MiniMax H3 Image: What It Is and How to Use It
MiniMax H3-Image is MiniMax’s open-weight image model for text-to-image, reference-based generation, and image editing, with native outputs around 16 megapixels, including 4096×4096.
What is MiniMax H3 Image?
MiniMax H3-Image (written MiniMax-H3-Image on Hugging Face) is an open-weight image generation and editing model in the MiniMax H3 family. The company announced it on its research blog as supporting text-to-image, reference-based generation, and image editing from one checkpoint. MiniMax says it can generate a 4K (4096×4096) image in under 10 seconds in its own setup, and that it can work with up to 16 reference images.
The official model card adds the technical picture. Generation and editing are unified: text-to-image and edits of one or more reference images share a checkpoint. Images are produced at about 16 megapixels across several aspect ratios, from 4096×4096 (1:1) up to 6272×2688 (21:9), “directly and without an upscaler.” The diffusion transformer is distilled to six sampling steps with no classifier-free guidance. The DiT has 33.1 billion parameters in BF16 and uses the MiniMax-H3 / Qwen3-VL text encoder.
Weights are released under the MiniMax-H3-Image Community License. An optional prompt expander, MiniMax-H3-Image-PE, is documented on the same card. The model is also offered as a hosted service through the MiniMax API, MiniMax Code, and Hailuo AI. The blog says Standard mode is positioned as 1.5–2× as fast as “mainstream alternatives,” Highspeed mode as 3–4×, and pricing as 50–70% of comparable offerings, with a launch-week 50% discount on 4K and above — those commercial claims come from MiniMax, not from an independent test on this site.
How to use it
You can run H3-Image in three official ways: call MiniMax’s API or apps, serve the open weights yourself, or use the native ComfyUI templates.
Hosted API and apps
The model card points to the global platform at platform.minimax.io and a China platform at platform.minimax.cn, plus Hailuo web and MiniMax desktop apps. Official request routes include text-to-image at `/v1/images/generations` and image editing at `/v1/images/edits`. Prompts are truncated at 4096 tokens. The documented sampler settings are fixed: 6 steps and guidance 1; other values are rejected, and negative prompts are not supported. Write anything you want to avoid into the positive prompt instead.
MiniMax recommends a detailed English description of the whole image. Short prompts can be expanded with H3-Image-Context-IR on the API, or locally with MiniMax-H3-Image-PE.
Local serving
The card documents downloads via `hf download MiniMaxAI/MiniMax-H3-Image` plus the MiniMax-H3 text encoder, and serving examples for SGLang, vLLM-Omni, and diffusers. Full BF16 weights are large (the card lists tens of gigabytes for the DiT and the text encoder), so check VRAM before you try a 16 MP local run.
ComfyUI
ComfyUI’s official tutorial requires ComfyUI 0.45.0 or newer and ships Text-to-Image and Image Edit templates. The tutorial uses Comfy-Org repacks (including a pruned INT8 DiT) and the same 6-step, CFG 1 schedule. Templates may open at about 4 MP so smaller GPUs do not run out of memory; you can raise resolution when the machine has room. In an edit, the first image is the one being changed, and further inputs are addressed in order in the prompt.
Practical tips from the official docs
- Use the published sizes when you want native 16 MP output: 4096×4096 (1:1), 4736×3552 (4:3), 5024×3360 (3:2), 5472×3072 (16:9), 6272×2688 (21:9), and the matching portrait sizes.
- Do not fight the distilled schedule. Official serving returns an error if you set a step count other than 6 or a guidance value other than 1.
- Prefer long, complete descriptions. Both MiniMax and ComfyUI say the model is built for detailed captions; turn the expander on when your prompt is short.
- For edits, name references in order. The docs describe the first image as the base and later images as identity, props, or layout references.
- If a local 16 MP job runs out of memory, follow the ComfyUI note to start at a smaller megapixel target or enable the resize node in the template.
- Read the Community License before you ship a product on top of the weights, including any extra authorization MiniMax asks for in Section IV of that license.
Where to get it
Announcement: minimax.io/blog/minimax-h3-image. Weights and cards: MiniMaxAI/MiniMax-H3-Image. ComfyUI guide: docs.comfy.org. Hosted API: platform.minimax.io. A China mirror is listed on ModelScope on the ComfyUI page.
FAQ
Is MiniMax H3 Image open source?
MiniMax released open weights under the MiniMax-H3-Image Community License. Read that license before commercial use. Hosted API access is separate and priced by MiniMax.
Can one checkpoint both generate and edit images?
Yes. Official docs say text-to-image and image editing are unified in a single checkpoint, including up to 16 reference images.
What sampler settings should I use?
Six inference steps and guidance (CFG) of 1. Official serving rejects other values. Negative prompts are not supported.
What resolution does it generate?
Native output is about 16 megapixels, for example 4096×4096, without an upscale pass. Smaller canvases also work and use less memory.
How do I run it in ComfyUI?
Update to ComfyUI 0.45.0 or later, download the documented model files, and load the official Text-to-Image or Image Edit template.
Sources: MiniMax H3-Image blog post · Hugging Face model card · ComfyUI official tutorial
Unofficial guide based on MiniMax and ComfyUI documentation. We are not affiliated with MiniMax or Comfy Org. Speed and pricing claims in the announcement are the company’s own.