MiniMax open-sources H3-Image: native 16MP output, up to 16 reference images
xiaohu · x · 2026-11-13
MiniMax released and open-sourced H3-Image, an image generation and editing model whose single set of weights handles both text-to-image and multi-reference editing, accepting up to 16 reference images and natively producing 16MP output with no separate upscaler. Distilled to 6 sampling steps.
- Native 4K: 4096×4096 square, 5472×3072 for 16:9 — 4x the pixels of the 2048 tier, generated directly at target size
- References each control one thing: official demos combine one empty-room photo with nine furniture images into a single living-room render; a gray render sets composition while the model adds materials and lighting; a red arrow on a map yields a first-person view
- Chinese prompts work: the API expands short requests into detailed English scene descriptions
- Local deployment: paths exist for SGLang, vLLM and ComfyUI (135 GB download); community license excludes the US, EU and other regions, with separate licensing above $1M annual revenue
- API pricing: $0.069 per 4096×4096 image, half price ($0.034) until Nov 18
More from Multimodal
- Creator Uses Seedance 2.5 to Generate a Heartwarming Short Film — SimplyAnnisa · 2026-11-13
- Fable 5.2 composes a symphony mimicking how AI learns, from noise to melody — FinanceYF5 · 2026-11-13
- Ghibli-style playable scene built with Fable 5.2, open sourced — FinanceYF5 · 2026-11-13
- Local pipeline showcase: MiniMax H3 refMod generations from reference image to 4K upscale — edisson75 · 2026-11-13
- DomoAI launches InkoAI Agent: describe a video in one sentence, get a ready-to-post clip — heyshrutimishra · 2026-11-13
- Gaussian splats may be the best way to capture how 3D displays like LKG Portrait look IRL — nptacek · 2026-11-13