KREA 2 LoRA training with AI Toolkit and ComfyUI

This tutorial covers KREA 2 LoRA training in AI Toolkit, followed by checkpoint comparisons in ComfyUI. The presenter describes local installation but demonstrates training on a rented RunPod GPU. His installer and comparison workflow are available through his Patreon; the description also links a manual installation guide.

Dataset preparation gets substantial attention. He recommends a small set of high-resolution images, explains matching image and caption filenames, and demonstrates automatic captioning. He advises adding the subject name or style token to every caption. ChatGPT ZIP uploads provide another captioning method in his demonstration.

For training, he recommends the raw model architecture over Turbo with a training adapter, based on his own results. He keeps rank at 32 and favors linear balanced timestep settings. He reports that layer offloading allows training with 10 GB VRAM and 64 GB system RAM, at the cost of speed. A 512-resolution setup reportedly reduces VRAM needs to 6 GB with the same RAM requirement, but he finds its image quality worse than training at 1024.

He disables training samples and evaluates saved checkpoints in ComfyUI with matching prompts and seeds. Most of his preferred results fall between 1,000 and 2,000 steps. Tests in another visual style help assess flexibility, while a final example explores stronger LoRA weights for a manga style.