Moss image edit methods compared: a 6,000-step LoRA makes all the difference

Sixhaunt · reddit · 2026-10-22

The author compared Moss Image Edit's two reference-input methods and trained a rank-16 LoRA (6,000 steps) with Ostris's AI toolkit:

Simple ref mode: feeds reference images directly; without a LoRA it often mashes inputs together or fails to edit properly; 15s per render.
Ostris edit mode: injects references more thoroughly, but without a LoRA produces an overlaid blend of the images; 45s per render and 3x longer training.
With method-specific LoRAs (same dataset, same steps) both improve markedly — the LoRA versions transfer the hat far more faithfully in the side-by-side grid.
Next: larger dataset, a bigger LoKr and at least 2x training time.

More from Multimodal