What is a LoRA in Stable Diffusion?

By Updated 3 min read

A LoRA (Low-Rank Adaptation) is a small file, from a few megabytes to a few hundred, that adds a style, a character, a product or a concept to an existing Stable Diffusion model without retraining it. You load it next to the base model, give it a weight, and can combine several at once.

How a LoRA works

Fine-tuning a whole model means rewriting hundreds of millions to billions of weights and producing a 2 to 7 GB file for SD 1.5 or SDXL. A LoRA freezes the model and trains only pairs of small matrices, mostly next to the attention layers; the product of each pair approximates the change. The result captures "what is new" in a small file and is applied at generation time, with a multiplier. Ten to thirty images are often enough to train one on a consumer GPU, in minutes to a few hours depending on the base model.

How to use a LoRA

  1. Download a LoRA made for your base model: an SD 1.5 LoRA does not work with SDXL, and vice versa; SD 3.5 has its own.
  2. Put the file in Automatic1111's models/Lora folder (or its equivalent in ComfyUI, Forge or Fooocus).
  3. In Automatic1111, Forge and Fooocus, call it in the prompt with <lora:filename:0.8>; in ComfyUI, add a Load LoRA node instead. The number is the weight; 0.6 to 1.0 is the usual range.
  4. Add the trigger word given by the author if there is one (a character name, a style token).
  5. Adjust: too strong and every image looks the same, too weak and nothing changes.

Combining LoRAs

Several LoRAs can be stacked: a style at 0.7, a character at 0.8, a detail enhancer at 0.4. Lower each weight as you add more, and fix the seed while you tune them, otherwise you cannot tell what each one does. LoRAs trained on the same concept fight each other; pick one.

LoRA, embeddings, checkpoints: which to choose?

TypeSizeBest forUsed how
Checkpoint (fine-tune)2–7 GB (SD 1.5, SDXL)a whole aesthetic (realistic, anime)replaces the base model
LoRAa few MB to a few hundred MBa style, character, object, pose<lora:name:weight> in the prompt, or a Load LoRA node in ComfyUI
Textual inversion (embedding)a few KB to 100 KBa single token, often a negativethe token name in the prompt

A common setup is one good checkpoint, with LoRAs swapped per image.

Where to find LoRAs

Civitai and Hugging Face host tens of thousands, listed with sample images, trigger words and the base model they need. Check the licence: some forbid commercial use, and Civitai has removed LoRAs of real people since May 2025. Our artist style comparison is a good alternative for styles: a name in the prompt often does what a style LoRA would, with nothing to download.

Try it now, freeCompare 1,731 artist styles without any LoRA →

Questions people also ask

Does a LoRA slow down generation?

Barely: Automatic1111 and ComfyUI merge the LoRA into the model's weights before sampling, so each step costs the same. Only that merge, when you load or change a LoRA, takes a moment.

Can I use LoRAs on this site?

Not your own. The tools run fixed models, to which the site already adds small detail and noise-offset LoRAs. Members' txt2img pages may expose specific LoRAs in the future.

What does the number in <lora:name:0.8> mean?

The strength with which the LoRA's changes are applied: 1.0 is what the author trained, 0.5 is half the effect, values above 1.2 usually distort the image.

All guides