PANews: MiniMax open-sources image generation and editing model MiniMax-H3-Image
BIBIBI
AT A GLANCE
According to PANews, MiniMax has open-sourced the image generation and editing model MiniMax-H3-Image, which can directly generate images of approximately 16000000 pixels.
Article
MiniMax Open-Sources MiniMax-H3-Image Model, Capable of Directly Generating 16000000-Pixel Images
PANews November 10 — According to Hugging Face, MiniMax has open-sourced its image generation and editing model MiniMax-H3-Image. The same set of weights supports both text-to-image generation and image editing, with up to 16 reference images as input. The model does not require a separate super-resolution model and can directly generate images of approximately 16000000 pixels, with resolutions ranging from 4096×4096 to 6272×2688, and sampling steps distilled down to 6. Its diffusion Transformer has 33100000000 parameters, with weights of approximately 68.65GB, and can be deployed using SGLang, vLLM-Omni, diffusers, and ComfyUI. MiniMax claims that the model's instruction-following capability is the strongest among open-source models. The model has also been simultaneously launched on Mi...
(Click the link below to read the full article)
🔗 https://www.panewslab.com/zh/articles/01a1c625-9021-743f-960d-733355662823
Key points
01
PANews November 10 news, citing a Hugging Face report, said MiniMax has open-sourced MiniMax-H3-Image.
02
The same set of weights supports text-to-image generation and image editing, with support for up to 16 reference images as input.
03
The model does not need to be connected to an additional super-resolution model and can directly generate images of approximately 16000000 pixels, with resolutions covering 4096×4096 to 6272×2688.
04
The number of sampling steps has been distilled to 6; the diffusion Transformer has 33100000000 parameters, with weights totaling approximately 68.65GB.
05
The deployment methods listed in the report include SGLang, vLLM-Omni, diffusers, and ComfyUI.
06
MiniMax claims that this model has the strongest instruction-following capability among open-source models; this is its own statement.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a model that can both generate images from text and edit images. The report emphasizes that it can directly generate images of approximately 16000000 pixels without attaching a separate super-resolution model, and supports up to 16 reference images. The 6-step sampling and weight size mentioned in the text are technical specifications and cannot be used to confirm actual generation speed or deployment costs.
Why it matters to readers
It is worth noting that the same set of weights supports both generation and editing, and can directly output high-resolution images; however, the actual performance and the barrier to use still need to be verified.
You can first focus on whether it meets the needs of text-to-image generation, reference image input, and image editing. Do not directly equate high pixel counts or self-proclaimed claims of being the "strongest" with actually better results.
Risks and unknowns
The end of the main text is truncated, and the complete information for "the model has been synchronously launched on Mi..." cannot be confirmed.
The provided text does not include the original model page or an open-source license, so specific usage conditions cannot be confirmed.
"The strongest instruction-following capability among open-source models" is a self-description by MiniMax; the text does not provide independent evaluations.
The text does not provide actual generation speed, hardware requirements, deployment costs, or commercial adoption data.
Related Developments
Loading event timeline…
Related concepts
Text-to-image
This term is not in the glossary yet. Browse related concepts in the glossary.
