- 🆕 Text-to-image model from MiniMax's H3 omni-modal family, released 2026
- 📏 Native output around 16 megapixels, up to 4096×4096 without upscaling
- 🎯 Aspect ratios span 21:9 through 9:21, plus 1K and 2K sizes
- ⚡ Checkpoint distilled to six sampling steps, no classifier-free guidance
- 🧠 Diffusion transformer reuses H3 architecture minus the audio layers
- 🔧 Paired with MiniMax H3 Image Edit for editing and combining images
- 📚 Open weights published by MiniMax with a prompt-expansion companion model
- 🌐 Prompts expanded into detailed English descriptions before generation
MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects an ambitious and…
Explore 10 more models by Minimax →MiniMax H3 Image is the still-image member of MiniMax's H3 generative family, offering text-to-image rendering at 1K, 2K and 4K across aspect ratios from ultrawide 21:9 to tall 9:21. It is designed to work alongside MiniMax H3 Image Edit, which handles instruction-driven editing and combining of reference pictures.
Architecturally it is a direct descendant of the H3 base system that powers the video siblings such as MiniMax H3 and MiniMax H3 Max. MiniMax states that the image model's diffusion transformer uses the H3 architecture with the audio layers removed, and keeps the MiniMax-H3 text encoder unchanged. The main generational shift is resolution strategy: where the H3 video line reaches 2K by regenerating a 768p result in context, the image model produces roughly 16-megapixel output directly, without a separate upscaler, at sizes from 4096×4096 (1:1) to 6272×2688 (21:9).
Efficiency also improves over the base H3 pipeline. The released checkpoint is distilled to six sampling steps with no classifier-free guidance pass, and MiniMax reports that H3-Image can generate a 4K image in under ten seconds, attributing the gains to post-training changes spanning data, reinforcement learning, sampling-step reduction, prompt expansion and deployment optimisation. The company also reports support for up to 16 reference images.
Prompting differs from older MiniMax image workflows: the model expects a detailed English description of the whole scene, and a companion prompt-expansion model, MiniMax-H3-Image-PE, converts a short request in any language into that description plus an aspect-ratio recommendation. The effective working range is roughly 4 to 20 megapixels, with best results near 16 megapixels.
This About section is AI-generated from public sources via VeniceStats + Venice inference, with no human editing. It may contain inaccuracies.
| Seller | Rep↓ | Sales | Images | Routing | $ / img | Save |
|---|---|---|---|---|---|---|
| ▲ Apex Ant 0x73b4…e736 | 81.71 | — | — | #1 | $0.02 | −50% |
"Best price" and the seller table are live Antseed catalog data (advertised $/1M tokens — or $ per generated image for unit-billed image models — not settled amounts). Reputation = buyer trust score (0-100, the Antseed SDK's own formula). "Routing" = the SDK's default buyer routing (what the VPR desktop app ships with): a trust ≥ 60 gate on the effective reputation, then cheapest-first among routable sellers; live failover state (per-peer cooldowns) is buyer-side runtime and not included. Model knowledge (TLDR, provider, About) via the VeniceStats enrichment layer. Advertised catalog, not the model used in any specific purchase. "Usage on Antseed" counts only settlements whose buyers share the per-model split on-chain (metadata v2/v3, opt-in), so every usage figure is a lower bound.