Smart LoRA Trainer v7.0 + Smart mixer
Ko-fi Supporter early build updated to v9.0
Roadmap next Z-Image turbo DoRa/Lora Trainer
Supporter version : https://ko-fi.com/madscientiststudio
Research-Grade Features:
Aspect Ratio Bucketing - No distortion, proper image handling
V-Prediction Support - For Pony/Illustrious models
DoRA - Weight-Decomposed LoRA for better quality
Noise Offset - Fixes gray/plastic skin (the SECRET SAUCE!)
Min-SNR Gamma - Focus on fine details
YOLO Masked Loss - Subject isolation
InsightFace Validation - Face similarity scoring
Correct Kohya Format - Works in ComfyUI/A1111
Description
FAQ
Comments (1)
So it works. I tried this version as well as the version made for Ko-Fi supporters. The smart mixer is a necessary tool, but it is only available to supporters. I think teaching is faster than with Koyha. I like it and I'm looking forward to seeing how it develops. Although I don't think the money will come together for the Z-image turbo part.
Nice work, test below.