MiniMax Drops MiniMax-H3-Image, a 33B Open Model Generating 16MP Photos in 6 Steps
MiniMax released a 33B parameter open-weights image model that generates and edits 16 megapixel images in just 6 sampling steps.
- MiniMax released MiniMax-H3-Image, a 33.1B parameter open-weights diffusion transformer for text-to-image and editing.
- Generates native 16 megapixel images up to 6272x2688 without any upscaler in the pipeline.
- Distilled to just 6 sampling steps with no classifier-free guidance, cutting inference cost sharply.
- Unified checkpoint handles text-to-image plus editing with up to 16 reference images via
<Picture N>tags. - Deployable through SGLang, vLLM-Omni, Diffusers and ComfyUI, with hosted API on platform.minimax.io.
- Requires long structured English prompts, aided by the H3-Image-Context-IR expander or local MiniMax-H3-Image-PE model.
MiniMax has released MiniMax-H3-Image, an open-weights model for text-to-image generation and reference-based editing. It produces images at roughly 16 megapixels across supported aspect ratios, runs a six-step denoising schedule, and accepts as many as 16 reference images.
The checkpoint combines generation and editing, removing the need to maintain separate models for those workflows. Its native resolution also removes a separate upscaling stage, although the 33.1-billion-parameter diffusion transformer and large text encoder require substantial GPU memory.
Six steps, 16 megapixels
| Capability | Specification |
|---|---|
| Native output | About 16.8 megapixels |
| Example resolutions | 4096 × 4096 at 1:1; 6272 × 2688 at 21:9 |
| Sampling | 6 steps, or 6 neural function evaluations |
| Guidance | No classifier-free guidance; guidance scale fixed at 1 in the hosted API |
| Reference inputs | Up to 16 images |
| Diffusion transformer | 33.1 billion parameters |
| DiT checkpoint size | 68.65 GB in BF16 |
| Primary workflows | Text-to-image generation and image editing |
A six-step schedule means the diffusion transformer evaluates each request six times while progressively converting noise into an image. Classifier-free guidance commonly adds conditional and unconditional evaluations to each step, roughly doubling transformer work. MiniMax-H3-Image omits that process, reducing compute per image and eliminating negative prompts from the hosted API.
How the model fits 16 MP
The 33.1B-parameter diffusion transformer uses a dense, single-stream design derived from the MiniMax H3 family. Its 50 layers have a hidden size of 5,376, 56 attention heads, and a feed-forward dimension of 14,336. Three-dimensional multimodal rotary position embeddings provide positional information across the model’s text and image inputs.
The text component is the full Qwen3-VL encoder used by MiniMax-H3, with the diffusion transformer consuming hidden states from its 50th layer. The image variational autoencoder, or VAE, compresses pixels into a smaller latent representation before diffusion and reconstructs the final image afterward. It combines a convolutional encoder with a two-dimensional vision-transformer decoder, uses 16× spatial downsampling, and produces 24 latent channels.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
