MiniMax H3 Image hands-on: fast plastic-free T2I, but edits and consistency fail
Divineboob · reddit · 2026-11-11
A Reddit user tested the MiniMax H3 Image ComfyUI template workflow. Image editing largely ignores the prompt (base image vs. replacement reference), Refmod silently fails, and VNCCS pose control follows pose but mangles character consistency. Positives: T2I is notably faster than QI2.1 and MOSS on an RTX 3070 Ti at 16MP, with no plastic-skin artifacts and good world knowledge. They hope for working Refmod, native VNCCS LoRA support, and ref2va-style prompting for edits.
More from Multimodal
- Fable 5.2 one-shots a photorealistic Central Park night photo recreation with p5.js brushes — yuwen_lu_ · 2026-11-11
- CHIMERA: An AI Cinematic Short Film Dubbed "Nature's Darkest Creation" — adanperezplaza · 2026-11-11
- Fuzzy 3D toy characters showcase AI image generation's stylized rendering — OVolosin82152 · 2026-11-11
- Midjourney vs MiniMax H3 Image: photorealistic portrait face-off — miilesus · 2026-11-11
- WIP: 4DGS Scene + Subject Reconstruction Running in VR on Quest 3 — mickmumpitz · 2026-11-11
- MiniMax H3-Image lands on LMArena Battle Mode for blind testing — arena · 2026-11-11