LoRA

A small add-on that teaches a model a style, a character, an object or a kind of motion.

#In plain words

A LoRA (low-rank adaptation) is a small file, usually tens to a few hundred megabytes, trained on top of one base model. It does not replace the model; it nudges it. One LoRA teaches one thing: a drawing style, a person's look (with consent), a costume, a camera move, or speed (see Speed-up LoRAs).

A LoRA works only with the family it was trained for: an SDXL LoRA on an SDXL-based model, a Wan 2.2 14B LoRA on Wan 2.2 14B.

#Where you meet it in a graph

• Load LoRA (LoraLoader): model and clip in, changed MODEL and CLIP out; fields lora_name, strength_model, strength_clip. For SD 1.5 and SDXL, whose LoRAs often change the text encoder too.
• LoraLoaderModelOnly: model in, MODEL out, strength_model only. Common for Flux, Qwen-Image, Z-Image and Wan.
• It stands in the path of the model: between the loader and whatever uses the model (the sampler, or ModelSamplingSD3 in Wan graphs). With Load LoRA, the text nodes must take CLIP from the LoRA node, not from the checkpoint.
• Several at once: chain them, the output of one into the next. Order rarely matters.
• Two-model graphs: Wan 2.2 14B LoRAs often come as a high and a low file, one for each model.
• Files live in models/loras.

#Values and what they change

• strength_model 1 is full strength. If the LoRA overrides everything or breaks the picture, try 0.6 to 0.8. Above about 1.2 most LoRAs distort.
• strength_clip changes how the prompt is read; usually the same as strength_model.
• Several strong LoRAs fight each other; lower each.
• Trigger words: many LoRAs act only when the prompt contains the words they were trained with. Authors often write them into the file's metadata or on the download page.

#Typical mistakes

• A LoRA of another family: ComfyUI loads it with "lora key not loaded" warnings in its log and the picture does not change. It does not stop.
• No trigger words: the LoRA barely acts.
• <lora:name:1> in the prompt: does nothing in ComfyUI.
• The text nodes still fed from the checkpoint after a Load LoRA: the CLIP part of the LoRA is skipped.

#How the app helps

• The Models tab card of a LoRA shows its Family, its Trigger words (click to copy), the strength the author advises, and which of your checkpoints it fits.
• Check reports models of different families in one path and a prompt missing a LoRA's trigger words.
• Lessons: LoRA: add a style or a subject, Fast video in four steps, LoRA for pictures, LoRA for a character: face, body, clothes, pose. A lesson checks that the LoRA is made for the family of the checkpoint and that its trigger words are in the prompt.

#See also

• Speed-up LoRAs
• Why families do not mix
• Model families
• Prompt syntax and weights
• The model card