Pruna strips a video model down to an ultrafast image generator

Training state-of-the-art image models usually means gathering massive datasets and burning through compute. Pruna AI took an entirely different route: they carved a fast image generator out of a massive open-source video model.

The team just released P-H3-Image, an open-weights model built directly on top of the MiniMax-H3 video architecture. By rebuilding nearly every critical block under the hood, they turned a heavyweight 143 GB video engine into an image model that runs in just 8 steps and shrinks down to a 20 GB download.

Stripping down the video engine

Starting from video sounds counterintuitive, but it solves a classic visual problem. To keep sequential frames coherent, video models have to learn how light wraps around three-dimensional objects, how materials behave, and how perspective shifts. Pure image models often struggle with those physical priors.

The catch was that MiniMax-H3 was never meant for single snapshots. Its raw single-frame outputs came out soft and blurry, and running 50 diffusion steps on a 143 GB model was impractical for anyone outside high-end server farms.

Pruna systematically re-engineered the pipeline. First, they distilled the massive 32B text encoder down to 4B, cutting its footprint from 67 GB to just 9 GB. Then, they replaced the timestep processing blocks—which made up 40% of the transformer backbone—with a far lighter equivalent, shrinking the backbone from 33B to 20B parameters without retraining.

To fix the speed issue, they used DMD distillation to slash diffusion steps from 50 to 8. They also swapped out the multi-frame video decoder for a sharp single-frame image decoder, and released 8-bit and 4-bit quantized versions. The smallest 4-bit package sits at 20 GB—roughly a 7x reduction in file size.

5x faster, with 8-reference composition

On hardware, the trade-off pays off immediately. On an Nvidia B200, generating a 4-megapixel image dropped from 12.4 seconds to 2.6 seconds, while an 8-megapixel render takes just 5.8 seconds—roughly five times faster than the original model. In 4-bit mode, peak VRAM for an 8-megapixel generation hovers around 21 GB, bringing it within reach of top-tier consumer GPUs with 24 GB of memory.

The output leans heavily toward film stills, documentary photography, natural skin textures, and anime art. Because the underlying model was built to track references across video frames, P-H3-Image can also blend up to 8 reference portraits or objects into a single cohesive scene based on a short prompt, or handle multi-turn edits.

It still has limits. Tiny text, code blocks, hands, and small background figures can distort easily below 8 megapixels. The weights are up on Hugging Face, and ComfyUI already supports the model natively without custom nodes.

Why it matters

Pruna’s approach points to an attractive shortcut in open-source AI: downcycling frontier video models into efficient image engines. Video backbones have already absorbed the hardest spatial reasoning lessons. Pruning and distilling them gives creators cinematic lighting and perspective without the massive overhead of training an image foundation model from scratch.

For digital creators and prompt engineers, the ability to blend up to 8 reference images in 8 diffusion steps is a massive workflow win, especially now that ComfyUI supports it out of the box.

That said, we think the biggest barrier isn't the hardware requirements, but the licensing terms. P-H3-Image inherits the MiniMax H3 Community License, which explicitly excludes commercial use in regions like the US, EU, UK, and South Korea without a separate agreement. For now, it is a brilliant local experiment for enthusiasts, but enterprise adoption faces a legal hurdle.