Qwen_Image_2.1 Ai-Toolkit Settings
retained text
I've tried almost all learning rates, optimizers, and timesteptypes now, and this combination created photorealistic loras for me. Try the combination, I'm really more than thrilled, with character loras I got perfect results with 1500-2000 steps in just 12 photos. It really couldn't be better, the likeness is 1:1. Loras with different things also work perfectly. The standard settings of the Ai toolkit are not really good; I've had almost nothing but bad experiences with them. Ema brings in even more quality and makes prompt follow better.
document fetched 2026-09-26 00:41 · canonical https://reddit.com/r/comfyui/comments/1wpqt2c/qwen_image_21_aitoolkit_settings