Pruna turns MiniMax H3 into a 5x faster open-weight image model
P-H3-Image generates and edits 2–16 MP images, while its smallest build is a 20 GB download and uses about 21 GB of VRAM at 8 MP.
Pruna AI has released P-H3-Image, an open-weight image generator and editor derived from the MiniMax H3 video model. It creates images from 2 to 16 megapixels and can edit or combine up to eight input images.
The company rebuilt the prompt upsampler, text encoder, diffusion schedule, transformer and decoder for single-frame output. Distillation reduces generation from 50 diffusion steps to eight. In Pruna benchmarks on one Nvidia B200, a 4 MP image took 2.6 seconds instead of 12.4 seconds for MiniMax H3; at 8 MP, the comparison was 5.8 seconds versus 30.6 seconds.
The smallest quantized package is a 20 GB download, down from 143 GB for H3, and Pruna reports about 21 GB of peak VRAM for an 8 MP image. That makes the 4-bit build practical on a single high-end consumer or workstation GPU. Full-precision inference remains much heavier. The model is available through Diffusers, while ComfyUI has native support for the quantized versions.
P-H3-Image is strongest at documentary and editorial photography, film stills and graphic styles with short quoted text. Pruna warns that hands, small background figures and dense text can still fail, especially below 8 MP.
Open-weight does not mean unrestricted open source here. The MiniMax H3 Community License excludes the European Union, United Kingdom, South Korea and United States unless users have a separate MiniMax license, and products must display a MiniMax attribution.
