MiniMax Releases Open H3-Image Model: 16 MP Without Upscaling in 6 Steps

Most open image models draw at 1–4 megapixels. Anything larger has to be run through an upscaler, which fills in fine details and sometimes ruins text or faces. MiniMax has released an official open model, MiniMax-H3-Image, built on H3, that works with high resolutions directly.

What the model can do

The headline figure is 16 MP without an upscaler. That covers formats from 4096×4096 to 6272×2688, and the stated working range is 4–20 MP. For print, large banners, and big previews, this saves one step in the pipeline and removes upscaling artifacts.

The second feature is generation speed. The checkpoint is distilled, so 6 steps (NFEs) are enough. No CFG is needed, and no negative prompt either. That leaves fewer settings to tune by hand.

The third is reference-based editing. The model accepts up to 16 images. The first is the one you are editing. The rest define characters, objects, materials, and composition. For example, you take a photo of an interior and add a specific sofa and a fabric texture to the frame, while pulling in a character from a separate image.

MiniMax also claims the best instruction following among open models and strong text rendering. In the examples, the model solves logic puzzles and draws boards with conditions. This is the company's own claim, and the post contains no independent comparisons, so test it on your own prompts.

Hardware requirements

The model is heavy, and that is its main limitation. According to the description:

• 24 GB VRAM is the practical minimum. The int8 DiT takes 21 GB and handles 4 MP, while the text encoder is offloaded to system RAM. The headroom is very small.
• 16 GB works only with w4a8 weights (12.5 GB) and 2–4 MP, with no guarantees.
• 32 GB is already comfortable, at 4–8 MP.
• 48 GB or more lets you do 8–16 MP without offloading.

You will need about 50 GB of disk space for the INT8 package and about 93 GB for BF16. You need at least 64 GB of RAM: when offloaded, the encoder takes 27 GB, and there is nowhere else to put it.

There are no speed measurements on consumer GPUs yet. The author of the post asks anyone who runs the model to share numbers. So the real time per image on an RTX 4090 or 5090 is still unknown.

How to get started and who it's for

Starting settings are: 4 MP, 6 steps, CFG 1, euler sampler, and simple scheduler. The model is available on Hugging Face, there is a demo space, and Comfy-Org has published a repack and a guide for ComfyUI. If you already work in Comfy, you can plug it in without separate infrastructure.

The takeaway for developers and product managers is simple. If your product generates large images (posters, ad layouts, product cards, concept art), H3-Image is worth testing. It is especially useful where an upscaler currently sits in your pipeline or where you need character consistency through references. Open weights let you keep everything in-house and avoid paying for an API per image.

There are caveats. The memory requirements rule out most home GPUs. Speed and quality on real tasks have not been independently verified. The claim of best instruction following remains marketing until someone runs a comparison against other open models. A sensible plan: use the demo or rent a 32–48 GB GPU, run 20–30 of your typical prompts at 4 MP, and see whether the model holds up on text and composition. Only then decide whether to move production to it.