AI News
All stories
Reported

Pruna AI open-sources P-H3-Image, an image model built from MiniMax H3 video model

Pruna AI says its open P-H3-Image generates or edits images five times faster in eight steps, and is seven times smaller than the H3 video model it is built from.

Models · · Updated · 1 min read

Pruna AI has released P-H3-Image, an open-weights image generation and editing model built from MiniMax H3, a video model. The weights are on Hugging Face, and the company published a technical report on its blog. The company says the model generates or edits images five times faster, using eight steps where the original uses 50. Matthew Berman, who covered the release on his Forward Future channel, cited a 4-megapixel image taking 2.6 seconds against 12.4 seconds for the original.

The speedup comes from cutting the number of denoising steps, the iterations a diffusion model runs to turn noise into an image. Fewer steps means fewer passes through the network for each image. Berman's side-by-side comparison of H3 at 50 steps against P-H3-Image at eight steps shows aliens and humans in space. He said the outputs looked almost identical.

The size claim is less clear. Pruna says the model is seven times smaller than H3, but the file sizes Berman read out are 143 GB for H3 and 55 GB for P-H3-Image at BF16, with about 30 GB at int8. Berman's reading is that high-resolution images become a light download and run on a much smaller GPU. That depends on which of those measures is meant.

Pruna lists its strengths as photorealism, documentary and editorial photography, film stills and graphic styles with short quoted text. Berman showed a photorealistic sample from the upsampler stage using the prompt an old fisherman at sunrise. He said the handwritten examples were convincing, including pencil and ink lettering and a coffee stain in one corner.

What changed

Pruna AI released P-H3-Image, an open-weights image generation and editing model built from MiniMax H3.

Type: New capability · Applies to: P-H3-Image

Why it matters

Companies generating images at high resolution could cut GPU time and download size if the speed and quality claims hold up in their own tests.

What you can do

Teams running image pipelines could test the model against their current generator on latency and quality for their own prompts.

How the story developed

  1. The commentary adds the specific speed timings, model file sizes and the presenter's quality comparison with MiniMax H3.

Original source

We have no link to the original source for this story yet.

Where we saw it

1 source

Forward Future (Matthew Berman)

MiniMaxPruna AI

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works