P-H3-Image: Pruna turns MiniMax H3 into an 8-step image model

What it is. Per Pruna, they rebuilt every part of H3 for still images: a Qwen3.5-9B prompt upsampler, a 4B text encoder distilled from H3's 32B one, an 8-step schedule instead of 50, a slimmer 20B backbone and a single-frame decoder that gives sharp images instead of the soft frames H3 produces on its own. Images run from 2 to 16 megapixels, and edits can combine up to 8 or 9 reference images (the model card and the blog give different numbers).

What it needs. The smallest ComfyUI set is a 20 GB download that peaks around 21 GB of VRAM at 8 MP, per Pruna. bf16 is about 55 GB to download and 56 GB of VRAM. The vendor's speed figures are 2.6 s for a 4 MP image on a B200 against 12.4 s for H3, and 6.3 s on a 48 GB L40S with the 4-bit set. The graph uses only built-in ComfyUI nodes, so you need a current ComfyUI build rather than a custom node pack.

Prompting, per Pruna. Open by saying what the image is ('a frame from an anime TV series', 'a documentary photograph'), describe one scene with a viewpoint, one subject and a light source, put any on-image text in quotes, and drop hype words like '8k' or 'masterpiece', which they say make results glossy. For edits, describe the finished image and say what must stay the same. Pruna lists hands, small background figures and dense text as weak spots, especially below 8 MP.

Licence. The model inherits the MiniMax H3 Community License, which per Pruna excludes the EU, UK, South Korea and the US unless you have a separate licence from MiniMax, and requires a 'Powered by MiniMax H3' credit. The upsampler and text-encoder parts are Apache-2.0. Read it before using P-H3-Image in client work.

Our take: if you already animate with H3 or the Pika H3 Flex adapters, stills from the same model family should hand off to video more cleanly than mixing in another image model's look. The licence is the catch for US and European studios, the same one that applied to H3 itself.