Load LoRA

Here's the thing nobody tells you when they hand you a LoRA: most of the time, the only node you need for it is this one. Load LoRA (class LoraLoaderModelOnly) takes your diffusion model, your LoRA file, and a strength, and returns a modified MODEL. No CLIP input, no CLIP output. Just the model patch. It's the version the community converged on because modern architectures mostly don't have a CLIP side to patch - Flux LoRAs are model-only, and so are the vast majority of things you'll actually download.

How it works

Three inputs:

• model - your diffusion model wire, from Load Diffusion Model, Load Checkpoint, or another LoRA node.
• lora_name - the file, from ComfyUI/models/loras.
• strength_model - how hard to apply it. Default 1.0, range −100 to 100.

The node loads the LoRA file, patches the model's weights, and gives you back a new MODEL. Crucially, it's chainable: the output is itself a valid model input, so you can stack a character LoRA, then a style LoRA, then a detail LoRA in series - the node description calls this out explicitly. Each link applies on top of the previous one's output.

The mechanism is worth one paragraph because it explains every LoRA gotcha you'll ever hit. A LoRA is a small set of low-rank matrices trained against a specific base architecture. Loading it means adding those matrices into the frozen checkpoint's weights. It is a patch, not a model - a LoRA file generates nothing on its own. And it's architecture-bound: an SDXL LoRA does nothing on Flux, and even Pony/Illustrious/SDXL LoRAs cross over unreliably despite sharing code.

The strength dial and the modern reversal

The old advice - "start at 0.5–0.8, 1.0 overpowers the prompt" - is SDXL-era and it does not generalize. On Z-Image people run character LoRAs at 1.0 to 1.5 because dropping the weight loses likeness. The right move is to read the model page and set what it publishes, then tune from there with a fixed seed. Change one thing at a time; randomizing the seed while you tune tells you nothing.

Common issues

The "my LoRA does nothing" checklist, in order:

1. Wrong base model. The #1 cause, by a mile. Check which architecture the LoRA page names.
2. Missing trigger word. Many LoRAs need their trained trigger in the prompt. The page will say so.
3. Strength too low - or, on newer models, too high. See above.
4. Stacking order. If stacking produces black images or NaNs, reorder with the larger LoRAs first.
5. CLIP skip on anime LoRAs. Pony/Illustrious-era LoRAs expect CLIP skip 2; in ComfyUI that's the Clip Set Last Layer node set to −2, applied before text encode.

If you came from a workflow that uses the old Load LoRA (Model and CLIP) node instead - the one with strength_clip - that's the pre-modern version for SD/SDXL LoRAs that patch the text encoder too. This node is the same idea minus the CLIP side. For 2026 models, that's a feature, not a loss. It ships with ComfyUI core; drop files in models/loras and go.