Personal Insights on Training a Single Multi-Character LoRA for Direct Multi-Character Generation
Preventing Feature Bleeding in Multi-Character Outputs
To prevent feature bleeding when generating multi-character images using a single LoRA, you must include a sufficient number of multi-character images in your dataset (composed of different characters from the dataset). While you can stitch single-character images together as a quick fix, it is highly recommended to collect genuine multi-character interaction images. Relying solely on stitched images often leads to overfitting, resulting in outputs with rigid side-by-side splits or poorly integrated compositions. For niche characters with scarce source material, you can use tools like NanoBanana or GPT-image-2 to generate 2-panel or 4-panel split grid compositions, which helps lower costs and improve efficiency.
Dataset Sizing and Repeat Configurations
Typically, 10 to 50 multi-character images are enough, and you don't need to cover every single character permutation. However, if you have a massive dataset of several thousand images (like mine), aim for around 200 multi-character images. It is crucial to configure distinct repeat counts for the multi-character subset and each single-character folder. I recommend setting the repeat count for the multi-character dataset to at least 3. During training, use a trainer that supports sample generation previews. Test with multi-character prompts to monitor the process; if character features begin to bleed, increase the repeat count for the multi-character subset. The repeats for single-character subsets can be adjusted based on their individual learning difficulty.
Avoiding Style Overfitting
To prevent style overfitting, you should introduce images of varied art styles for each character outfit or variant. For smaller datasets, adding two alternative art styles—with a volume equal to or about half of the original images—should suffice. However, for large-scale datasets where all characters share the same original art style, you will need a broader variety of external art styles to effectively decouple the characters from the specific style.
Reference Example
Here is a 90+ character all-in-one LoRA along with its dataset that I trained. Feel free to use it as a reference (though the quality is modest):
https://huggingface.co/kongbai-84/soultide_lora_Anima
What models are you using for tagging/making captions, when you stitch single-character images together?
What models are you using for tagging/making captions, when you stitch single-character images together?
I had an AI write a script for me that stitches images with a single click while automatically generating caption content for the combined image based on the original images' .txt files.
The specific rules are: depending on how many images are selected to stitch, it prepends "X girls" at the very beginning; then, based on the stitching sequence, it prepends "left / middle / right" to each original image's tags before merging them into the final caption.
You can ask GPT or another model to generate a script for you following these rules.
What models are you using for tagging/making captions, when you stitch single-character images together?
I had an AI write a script for me that stitches images with a single click while automatically generating caption content for the combined image based on the original images' .txt files.
I'm making captions on natural language, so I think script doesn't work, but ok I thing with 2-6 characters loras it will working normal too with caption model that I'm using
What models are you using for tagging/making captions, when you stitch single-character images together?
I had an AI write a script for me that stitches images with a single click while automatically generating caption content for the combined image based on the original images' .txt files.
I'm making captions on natural language, so I think script doesn't work, but ok I thing with 2-6 characters loras it will working normal too with caption model that I'm using
Natural language can work as well. The purpose of stitching multiple images together is to help the LoRA avoid feature bleeding when generating multi-person images, so it doesn't require extremely precise captions. If you want higher-quality captioning, you can use Gemini-3.5-flash—it's fast, cheap, and produces great results.