Testing Krea2 Image Training: How LoRA and Lokr Parameters Affect Quality

bhanvadia · reddit · 2026-08-06

A developer tested various LoRA and Lokr training configurations on the Krea2 model using an identical dataset of 38 character images.

Running approximately 3,000 steps across five different setups, the author compared variables like resolution (512px to 1536px), optimizers, and file sizes (27MB to 372MB). The experiment revealed counter-intuitive results, such as a 1536px Lokr factor 4 performing worse than a 512px Lokr factor 8, offering valuable insights for custom image training.

Original post →

More from Multimodal

Multimodal channel →