Flux-3 image edit tested in ComfyUI: great details, faces still break
EstablishmentFun6629 · reddit · 2026-10-19
A Redditor tested Flux-3 image editing in ComfyUI v0.41.3 with several challenges (first image = input, second = output):
- Reflective dress change: editing clothes while keeping them consistent in reflections
- Mirror challenge: consistency of mirrored content after edits
- Environment swap and camera angle change: subject preservation across backgrounds and viewpoints
Verdict: details are good, faces are cooked. Reflection consistency has long been a weak spot for edit models, so these tests are a useful reference for comparing Flux-3's performance.
More from Multimodal
- Nano Banana 'disposable camera' prompt produces strikingly filmic portraits — michaelrabone · 2026-10-19
- Three new MOSS-Image finetunes drop: 1024px deblur tune and two 8-step DMD distills, Apache-2.0 — linoy_tsaban · 2026-10-19
- Midjourney SREF share: style code 969064757 yields screen-printed surreal poster look — tisch_eins · 2026-10-19
- Running an Ollama + Flux.1 text-to-image workflow on a 12GB 5070 Ti — RYuu-1120 · 2026-10-19
- Moss LoRA training in AI-Toolkit: default rank 32 looks bad, rank 8 lacks detail — Lucaspittol · 2026-10-19
- Benj Edwards turns voice memos into funk with Suno: AI ushers in personal media — binarybits · 2026-10-19