Flux Klein consistency LoRA holds faces but distorts bodies across viewing angles

rolens184 · reddit · 2026-09-10

The author has been using Flux Klein to generate LoRA training datasets from one or more character photos, plus a consistency LoRA. Faces stay consistent, but bodies drift: side views flatten the breasts and rear views impose a standard hip/thigh shape that doesn't match the source. They ask for fixes, noting step counts up to 8 are rumored to help.

Original post →

More from Multimodal

Multimodal channel →