konakona666/krea-2

This repository is a fork of Krea AI's Krea 2 image generation model, adapted specifically for training custom character models quickly. Krea 2 itself is a model that generates images, and this fork focuses on teaching it to reproduce a specific character or person using a technique called LoRA training, combined with two training methods named SFT and DRaFT-K. To train a character, you install the training dependencies, point the tool at a Krea 2 checkpoint file, and provide a folder of images of the character you want to train, in formats like JPG, PNG, WebP, BMP, or AVIF. You can optionally give the character a trigger word to invoke it later. By default training runs 500 SFT update steps followed by 60 more focused DRaFT steps, and any missing face detection models needed for training are downloaded automatically. There are a few ways to supply the training data. The tool can automatically write captions for your images and generate its own set of training prompts, or you can supply your own captions as a CSV file or as a matching text file next to each image, along with your own list of prompts. This lets you train fully offline without relying on an external captioning service. Once a character model is trained, there is a script included to convert the resulting file into a format that works with ComfyUI, a popular tool for building image generation workflows. The code in this repository is released under the Apache 2.0 license, though the Krea 2 model weights themselves follow a separate Krea 2 Community License with its own terms for commercial use.