras, same as always.

Where people get burned

Base model mismatch. A Flux LoRA won't run on SDXL, and even within the SDXL-derived family, Pony, Illustrious, and SDXL checkpoints share enough architecture that a LoRA will show up as selectable across all of them without actually being cross-compatible. If output looks wrong in a way that isn't just strength, check that the LoRA was trained for the checkpoint you're running it on.

Trigger word missing. Many LoRAs need a specific word or phrase in the prompt to activate their effect - check the model's page for it before assuming the LoRA itself isn't working.

Weight too high. As above - 1.0 is a starting guess, not a correct default. If you're seeing artifacts or deformation, drop strength_model to 0.5–0.8 and compare with a fixed seed before concluding the LoRA is broken.

Since this node skips CLIP entirely, CLIP-skip-related issues (a common troubleshooting step for anime-leaning checkpoints) don't apply here - if you need that, you want the full LoraLoader_Any instead.