Load LoRA (Any)

This is the LoRA node most people reach for first, and for good reason: it patches both halves a LoRA is typically trained to touch - the diffusion model and the CLIP text encoder - so it's a genuine drop-in for almost any LoRA you download, whether it's a style, a character, or a concept. LoraLoader_Any is the pack's wrapper around that same node, plus the one optional input every loader here gets.

The inputs that matter

Five fields, most of which have sensible defaults:

• model and clip - the diffusion model and CLIP encoder the LoRA will be applied to. Both required.
• lora_name - the LoRA file, from models/loras.
• strength_model and strength_clip - how strongly to modify each. Both default to 1, range -100 to 100, step 0.01, and both can go negative to subtract a LoRA's effect rather than add it.

The default of 1.0 on both strengths is a starting guess, not a correct value - community consensus, repeated across essentially every LoRA troubleshooting thread, is that 1.0 is often too much. If output looks artifacted, deformed, or just "off" with a LoRA in the graph, try 0.5-0.8 before assuming anything else is wrong, and change one thing at a time with a fixed seed so you can actually tell what caused the difference.

LoRAs "can be linked together," per the node's own description - chain multiple LoraLoader_Any nodes to stack a style LoRA with a character LoRA with a detail LoRA, each node's MODEL/CLIP outputs feeding straight into the next one's inputs. Outputs here are exactly that pair: the modified MODEL for your sampler, the modified CLIP for your text encode.

The any input

The pack's usual trick: an optional any input that accepts anything and does nothing with the value passed in - it exists solely to make this node wait for whatever feeds it before applying the LoRA. ComfyUI's executor otherwise runs a node the instant its required inputs are ready, and since LoRA patching only depends on an already-loaded model and CLIP, it can happen earlier than you'd expect in a graph with multiple stacked LoRAs. Staging this behind whichever node should genuinely finish first gives you deliberate control over that order instead of leaving it to the scheduler.

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 LoraLoader class. LoRA files go in models/loras as always.

Where people get burned

Base model mismatch. A Flux LoRA won't run on SDXL, and even within the SDXL family, Pony/Illustrious/SDXL checkpoints share enough of the underlying architecture that a LoRA shows up as selectable everywhere without being cross-compatible with all of them. Check the LoRA's page for what it was actually trained against.

Trigger word missing. A lot of LoRAs need a specific word or phrase in the prompt before their effect shows up at all - check the model page if a LoRA seems to be doing nothing.

CLIP skip. Anime-leaning checkpoints - Illustrious, Pony, NoobAI - often expect CLIP skip 2 for correct results; in ComfyUI that's a "Clip Set Last Layer" node set to -2, separate from this loader but frequently the missing piece when an anime LoRA looks subtly wrong rather than absent.

Weight too high. Covered above, and still the single most common LoRA complaint - try 0.5-0.8 before anything else.

If you only need the model half patched and don't care about CLIP, the pack's sibling node, LoraLoaderModelOnly_Any, skips the CLIP patch and the clip input entirely.