MiniMax Releases H3 Image Model

MiniMax released MiniMax-H3-Image as an open-source model for text-to-image generation and image editing. The model accepts text prompts and one or more reference images in a single checkpoint. MiniMax says it follows instructions precisely, supports up to 16 reference images, and produces images of about 16 megapixels across aspect ratios from 4096x4096 (1:1) to 6272x2688 (21:9), without an upscaler. Its effective working range is about 4 MP to 20 MP, with best results generally around 16 MP.

The model is distilled to 6 sampling steps and runs without classifier-free guidance. Its diffusion transformer (a neural network that generates images through iterative denoising) uses the H3 architecture with its audio layers removed; the MiniMax-H3 text encoder is unchanged. The model has 33.1B parameters, 50 layers, a hidden size of 5376, 56 attention heads, and a feed-forward network size of 14336. Its weights use BF16, while the embedding, timestep, and output layers use FP32.

MiniMax describes the model as combining speed, image quality, and instruction following. It says the 16-megapixel output provides more detail than the 1–4 MP images it describes as common, while keeping computational cost manageable. The company also claims strong realism and instruction following among open-source models. Those are MiniMax's stated comparisons; the page provides no benchmark scores to quantify them.

The model's image VAE (a component that encodes and decodes image representations) combines a CNN encoder with a 24-layer 2D ViT decoder, 16x spatial downsampling, and 24 latent channels. Its transformer processes prompt and image tokens together to predict the target image representation, which the VAE then decodes. The repository lists the main BF16 model at 68.65 GB and an optional prompt-expansion model, MiniMax-H3-Image-PE, at 17.9 GB.

Users can access MiniMax-H3-Image through MiniMax's API or app, or deploy it locally with SGLang, vLLM-Omni, diffusers, LightX2V, or ComfyUI. The page gives one-GPU serving examples and lower-latency four-GPU configurations using sequence parallelism. It says both NVIDIA and AMD accelerators are supported, including NVIDIA H200, RTX PRO 6000 Blackwell, and AMD Instinct MI350X. Text-to-image and editing requests use separate API endpoints; edits accept one or more uploaded reference images, whose upload order determines their labels in the prompt.

Prompts are limited to 4096 tokens. Generation uses 6 inference steps and a guidance scale of 1; other values for either parameter return HTTP 400, and negative prompts are unsupported. The optional minimum generation area ranges from 1024 to 16777216 pixels and defaults to 4194304 pixels (2048x2048). The page says the model expects a detailed English description; MiniMax's API can expand a short prompt in any language before generation, and the optional prompt-expansion model can be deployed locally. The model is released under the MiniMax-H3-Image Community License Agreement. Hugging Face listed 142,989 downloads in the last month.

MiniMax released MiniMax-H3-Image as an open-source model for text-to-image generation and image editing. The model accepts text prompts and one or more reference images in a single checkpoint. MiniMax says it follows instructions precisely, supports up to 16 reference images, and produces images of about 16 megapixels across aspect ratios from 4096x4096 (1:1) to 6272x2688 (21:9), without an upscaler. Its effective working range is about 4 MP to 20 MP, with best results generally around 16 MP.

The model is distilled to 6 sampling steps and runs without classifier-free guidance. Its diffusion transformer (a neural network that generates images through iterative denoising) uses the H3 architecture with its audio layers removed; the MiniMax-H3 text encoder is unchanged. The model has 33.1B parameters, 50 layers, a hidden size of 5376, 56 attention heads, and a feed-forward network size of 14336. Its weights use BF16, while the embedding, timestep, and output layers use FP32.

MiniMax describes the model as combining speed, image quality, and instruction following. It says the 16-megapixel output provides more detail than the 1–4 MP images it describes as common, while keeping computational cost manageable. The company also claims strong realism and instruction following among open-source models. Those are MiniMax's stated comparisons; the page provides no benchmark scores to quantify them.

The model's image VAE (a component that encodes and decodes image representations) combines a CNN encoder with a 24-layer 2D ViT decoder, 16x spatial downsampling, and 24 latent channels. Its transformer processes prompt and image tokens together to predict the target image representation, which the VAE then decodes. The repository lists the main BF16 model at 68.65 GB and an optional prompt-expansion model, MiniMax-H3-Image-PE, at 17.9 GB.

Users can access MiniMax-H3-Image through MiniMax's API or app, or deploy it locally with SGLang, vLLM-Omni, diffusers, LightX2V, or ComfyUI. The page gives one-GPU serving examples and lower-latency four-GPU configurations using sequence parallelism. It says both NVIDIA and AMD accelerators are supported, including NVIDIA H200, RTX PRO 6000 Blackwell, and AMD Instinct MI350X. Text-to-image and editing requests use separate API endpoints; edits accept one or more uploaded reference images, whose upload order determines their labels in the prompt.

Prompts are limited to 4096 tokens. Generation uses 6 inference steps and a guidance scale of 1; other values for either parameter return HTTP 400, and negative prompts are unsupported. The optional minimum generation area ranges from 1024 to 16777216 pixels and defaults to 4194304 pixels (2048x2048). The page says the model expects a detailed English description; MiniMax's API can expand a short prompt in any language before generation, and the optional prompt-expansion model can be deployed locally. The model is released under the MiniMax-H3-Image Community License Agreement. Hugging Face listed 142,989 downloads in the last month.