MiniMax H3's official images look stunning—why can't ComfyUI users reproduce them?
cptdurr · reddit · 2026-11-16
A Reddit user posts a detailed complaint: MiniMax H3's official GitHub repo shows 50+ spectacular sample images, but the same prompts in ComfyUI workflows produce clearly worse results—side-by-side comparisons show artifacts on a beetle image, wrong framing on a grass group shot, a badly held bottle in the cow-feeding image, a completely different wizard composition, a mangled lion face, and broken fire physics.
The poster has tried multiple samplers, resolutions, and seeds and suspects MiniMax used different sampling/inference settings than the public workflows, asking anyone who understands the underlying tech to help reproduce the official quality.
More from Multimodal
- New image model nails painterly illustration style so well it looks hand-drawn — gizakdag · 2026-11-16
- Bicycle glissando: an unusual MIDI editor built in Antigravity with Gemini 4 Carbon — alexanderchen · 2026-11-16
- Synthesia launches Syren, an AI agent that turns one prompt into on-brand video — synthesiaIO · 2026-11-16
- Open Vision Banana: Flux 2 Klein finetunes give generative models vision, 7 tasks in one — linoy_tsaban · 2026-11-16
- AI-animated short "Three Dollars, Two Wants" debuts its first episode trailer — Intelligent-Past1633 · 2026-11-16
- Hugging Face releases definitive fine-tuning guide for 3D generation model TRELLIS.2 — ariG23498 · 2026-11-16