Skip to main content
Back to timeline
openbiox BizardSource publication:

MiniMax opens the weights of a 16-megapixel image model: six sampling steps, every aspect ratio, no upscaler

Synopsis

MiniMax has released MiniMax-H3-Image on GitHub as an open-weight model that puts text-to-image and multi-image editing in one checkpoint, generating roughly 16-megapixel images directly at aspect ratios from 1:3 to 3:1 without a separate upscaling step, in six sampling steps. It accepts up to 16 reference images, and MiniMax describes its instruction following as the strongest among open-source models, with local deployment paths for SGLang, vLLM-Omni, diffusers and ComfyUI.

Interpretation

An open-weight image model now combines text-to-image and multi-image editing in a single checkpoint and outputs roughly 16-megapixel images directly across aspect ratios from 1:3 to 3:1. The common earlier route produced a smaller image and handed it to an upscaling model to add detail, whereas this model generates about 16 megapixels directly at each aspect ratio without an upscaling step. The stated specification is six sampling steps, output areas of 1K, 2K or 4K, custom width and height covering 655,360 to 20,000,000 pixels, and best results at about 16 megapixels.

The model encodes the text prompt and reference images, packs them into one sequence, and a single diffusion transformer predicts the latent of the target image, which is then decoded into pixels. The image encoder is a VAE with 16x spatial downsampling and the decoder is a 24-layer 2D ViT, so each transformer token covers 32x32 pixels. The diffusion transformer is a 33.1B-parameter dense single-stream model with 50 layers, hidden size 5376 and 56 attention heads, and the text encoder is MiniMax-H3's Qwen3-VL, read at its 50th layer.

Local deployment has ready paths: SGLang and vLLM-Omni need only the original checkpoint plus the MiniMax-H3 text encoder, about 135 GB. The full Hugging Face repository holds both the original checkpoint and the diffusers format, about 204 GB in total, so selective downloads avoid fetching the same parts twice. MiniMax recommends serving on one GPU, or letting SGLang offload model components when less than 130 GiB is free, and splitting a request across several GPUs with sequence parallelism for lower latency.

Perspective

The demonstrated scope is the repository's own model specification, architecture description, deployment commands and showcase images, aimed at developers who want to run or call the model on their own GPUs; the material gives no quantitative comparison with other models.

The repository calls its instruction following the strongest among open-source models, but the material gives no comparative evaluation supporting that claim; mask-based inpainting, transparent backgrounds, user-specified seeds and preview images are listed as not yet supported.