MiniMax-H3-Image: Unified text-to-image and image editing, directly generating 16 million pixels

MiniMax's open-source image generation and editing model, part of the MiniMax H3 family: The DiT follows the H3 architecture but removes the audio layers, has 331 billion parameters, and uses the text encoder from MiniMax-H3. The same checkpoint supports both text-to-image generation and editing based on one or more reference images, with official claims supporting up to 16 reference images. Officially, the model can directly generate images of approximately 16 MP across various aspect ratios (1:1 is 4096×4096, 21:9 is 6272×2688) without needing an upscaler; the model is distilled to 6 steps and does not use CFG. Model card examples include realistic photography, concept design, text rendering for Chinese and English logic puzzle solution boards, converting white-model renders into cinematic stills, and compositing a living room from ten reference images consisting of an empty room plus nine pieces of furniture.

Original title: MiniMaxAI/MiniMax-H3-Image

MiniMax's open-source image generation and editing model, part of the MiniMax H3 family: The DiT follows the H3 architecture but removes the audio layers, has 331 billion parameters, and uses the text encoder from MiniMax-H3. The same checkpoint supports both text-to-image generation and editing based on one or more reference images, with official claims supporting up to 16 reference images. Officially, the model can directly generate images of approximately 16 MP across various aspect ratios (1:1 is 4096×4096, 21:9 is 6272×2688) without needing an upscaler; the model is distilled to 6 steps and does not use CFG. Model card examples include realistic photography, concept design, text rendering for Chinese and English logic puzzle solution boards, converting white-model renders into cinematic stills, and compositing a living room from ten reference images consisting of an empty room plus nine pieces of furniture.

Why it matters Directly generates approximately 16 million pixels without an upscaler, uses six-step sampling, supports up to sixteen reference images, and is already supported by mainstream inference frameworks and ComfyUI

Rising score 471 · Likes 476 · Downloads 667

Reasons

  • Officially claims the strongest instruction following among open-source models; short prompts can first be expanded into full English descriptions via H3-Image-Context-IR (API) or the locally deployable MiniMax-H3-Image-PE before generation
  • Comprehensive deployment paths: Official documentation or templates are available for SGLang, vLLM-Omni, diffusers, and ComfyUI; officially listed hardware includes H200, RTX PRO 6000 Blackwell, and AMD MI350X
  • The model card provides a comparison of images generated using the same prompt via local SGLang and the official API, along with ready-to-run example scripts

Usage

  • Run hf download MiniMaxAI/MiniMax-H3-Image, then download the text_encoder/ from the MiniMax-H3 repository; start the service using sglang serve or vllm serve --omni. Send text-to-image requests to /v1/images/generations and editing requests to /v1/images/edits (reference images correspond to <Picture 1>, <Picture 2> in the prompt in upload order)
  • Specify size as width x height; official recommendation is approximately 16 MP (e.g., 4096x4096, 5472x3072); steps must be 6, guidance_scale must be 1, and negative prompts are not supported
  • Prompts should be complete English scene descriptions; short prompts should first be expanded using MiniMax-H3-Image-PE
  • ComfyUI: Use the packaged version from Comfy-Org/MiniMax-H3-Image and follow the official tutorial to apply text-to-image or editing templates

Risks

  • Large size: DiT weights are approximately 66 GB, plus approximately 67 GB for the MiniMax-H3 text encoder, requiring data-center-grade high-VRAM GPUs
  • Released under the MiniMax-H3-Image community license; written authorization must be separately requested for scenarios listed in Section 4 of the license, so read the terms before commercial use
  • "Strongest instruction following among open-source models" is an official claim; the model card does not include third-party evaluations
  • This entry has not been run; refer to the model card for actual performance

Open

Hugging Face Trending

Explore the latest AI Radar updates