Hunting for a mask + reference image inpainting workflow that doesn't sacrifice quality

NeoHisto · reddit · 2026-09-08

The author is looking for high-quality inpainting that combines a mask on a base image with separate reference images. Flux.2's Reference Latent+ node technically works but quality falls short; GPT Image 2 nails quality but can't take a mask and a reference image simultaneously. A workaround — cropping the region, inpainting it, pasting back — breaks seamlessness because the model freely regenerates the whole crop, leaving lighting and positions slightly off. They're asking for a node or workflow that handles both without the tradeoff.

Original post →

More from Multimodal

Multimodal channel →