Today, we open-source P-H3-Image, an image generation and editing model built from MiniMax-H3 by MiniMax (official). A SOTA video model, turned into an efficient open image model. - 𝗛𝗼𝘄 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗶𝘀 𝗶𝘁? Generate or edit images 5× faster, in 8 steps instead of 50. A 4 MP image takes... 2.6 s instead of 12.4 s. The model is 7× smaller than H3, so high-resolution images are a light download and run on a much smaller GPU. - 𝗛𝗼𝘄 𝗴𝗼𝗼𝗱 𝗶𝘀 𝗶𝘁? Strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It renders 2 to 16 MP and edits with up to 8 input images. -𝗛𝗼𝘄 𝗱𝗼𝗲𝘀 𝗶𝘁 𝘄𝗼𝗿𝗸? We reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer, and the decoder. Use it with Diffusers or ComfyUI. 📄 Tech Report: 🤗 Hugging Face Open Weights: 👉 ComfyUI Native Support:show more

Try P-H3-Image on the HF Space:

@MiniMax_AI Exciting! 👀

@MiniMax_AI How does this compare to qwen 2.1 for image editing? thats what I've been using mainly

@MiniMax_AI 5× faster generation with open weights? This is a huge win for the AI image generation community! 🔥

@MiniMax_AI A great initiative. I'm excited to try it.

@MiniMax_AI 8 steps down from 50 is a big cut. Distilled schedule or just fewer sampling steps? That distinction matters for how much to trust quality at higher resolutions.

@MiniMax_AI I think it's amusing that you got the image model variant of their model out before they did

@MiniMax_AI @Presidentlin

@MiniMax_AI Tried the free Space with a red mug labeled OPEN LATE plus a brass key and green notebook. One 4 MP run got the words and all three objects right. The Space reported 3.9s rendering plus 4.0s prompt upsampling on GPU. Queue time was separate.

@MiniMax_AI mini max going from video to an open image model is wild

@MiniMax_AI Going from 50 steps to 8 is the part that matters for local use. I would be curious how the quality holds up on image editing compared to plain generation.

@MiniMax_AI does prompting usesame syntax as H3 minimax ?

@MiniMax_AI @realrebelai @Zanzibased

@MiniMax_AI @grok will 16gb ram with 12gb rtx 3060 run it fine and in speed?

@MiniMax_AI More faster 🔥🔥

@MiniMax_AI 8 steps instead of 50, without compromising on quality. That’s a serious leap in making high-quality image generation more accessible. Open-source efficiency and photorealism are a powerful combination. Excited to see what the community builds

@MiniMax_AI Checked the repo: the transformer is 33.1B params shipped in BF16, so 66 GB of the ~100 GB download. The prompt upsampler adds another 19 GB. Is an FP8 or NVFP4 build planned? At ~4.8 real bits the transformer would be ~20 GB.

@MiniMax_AI 8 steps instead of 50" is the kind of math I'd love to see in my own pipelines.