A Krea2 LoRA training trick keeps face detail at 512px by adding crops

More_Bid_2197 · reddit · 2026-07-27

A LoRA training trick for Krea2: crop faces and keep detail at 512px

The author says they train Krea2 LoRAs faster by pairing a normal full photo with a cropped face image. That lets them use 512-resolution training without losing facial detail, because the face can stay large enough in the crop.

They note that similar tricks did not work well with older models, but Krea2 appears more resistant to overfitting. Example: for a beach photo, they use both the original and a face-only crop so the model can learn fine facial features without needing high-resolution training.

Original post →

More from Multimodal

Multimodal channel →