MiniMax Releases H3-Image, an Open-Weight Model for 4K Image Generation and Editing From Up to 16 References

MiniMax released H3-Image, an open-weight model that generates 16-megapixel images and edits from up to 16 reference images, claiming 4K generations in 10 seconds in its launch post on X. High-resolution stills and reference-driven edits come from a model you can download and run yourself or use through MiniMax's hosted services.

• 16MP straight out of the model. The model card lists native sizes from 4096x4096 at 1:1 to 6272x2688 at 21:9, generated directly without an upscaler.
• Generation and editing share one checkpoint. The same weights handle text-to-image and edits that pull from one to 16 reference images.
• Six sampling steps. The model is distilled to six steps with no classifier-free guidance, so there is no guidance value to tune.
• Open weights with several deployment routes. The model card recommends SGLang, vLLM-Omni, diffusers and ComfyUI, and the model ships under a community license, according to PANews.
• Hosted access through three channels. H3-Image is available through MiniMax's API, Hailuo AI and MiniMax Code.

H3-Image is the H3 architecture minus the audio layers, built around a 33.1B-parameter transformer

The image model comes from the MiniMax H3 family, which we covered as a single model for text, image, video and audio generation.

It follows the H3-Base design from MiniMax with the audio branch removed. It reuses MiniMax-H3's full Qwen3-VL text encoder unchanged.

The diffusion transformer has 33.1 billion parameters across 50 layers, and each transformer token covers a 32x32-pixel patch.

MiniMax says 16MP images carry far more detail than the 1-4MP outputs common in image generation but usually cost substantially more compute. The company claims H3-Image keeps that cost manageable and delivers the strongest instruction following among open-source models.

Edits name each reference image by number in the prompt

In ComfyUI, H3-Image runs natively on your own GPU or through Partner Nodes, and connecting images switches the same graph from text-to-image to editing, according to the ComfyUI blog. The editing workflow works like this:

• Numbered references. Prompts name references as <Picture 1>, <Picture 2> and so on, numbered by input order.
• image_1 is the edit target. The first input is the picture being edited; the other inputs act as references.
• The VAE anchors references. Connecting the VAE anchors reference images as latents in addition to feeding them through the text encoder.

ComfyUI prompts can run up to 4096 tokens. For text-to-image, the model card lists an optional prompt expander, MiniMax-H3-Image-PE, that writes longer descriptions.

The MiniMax API's edit endpoint accepts references as public URLs, uploaded files or base64 data, with a 64 MiB cap on the request body. MiniMax lists masked edits, transparent backgrounds and preview images as unsupported.

Local runs need about 135 GB of weights, while the API reaches 20MP

The transformer weighs 66.28 GB in BF16, the VAE adds 2.34 GB, and the text encoder, loaded separately from the MiniMax-H3 repository, adds 66.7 GB. The optional prompt expander adds 17.9 GB in NVFP4.

Everything downloads from Hugging Face, and anyone who already has MiniMax-H3 skips the text encoder download.

Through the API, output area can reach 20 million pixels with custom width and height, aspect ratios from 1:3 to 3:1, and 1K, 2K and 4K area presets, with 4K as the default. A MiniMax-H3-Image-Highspeed variant takes the same parameters and limits, and outside the API the model also runs in Hailuo AI and MiniMax Code.