zirk5/krea2-character-lora-recipe

Krea 2 Character LoRA - Training Recipe

This is an evolving end-to-end recipe for training a Krea 2 LoRA on a character.

Current Status

I'm in the middle of training my first LoRA. As of 7/21, I've done a trial training run on my RTX 3060 12GB followed by a proper training run on a RTX Pro 6000 WS 96GB via vast.ai. Results are good so far.

My goal here is to document and improve the process that I've found success with.

Some of the improvements on my mind at the moment:

Minimize Cloud GPU Time

My first cloud training run costed $1.25 for a $0.67/hr instance, during which the training itself only took about 15 minutes.

So, there was over 1 hour of just fiddling around to get the server setup + downloading the resulting model.

My ideal goal is to automate the whole process of renting a GPU, training on it and grabbing the result(s) by using VastAI's CLI, etc.

Dataset Tools

I've been using a custom/vibe-coded couple of PyQt apps for putting my dataset together.

I'd like to add these tools to this project and improve+unify them so that they're useful for others too.

This includes:

• Image Extractor: Point it to a folder with videos and scrub through them to grab 1:1 cropped frames.
• Captioner: Edit the .txt files next to images to manage captions.

Resources

This project is based on my reading, using and/or following the below informational sources:

• JahJedi/krea2-character-lora-recipe - The original repo for this recipe. Read the original README here.
• [VRAM 12GB] A Complete Guide to Creating Character LoRAs with the Image Generation AI "Krea2" (musubi-tuner) - I followed this first in order to do a test training run on my RTX 3060 12GB (48gb system RAM - not all of which was used)
• musubi-tuner dataset config - Info about setting up the dataset
• musubi-tuner Krea 2 doc
• Blog post: Krea 2 LoRA Training - Centered around KreaAI's training platform, but has some info about prompting/captioning.
• Reddit post: Training Krea 2 LoRA on RTX3060 12Gb: a slow, uncomfortable guide that actually works - Uses AI Toolkit instead of musubi-tuner. Has info about captions.

---

License

Recipe, scripts, configs, README, and preview images: CC BY 4.0. See LICENSE for details.

The trained LoRA file and training dataset referenced in this recipe are not part of this repository and are not licensed for distribution.