Pruna turns MiniMax-H3 into open image model
Pruna AI has released P-H3-Image, an open-weights image generation and editing model built from MiniMax's H3 video model, which it says is seven times smaller and five times faster than the original.
Pruna AI announced P-H3-Image, an open-weights model for text-to-image generation and image editing built on MiniMax-H3, a video model released with open weights earlier this year. It renders images from 2 to 16 megapixels and can edit using multiple input images. The post gives the limit as 9 in one place and 8 references in another.
Pruna's argument is that a video model, trained to keep frames consistent, has already learned how objects, lighting and bodies look and move. Used out of the box, though, H3 produces soft single frames and is too large for most users to run.
The company says it reworked every component of H3:
- Text encoder: distilled from 32B to 4B parameters (9 GB instead of 67 GB).
- Sampling: cut from 50 to 8 steps with DMD distillation, using a schedule adjusted to bring out fine detail.
- Backbone: reduced from 33B to 20B parameters by replacing the timestep block, which made up 40% of the transformer.
- Decoder: retrained to decode a single frame into a sharp image.
- Prompt upsampler: a new model based on Qwen3.5-9B.
On one B200, a 4 MP image takes 2.6 s instead of 12.4 s. The smallest 4-bit set is a 20 GB download, compared with 143 GB for H3, and peaks at about 21 GB of VRAM at 8 MP. Pruna says the 4-bit version can alter small details and sometimes the composition.
Pruna lists its own limitations: hands, small figures and dense text are unreliable, especially below 8 MP. Weights are available on Hugging Face, with a Modular Diffusers pipeline and native ComfyUI support.