FLUX.3 hands-on in ComfyUI: strong image editing, weak hands and complex scenes
cgpixel23 · reddit · 2026-10-18
A hands-on ComfyUI test of the new FLUX.3 model compares Base (CFG 3, 25 steps) with Turbo (CFG 1, 16 steps). Generation handles textures, artistic styles, wildlife, and natural environments well, but struggles with hands, feet, realistic skin, and complex scenes. Surprisingly, image editing is one of its strongest features — Turbo delivers good style transfer, outfit changes, background replacement, weather changes, and object removal. The author also demonstrates the BFL team's fixed workflow for rescuing low-quality outputs. Full tutorial and workflow on Civitai.
More from Multimodal
- MOSS-Image-SFT-4-steps v2 lands: legible text at 4 steps and 1.4s — Illustrious_Row_9971 · 2026-10-19
- The Hobbit Benchmark: can AI fix the trilogy's infamous glowy look? — Capable_Parsnip_156 · 2026-10-19
- Dev builds K-pop virtual group music video overnight with Claude Opus 5.5 and Hedra MCP — AIandDesign · 2026-10-19
- Creator shares AI vertical-video workflow using HeyGen, HyperFrames Studio and Opus — toolstelegraph · 2026-10-19
- Redditor shares AI-generated 3-minute music video, asks if 480p quality is good enough — jponlyf · 2026-10-19
- AI makeup tutorial goes hilariously wrong: extra limbs turn bathroom into a Resident Evil boss fight — alifcoder · 2026-10-19