MiniMax-AI/MiniMax-H3-Image

Introduction

MiniMax-H3-Image is an open-weights image model from the MiniMax H3 family that unifies text-to-image generation and reference-image editing in a single checkpoint. The DiT has 33.1B parameters, uses the unchanged Qwen3-VL text encoder from MiniMax-H3 (50th layer hidden states), and a CNN/2D-ViT image VAE with 16x downsampling. It is distilled to 6 sampling steps with no classifier-free guidance, and outputs roughly 16-megapixel images at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), without an upscaler. It targets creators and developers who want controllable high-resolution generation and edits, including up to 16 reference images per edit. See the model card on Hugging Face for weights.

Quick Start

Download the model and the shared H3 text encoder into the Hugging Face cache, then serve it with a supported inference framework. The documented SGLang path:

hf download MiniMaxAI/MiniMax-H3-Image
hf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*"
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode off

On four GPUs, add --num-gpus 4 --ulysses-degree 4 for sequence parallelism. To use a local text encoder folder instead of the cache, pass --text-encoder-path <path>. Text-to-image requests are JSON to /v1/images/generations; edits send reference images as multipart form data to /v1/images/edits, with upload order mapping to <Picture 1>, <Picture 2>, and so on. Both return a PNG in data[0].b64_json.

Key request constraints: size and prompt are required; num_inference_steps must be omitted or 6 and guidance_scale omitted or 1 (other values return HTTP 400); negative prompts are not supported. Set generator_device to cpu for reproducible noise from a seed. min_generation_area defaults to 4194304 (2048x2048); small sizes are generated on a larger canvas and resized with LANCZOS unless you pass 1024 to disable it.

For API-level image quality, the README uses the MiniMax H3-Image-Context-IR API to expand short prompts into detailed English descriptions, with a local option, MiniMax-H3-Image-PE. The runnable examples live in the examples/ folder; they expect MINIMAX_API_BASE (e.g. https://api.minimax.io) and MINIMAX_API_KEY. Full deployment instructions are in the SGLang cookbook; vLLM, diffusers and ComfyUI are also listed as serving options.

Constraints

• The full download is large: 66.28 GB for the transformer, 2.34 GB for the VAE, and the text encoder (66.7 GB) is pulled from the separate MiniMax-H3 repository if not already cached. An extra 17.9 GB is needed for the optional prompt expander.
• Generation is fixed to 6 NFEs and guidance scale 1; the model is not configured for other sampling schedules through the API.
• Effective working range is about 4–20 MP, with best results around 16 MP; sizes are selected from the documented aspect ratios.
• Serving requires a GPU and one of the supported inference stacks.
• Weights and the PE model are governed by the MiniMax-H3-Image Community License Agreement; authorization beyond it must be requested through MiniMax's license page.

HotRepo's take

This is a fresh release: HotRepo first recorded the repository on 2026-11-11 with 117 stars and 4 forks, and the new-project snapshot reports 15 stars gained over the one-hour observation window starting 06:15 UTC that day. With no language detected in the repository and no daily all-language ranking recorded yet, there is not enough same-metric history to draw a trend. What the record does show is a weights-focused release whose repository is mostly documentation and examples; the practical work happens through the Hugging Face checkpoints and inference frameworks. For readers deciding whether to try it, the 16 MP output and unified generate/edit checkpoint are the concrete differentiators to evaluate against the download and serving requirements above.