MiniMax H3 Image Open-Source Debut Hits ComfyUI Day 0, Community Tests Mixed

On November 11, MiniMax open-sourced its image generation model H3 Image, gaining official native ComfyUI support the same day along with Day 0 integration on the Fuobe platform. A single checkpoint covers both text-to-image and reference-image editing (up to 16 reference images), generating roughly 16 megapixels per output (native 4096×4096). Thanks to CFG distillation, it needs only 6 sampling steps with guidance 1 and no separate upscaler. Prominent Hugging Face developer multimodalart praised it after trying it out, saying it finally gives open-weight models speed, quality, and aesthetic novelty all at once.

Confirmed

• Single checkpoint, dual modes: ComfyUI officially provides two native workflows—plug reference images into the conditioning node for editing, leave them out for text-to-image (m1/m6/m8)
• Generation specs: 16MP, 6 sampling steps, guidance 1, no upscaler needed; Reddit user hakanavgin ran systematic tests using H3 Image Pruned W4A8 with the qwen3vl32b int4 text encoder and euler-simple at 16MP/6 steps, calling it the strongest open model for photos/posters, though prompts need "sanitizing" (m2/m6/m8)
• Editing ability widely praised: kfromHyperDraw tested pose control, repainting styles, and relighting and found them "work like magic," capable of editing from hand-drawn doodles all the way to macro close-ups with near-natural-language prompts; MaterialWestern3016 used the official workflow to successfully combine 5 reference images into a "woman walking a dog along a riverbank" (m4/m7)
• Speed: FrontPraline9683 reported that local 16MP 4-step editing takes about 1 minute, but with 1 reference image, prompt encoding generating 8k tokens takes roughly 25 minutes—a jarring experience (m5)

Unconfirmed

• Split verdicts on faces and text-to-image quality: MaterialWestern3016 said low-resolution text-to-image and people often fail (dogs and a leather jacket were fine); Divineboob said text-to-image is fast with no plastic faces, but image-to-image completely ignores prompts, Refmod doesn't work, and VNCCS follows pose while mangling character reference; kfromHyperDraw also found small edits easily ruin images. Conclusions vary, likely depending on workflow and parameters (m3/m4/m7)

Why it matters

• This is a rare image model that was open-sourced at launch and gained native support in mainstream toolchains on day one, lowering the barrier to trying it
• Sharing one set of weights for editing and generation simplifies deployment, and its editing ability is rated by multiple users as a standout in the open camp—but character consistency, image-to-image prompt adherence, and local PE speed remain commonly cited weaknesses that will affect real-world production readiness

2026-11-10 ~ 2026-11-11 · 10 related posts

Full story(4 episodes)→

• Episode 1: MiniMax Hailuo H3-Image Early Access: up to 16 reference images, 4K output(2026-11-10, 24 posts)
• Episode 2: MiniMax Releases and Open-Sources H3-Image: 16MP Output in 10-20 Seconds(2026-11-10, 28 posts)
• Episode 3: MiniMax H3 Image Open-Source Debut Hits ComfyUI Day 0, Community Tests Mixed(2026-11-10, 10 posts)
• Episode 4: MiniMax launches H3-Image model for game assets and 3D scenes(2026-11-11, 3 posts)

• MiniMax H3 Image open-weights image model praised for speed, quality and novel aesthetics — multimodalart · 2026-11-10
• [source] MiniMax H3 Image lands in ComfyUI day-one: 6-step sampling, native 4096×4096 output — Lexius2129 · 2026-11-11
• [source] MiniMax-H3-Image lands in ComfyUI: native 16MP, 6-step, no upscaler — Acrobatic-Example315 · 2026-11-11
• MiniMax H3 Image hands-on: fast plastic-free T2I, but edits and consistency fail — Divineboob · 2026-11-11
• Minimax H3 Image hands-on: superb edits, but small repeated edits ruin images — k_from_HyperDraw · 2026-11-11
• MiniMax H3-Image hands-on: strong editing, weak faces, resolution quirks in ComfyUI — MaterialWestern3016 · 2026-11-11
• Minimax H3 Image tested: best open-weights photo/poster model, but never say "realistic" — hakanavgin · 2026-11-11
• MiniMax H3 Image day-0 hands-on: workflow tips, parameters and prompt tricks — ggbro_itsover · 2026-11-11
• MiniMax H3 Image local image editing: 1-min edits but 25-min prompt encoding — Front_Praline9683 · 2026-11-11
• [source] MiniMax H3 Image day-one tips: one checkpoint for t2i and edits — ggbro_itsover · 2026-11-11