Sign In
home
models
images
videos
posts
articles
bounties
tools
challenges
events
shop

LoRA Training Recommendation Guide/Tutorial (WIP Revamp)

9
urbanlegendwiki's Avatar
urbanlegendwiki
Apr 3, 2025
training guide
opinionlorawiptrainingsdxl
LoRA Training Recommendation Guide/Tutorial (WIP Revamp)

Hello! In this article, I will provide recommendations for LoRA training, and it's also a tutorial. I wish you good luck and have fun!

Note that this is subjective and opinional. This article may work for someone but not for another. If it doesn't work for you, feel free to tell me or follow another tutorial/advice.

Name and Category

On Civitai, there are three pickable categories; Character, Style, and Concept.

Where to Find Dataset

You can use art sites or boorus to gather dataset images, but these are not the only sites where you can gather dataset images! You can also gather images from most kinds of sites, like video sharing sites and even chat platforms.

Be careful of where you gather images from for your training dataset. Some sites like Cara or Fusion (by Devoted CG) prohibit images posted there to be used for AI and are also anti-AI.

Here is the list of lists of dataset image sources:

Dataset Image Recommendations

General

Character

  • Crop sexually explicit NSFW parts, like sex, paizuri, etc. However, it is a good idea to include nudes in dataset, to let AI know how the character's body looks like.

  • Train your LoRA on official art of the character, so it can generate the character in the accurate artstyle!

  • If you are going to use a photo of a character as a plushie for a dataset, remove the plush tag (unless that's what the character has, or the LoRA is meant to generate the character as a plushie).

  • If you have an image of multiple characters and wanna use it in your dataset, crop the parts where the other characters are, so that only a solo character is visible in the image. You can also edit out the other characters; this will also make only one character visible in the image.

Concept

  • If the LoRA is meant to produce a characteristic of a character, do not use an image with multiple characters (except for conjoined bodies, ambient wildlife, or background characters) in the training dataset. This is crucial so AI can understand that this is a characteristic LoRA, not a multiple character LoRA.

Style

  • When making a single style LoRA, don't use images with different artstyles for your dataset. It causes confusion and may cause unpredictable outputs, too. It's best to use the images for the dataset of another LoRA. (Reference)

9

Comments