Custom LORAs

Train custom LORAs simply by uploading sample images.

Steps to train a LORA Model

Name(Required)

Add name for your LORA model

Base Model

Choose the base model you want to use as a starting point for your lora. Usually it's best to just choose the default option of stable-diffusion-v1.5 but you can choose any one you feel would work best. Just because you use a model as a base model when training your lora, does not mean your lora will only work with that model.

FineTuning steps(Required)

The number of iterations the AI should take when training your model. Must be a number between 100 and 5000. A good default value is around 10 steps per training image.

Learning Rate

How fast the model will be trained. 0.000001 (1e-6) is a good default. Between 1e-7 and 1e-4.

Instance prompt

The instance prompt, are the words you need to use when telling an image model to use this lora. If you are training the model on an object, set this to "picture of ukj object". If you are training the model on an animal, set this to "picture of ukj animal". Add this prompt to your character's appearance, and the lora will always be used for selfies.

Image Cropping

Training images are automatically cropped to 512x512

Center Crop will crop the center of the image

Face Crop will attempt to locate a face in the photo and crop around it. If no face is found, uses Center Crop instead.

Captions(optional)

You can include a JSON file to provide a description, or keywords for each of your training images. This can greatly increase the success rate of trainings.

{
  "myimage0001.jpeg": "ukj soda bottle on a wooden table in warm afternoon lighting, pine trees in the background",
  "myimage0002.jpeg": "two ukj soda bottles next to a tropical waterfall",
  ...
}

Upload training Images

Upload 5 to 200 images of the subject you want train for. If you want accurate face results, many images of closeups to the face are needed.

If you want full body, many images of the subject's full body will be needed as well.

A good number of images is around 25 to 50.

Last updated