H3 text-to-image fix discovered: feeding any reference image switches it to a higher-quality edit mode
ninjazombiemaster · reddit · 2026-11-13
A Reddit user found that MiniMax H3's text-to-image mode often produces artifact-ridden "crusty" images — but there's a simple fix: feed any image into the reference input (even a blank black one) and quality improves dramatically, often repairing low-res generations outright.
Key findings:
- Discovered while adding an earring via image reference; artifacts vanished, and a blank black image worked too
- Current theory (discussed with Kijai, the ComfyUI integration author): once the model receives 64+ image tokens — triggered by feeding the image conditioning node any image, VAE or not — it switches to an edit/reference mode with much higher quality
- Workaround: plug in a small black image (auto-resized to at least 1024) to trigger the switch, at the cost of many wasted tokens
- Kijai is considering adding a switch to force ref mode without the pointless tokens
The author says this "transformed the model" — worth retrying if you wrote off H3's t2i.
More from Multimodal
- Fable 5.2 compresses 40 years of computer graphics evolution into a 29-second animation — FinanceYF5 · 2026-11-13
- Ant's Ling-3.2-flash demo turns meal photos into illustrated food journals — pipecanon · 2026-11-13
- Minimax-generated video of Jensen Huang handing NVIDIA to Tim Cook goes viral — Riddy_93 · 2026-11-13
- A 'very convincing' grape arbor photo shows how real AI images now look — misovalko · 2026-11-13
- Krea demos 'Agentic Editing' to SF growth crowd, calling it Cursor moment for creative work — angrypenguinPNG · 2026-11-13
- Manga ink backgrounds LoRA for NoobAI V-Pred delivers solid blacks and real linework — Historical_Song1864 · 2026-11-13