MiniMax H3 Image tested: six characters into one 16 MP shot in ComfyUI, blind-tied with Qwen and FLUX Klein
AuzBoss · reddit · 2026-11-21
A Redditor spent a week testing open-weights MiniMax H3 Image (33B, 6 steps, no CFG) in ComfyUI on a 5090, compositing six game characters from six photos into one 16 MP couch photo in a single pass.
Key findings
- Multi-person compositing: name each photo, specify left-to-right positions, and state each person keeps their own face/hair/outfit. Six held up; nine drifted.
- Blind test vs Qwen Image 2.1 and FLUX 3 Klein (10 edits, 2 seeds): Qwen 32, Klein 32, H3 31 out of 40 — roughly even. At 1 MP, H3 came last (25) and Klein first (34); size matters a lot.
- Speed: 40s per 16 MP edit, 24s with the 2-step Fast LoRA; Klein 4s and Qwen 19s but at 1-2 MP.
- Failure modes: at 1 MP removing a car moved the whole scene; at 16 MP same seed only 2% moved; small faces and fine textures go soft.
Two habits fix most bad edits: say who stays ("keep the other two people" took success from 1/12 to 12/12) and run it big. The workflow (Generate + Edit v1.2) is on Civitai, needs ComfyUI 0.45+ and the AusBoss node pack. Note: MiniMax's license excludes the US, EU, UK and South Korea.
More from Multimodal
- Stripe residency dev draws all friends in JavaScript with GPT 6.2 Astra, results stun — gabrielchua · 2026-11-22
- Fable 5.2 plus WebGL demoed as a tool to "create anything" — hewarsaber · 2026-11-22
- Cinematic UFO folklore music video built with Suno and SLOPCANNON — DarkroomSignal · 2026-11-22
- Photographer crafts AI-generated digital muse Nous Girl on faux magazine cover — alexcovo_eth · 2026-11-22
- Turn any Google Maps location into a cinematic drone shot with AI — aziz4ai · 2026-11-22
- Open-source world model PAN 2 wows with GTA-like physics in driving demos — LinusEkenstam · 2026-11-22