Fine-tuning Custom Models
      Back to home
      1. Scenario Knowledge Center
      2. Scenario Basics
      3. Fine-tuning Custom Models

      Training a Character LoRA

      Training a Character Model has never been easier!

      Step 1: Generate a Character

      Find a Prompt and a Model that creates a character you want to train and download it or simply copy it to your clipboard.

      image (1)-1

      Step 2: Ask Chatgpt to Describe the Character

      Attach or Paste the Image into the Chatgpt prompt field and and a prompt asking Chatgpt to describe the image. We used the simple prompt "Describe precisely the character in this image"

      image (2)-2

      Step 3: Refine the Description into a Prompt and Add it to your Original Prompt

      Craft a prompt from the description provided by Chatgpt. Use a combination of the original prompt used to generate the image on Scenario, and the description given by Chatgpt. You can easily ask Chatgpt to condense the description into 30 words separated by commas and to forgo describing the style or the background.

      image (3)-1

      image (4)-1

      image (5)-1

      Step 4: Generate a Large Set of Images and Curate a Dataset

      Generate a large pool of images using the prompt. Add some different descriptors of the set and setting, clothing styles or emotions of the character to ensure you have plenty of options available for the dataset.

      image (7)

      Curate a selection of images that show a consistent character.

      image (8)

      Step 5: Train Your Model

      Go to the Model tab and Create a New Model. Upload your Dataset. For Advanced users, select a unique token for your character and add it to your Image Captions. For a guide on hand captioning your dataset go here: Guide to Manually Captioning Your Dataset for LoRA Training on SDXL

      We have developed the following settings that work well, but please experiment and try new things! There are many ways to create a successful training and we are always discovering new ones. We would love to see your results and experiments (the good and the bad) and one of the most enriching experiences on Scenario is the amount of space there is to discover and create!

      Screenshot 2024-01-15 at 1.58.06 PM

      Step 6: Use LoRA Compositions to have your Character Inhabit Unlimited Styles and Worlds!!!

      With LoRA Compositions, you can now add your character model to any Custom Blend, easily inserting your character concept into a near endless array of settings, concepts, contexts, and styles. For more on LoRA Composition visit: Introducing LoRA Compositions

      GBx5Qb2acAA1DYh     GByaeOGWgAAbV0W

      GByfEXobcAAjKLr     GByfGX-aIAAR1t4

      Get even more Control using ControlNet!

      GB2H9SqbQAA8qHH

      GBz-rANboAAkuba