Krea 2 LoRA needs strength 1.8-2.5 to work — how to fix the training config

DeltaWaffleSyrup · reddit · 2026-09-20

A Reddit user training a Krea 2 body-type LoRA with AI Toolkit on Runpod (100-300 images, 8000 steps) finds it only works at strength 1.8-2.5. They shared their full config — LoKr rank 32/alpha 64, multi-resolution 512-1280, flowmatch, adamw8bit, lr 2e-4, bf16, qfloat8 quantization — and asked the community how to make it effective at lower strengths.

Original post →

More from Multimodal

Multimodal channel →