A LoRA (Low-Rank Adaptation) is a small add-on file that changes how an AI image model draws one thing, such as a face or a style, while the model itself stays untouched. You load it beside the checkpoint it was trained for and set its strength. Before using one, check that it matches your base model, that it is a safetensors file, and that its licence and the base model's licence both allow your use.
A LoRA (Low-Rank Adaptation) is a small add-on file that changes how an AI image model draws one thing, such as a face, a style or an outfit, without replacing the model itself. You load it beside the base model it was trained for, set its strength, and that model now produces the subject or look the LoRA was trained on.
Checked 13 October 2026 against the LoRA paper and Microsoft's reference repository; the Diffusers LoRA and DreamBooth guides; ComfyUI's LoRA tutorial, multiple LoRAs tutorial, models page, LoRA examples and node source; the AUTOMATIC1111 feature list; the safetensors repository; Hugging Face's licence documentation; and the model cards and licence files for SDXL, Stable Diffusion 3.5, FLUX.1 [dev], FLUX.2 [klein] 4B and Z-Image.
This page explains the term. If you already know what a LoRA is and want to make one, go to our LoRA training guide, which covers the dataset, captions, trainer settings and fixes. The short definition also lives in our glossary under LoRA.
What is a LoRA?
LoRA stands for Low-Rank Adaptation. The paper that introduced it, written for language models in 2021, describes a method that "freezes the pre-trained model weights and injects trainable rank decomposition matrices" into the network. Image tools adopted the idea. Hugging Face's Diffusers guide puts it in plainer words: LoRA "works by inserting a smaller number of new weights into the model and only these are trained", which makes training faster and produces smaller files that are easier to store and share.
For someone making images, that means two files. The first is the base model, for example a Stable Diffusion checkpoint. The second is the LoRA. ComfyUI's examples describe the relationship exactly: LoRAs are "patches applied on top of the main MODEL and the CLIP model".
- 01Base model
The checkpoint. It can make images alone.
- 02LoRA
The patch. It adds one face, style or concept.
- 03Strength
How hard the patch pulls on the model.
- 04Prompt
Your text, plus the trigger word if the LoRA has one.
- 05Image
The base model's output, bent toward what the LoRA learned.
LoRA vs checkpoint vs fine-tuning
All three change what a model draws. They differ in what is trained and in what ends up on disk.
| LoRA | Checkpoint | Full fine-tune such as DreamBooth | |
|---|---|---|---|
| What it is | A small add-on file of extra weights | The model itself. ComfyUI's docs say it usually contains MODEL (UNet), CLIP and VAE | A training method that leaves you with a new model |
| What training changes | Only the added weights. The original model stays frozen | Nothing. It is the starting point | The entire model, in the words of the Diffusers DreamBooth guide |
| Example file size | 49.6 MB: the example LoRA in Stability AI's SDXL repository | 6.94 GB: the SDXL 1.0 base file | A copy of the whole model |
| Works alone | No. It needs the base model it was trained for | Yes | Yes |
| Folder in ComfyUI | models/loras | models/checkpoints | models/checkpoints |
A LoRA is itself a form of fine-tuning, the lightweight kind. The LoRA paper calls full fine-tuning the approach "which retrains all model parameters", and the Diffusers guide describes DreamBooth as a technique that "updates the entire diffusion model by training on just a few images of a subject or style". A LoRA trains only the small added part.
One more term sits nearby. A ControlNet is also a helper file, but it does a different job: ComfyUI's models page lists a LoRA as an add-on "tuned for a style, character, or concept" and ControlNet as "extra guidance from edges, depth, pose". A LoRA decides who or what is drawn. A ControlNet decides where things go.
What LoRAs are used for
| Kind of LoRA | What it adds | Typical use for a persona |
|---|---|---|
| Character | One face or body, called up with a trigger word | A persona that has to look the same across many images |
| Style | A look: a film stock, an illustration style, a colour treatment | A feed that should read as one visual world |
| Concept | An object, outfit or pose the base model draws badly | A recurring prop or a signature outfit |
For an AI persona the character LoRA matters most, because a model left alone draws a new face each time. A LoRA is one of several ways to hold a face. InstantID does it from a single reference image with no training, and our guide to what an AI influencer is shows where the face fits in the whole job.
How to use a LoRA in ComfyUI
ComfyUI loads a LoRA with one node. The steps below follow the official LoRA tutorial.
- 1Match the base model
Check which model family the LoRA was trained on. It has to be the family of the checkpoint you load.
- 2Place the file
Put the LoRA in ComfyUI/models/loras. Checkpoints go in ComfyUI/models/checkpoints.
- 3Add a Load LoRA node
Connect model and clip from Load Checkpoint into it, then use its model and clip outputs for the rest of the graph.
- 4Choose the file
Select it in lora_name. Refresh or restart if a file you just copied is not listed.
- 5Set the two strengths
strength_model acts on the image model and strength_clip on the text encoder.
- 6Prompt and queue
Add the trigger word if the LoRA's page gives one, then generate and compare against the same seed without the LoRA.
The tutorial notes that every LoRA variant (it names Lycoris, loha, lokr and locon) is used in the same way. To apply two at once, chain Load LoRA nodes in series. Some workflows for newer models use the LoraLoaderModelOnly node instead, which takes only the model and one strength_model value; ComfyUI's FLUX.1 depth LoRA example is wired that way. What ComfyUI needs to run in the first place is on our ComfyUI statistics page.
In AUTOMATIC1111 a LoRA is called from the prompt with <lora:filename:multiplier>. Its wiki adds that a LoRA cannot be added to the negative prompt.
LoRA strength: what the number does
Strength scales how much of the patch is applied. An early Stable Diffusion implementation, cloneofsimo/lora, explains it in one line: setting the scale to 0 "is same as using the original model", and setting it to 1 is the same as using the fully fine-tuned model. In between you get a blend.
In ComfyUI the Load LoRA node has strength_model, which the tutorial says "affects how strongly the LoRA influences the model weights", and strength_clip, which does the same for the text encoder. Both default to 1.0 in the node source, and the source notes that the value can be negative. In AUTOMATIC1111 the multiplier is "a number, generally from 0 to 1".
No value is correct for every file. Fix the seed, generate at a few strengths, and keep the lowest one that still gives the face or look.
Where to get LoRAs, and what to check first
ComfyUI's docs name the usual sources for model files: Hugging Face, Civitai, or a project's page on GitHub. Two notes on those. Hugging Face groups LoRAs as adapters in a base model's model tree, which tells you the family a file belongs to. Civitai shows a notice that it has not been accessible to visitors in Australia since 16 March 2026, so confirm it loads where you are before you rely on it.
- The base model it was trained on is one you run: an SDXL LoRA will not work on FLUX
- The file ends in .safetensors, a format built to store weights safely
- The page gives a trigger word, or says none is needed
- The LoRA's own licence allows your use
- The base model's licence allows your use too
- It was not trained on a real person's face
The file format is a safety matter. The safetensors project describes itself as a format for storing tensors "safely (as opposed to pickle)", and lists the older pickle format as "Unsafe, runs arbitrary code".
LoRA licences: read two, not one
A LoRA carries its own licence, set by whoever trained it. Hugging Face shows it in the model card and tells users to "seek out and respect a project's license". A LoRA is also a derivative of its base model, so the base model's terms come with it. These are the terms the main families state today.
| Base model | Weights licence | What it says |
|---|---|---|
| SDXL 1.0 | CreativeML Open RAIL++-M | The licence text says derivative versions must include, at minimum, the same use-based restrictions |
| Stable Diffusion 3.5 | Stability AI Community License | Free for commercial use below $1M in total annual revenue; an Enterprise licence above that |
| FLUX.1 [dev] | FLUX.1 [dev] Non-Commercial License | The model card says generated outputs can be used commercially as described in that licence |
| FLUX.2 [klein] 4B | Apache 2.0 | The model card says the open weights are available for commercial use |
| Z-Image and Z-Image-Turbo | Apache 2.0 | Stated on both model cards |
We could not open Civitai's licence pages from Australia, so its per-model permission settings are not described here. Read them on the model's own page. The family you are most likely to meet first is covered in our SDXL guide.
Should you download a LoRA or train your own?
- Right for styles, lighting and general concepts
- Nothing to train: match the base model and load it
- You inherit the uploader's licence and dataset choices
- It cannot be your own persona's face
- Right for a character that belongs to you
- You control the images, captions and trigger word
- Needs a dataset, a trainer and test runs
- You know exactly whose face is in it
A character LoRA for your own persona has to be trained, because nobody else has your character. Before you spend a run, the free LoRA training steps calculator shows how images, repeats, epochs and batch size turn into total steps. Once a LoRA holds the face, img2img is the usual way to refine a finished image without losing it. The AI Influencers program takes the same character from a written description to a trained face and on to publishing.
What is a LoRA: FAQ
What is a LoRA in AI image generation?
A LoRA is a small add-on file that changes how an image model draws one thing, such as a face, a style or an outfit. The name stands for Low-Rank Adaptation. The base model stays frozen and a small set of extra weights is trained beside it, so the result is a compact file that you load together with the model it was trained for.
What is the difference between a LoRA and a checkpoint?
A checkpoint is the model itself. ComfyUI's docs say it usually contains three parts: the image model, the text encoder and the VAE. A LoRA is a patch applied on top of a checkpoint and cannot make an image alone. On Hugging Face the SDXL 1.0 base file is 6.94 GB, and the example LoRA in the same repository is 49.6 MB.
Is a LoRA the same as fine-tuning?
It is a lightweight kind of fine-tuning. The LoRA paper describes full fine-tuning as retraining all of a model's parameters. LoRA freezes those weights and trains a much smaller set of added ones. Hugging Face says this makes training faster and more memory-efficient, and produces smaller files that are easier to store and share.
How do I use a LoRA in ComfyUI?
Put the file in ComfyUI/models/loras, add a Load LoRA node after Load Checkpoint, and pick the file in lora_name. Connect the node's model and clip outputs to the rest of the graph in place of the checkpoint's. To use more than one, chain several Load LoRA nodes in series, which the official multiple LoRAs tutorial shows.
What LoRA strength should I use?
Start at the node default and test. ComfyUI's Load LoRA node has two values, strength_model and strength_clip, and both default to 1.0. In AUTOMATIC1111 the multiplier in <lora:filename:multiplier> is, in its wiki's words, a number generally from 0 to 1. Compare a few values on one fixed seed and keep the lowest that still gives the look.
Can I use a LoRA commercially?
Check two licences: the one on the LoRA's own page and the one on the base model it was trained on. SDXL's licence says derivative versions must carry at least its use restrictions. FLUX.1 [dev] weights are under a non-commercial licence, while FLUX.2 [klein] 4B and Z-Image are Apache 2.0. A LoRA does not remove the base model's terms.
Can I make a LoRA of a real person?
Not without that person's written permission. A character LoRA should be trained on a fictional adult persona that you created, or on yourself. Reproducing a real person's face without consent raises likeness and privacy law problems whatever tool you use, and downloading a LoRA someone else trained on a real person carries the same problem.
Free · no card needed
Start with the free Starter Kit.
Five working files, one from each program, open in a free account. Then one short lesson a day by email, on the program you read.
Now you know what the file does. The program shows how to build a persona with it.
AI Influencers, included in All Access, covers the character from a written description to a trained LoRA, then ControlNet for the body, photorealism, video and the disclosure rules for publishing. The other three programs, live coaching and the private community come with the same subscription.
Planning to train one?
Enter images, repeats, epochs and batch size and the free calculator returns steps per epoch and total steps, the number every trainer setting depends on.