LoRA Training Help: Outputs Mismatch Sample Images

Kryimsson · reddit · 2026-08-03

A novice user is experiencing issues while training a character LoRA. They created a dataset using Image-to-Image (I2I) for accurate likeness and trained it on Civitai using the Krea2 model.

The Issue: The 3 sample pictures from the final epoch looked exactly like the target face. However, when the LoRA is put into action, the generated faces look nothing like the samples, regardless of adjusting the weight to 0.7, 0.8, 0.9, or even 1.0.

The author is asking the community for advice on specific settings or whether local training is necessary.

Original post →

More from Multimodal

Multimodal channel →