MiniMax Open-Sources H3 Image: Native 4K in 10 Seconds, Up to 16 Reference Images
Open-sourcing H3 Image puts a 4K-capable, multi-reference image generator into any local ComfyUI setup, cutting dependence on paid cloud APIs and enabling high-res asset generation without upload limits.
MiniMax has open-sourced H3 Image, a 4K image generation model that tops the open-source SOTA list. Officially, it renders native 4096×4096 (16MP) images in under 10 seconds; the author's own RTX 5090 took about a minute, but the weights are out and ComfyUI templates are ready, so local deployment is straightforward.
The model's headline feature is multi-reference control: it accepts up to 16 separate reference images, letting you assign characters, outfits, color palettes, or scene elements by number. Community examples show it turning map coordinates into game-style scenery, applying materials and lighting to untextured "white models," producing multi-view mech line art, arranging eight people with individual color schemes in one group photo, and even laying out furniture in a room from separate product shots.
For local use, ComfyUI 0.45.0+ and template pack 0.11.104+ are required. The author advises using an AI agent to pick a quantized version unless your VRAM is ample, since the official weights can be heavy. Chinese text generation remains a weak spot: characters occasionally break, needing retries. Cloud access through Hailuo AI is paid; the open weights and ComfyUI integration are the main draw.
The 16-reference-image capability is the standout differentiator, enabling precise control over multiple characters and assets, but it demands careful prompt construction to assign each reference correctly.
Chinese text rendering remains an unresolved pain point even as image quality improves, showing that multilingual text-in-image is a separate challenge from general generation.
Open weights and ComfyUI integration lower the barrier to local high-res generation, but VRAM constraints push users toward quantized versions, so practical deployment is still hardware-bound.
The example workflows (map-to-scene, white-model texturing, multi-view sheets) target production pipelines like game asset creation and storyboarding, not just casual image making.