← Back to blog
AiAbout 7 min read

Pruna Turned a Video Model Into an Image Model. The Interesting Part Is What It Threw Away.

Published Nov 7, 2026
Pruna Turned a Video Model Into an Image Model. The Interesting Part Is What It Threw Away.

Pruna AI released P-H3-Image on November 3. It generates and edits images, the weights are open, and it is built out of MiniMax-H3, a video model.

That last sentence is the whole design argument. A video model has to learn how an object looks from many angles, how light falls on a material as the camera moves, how a scene stays coherent frame to frame. A single photograph is a degenerate case of that problem. So in theory the hardest part of image generation is already paid for, and you only need to teach the model to render one frame well.

In practice, asking a video model to output one frame produces soft, blurry images. MiniMax's own H3 can do it, badly. Pruna's contribution is taking the model apart and rebuilding each piece for the still-image job.

Four cuts

The text encoder went from 32 billion parameters to 4 billion, using a distilled Qwen3-VL-4B. That single change drops 67 GB of weights to 9 GB, and it is the least visible part of the release.

Four progressively smaller glass prisms nested inside one another on a black surface.

The transformer backbone shrank from 33 billion to 20 billion parameters. The main surgical move was replacing the timestep block, which the blog describes as roughly 40 percent of the transformer, with something much smaller. Timestep conditioning is expensive because it has to modulate the whole network at every denoising step. For video you need fine control across time. For a still you need one number.

The decoder was retrained to produce a single sharp image across a 2 to 16 megapixel range instead of a sequence of video frames.

And the sampling schedule was distilled with DMD, taking generation from 50 steps down to 8.

A small prompt upsampler also sits in front, built on Qwen3.5-9B, rewriting short prompts into something H3 responds to well. It is on by default in the Diffusers pipeline, and it is the reason the bf16 memory numbers look worse than the model actually is.

The numbers, with the vendor label attached

Pruna's measurements: a 4 MP image takes 2.6 seconds on a B200, against 12.4 seconds for MiniMax-H3. An 8 MP image takes 5.8 seconds instead of 30.6. On an L40S with 4-bit weights, 4 MP takes about 6.3 seconds.

Downloads are 55 GB in bf16 or 20 GB in 4-bit, against 143 GB for H3. Peak VRAM at 8 MP is roughly 56 GB in bf16 or 21 GB in 4-bit.

That 21 GB figure is the one that matters for anyone reading this at a desk. It points at 24 GB cards, the 3090, 4090 and 5090, running the 4-bit build at full resolution. On a 16 GB card you will be reducing resolution or offloading.

One warning that is easy to miss: the Diffusers pipeline with the prompt upsampler in bf16 at 8 MP wants about 100 GB of GPU memory. That is not a typo and it is not the model's fault. If you are running Diffusers rather than ComfyUI, either use the quantized weights or turn the upsampler off.

The license excludes the countries most likely to use it

P-H3-Image inherits the MiniMax H3 Community License. In practice that means the weights are open to download and closed to deployment in the United States, the European Union, the United Kingdom and South Korea without a separate agreement with MiniMax. Products built on it must display "Powered by MiniMax H3."

This is now a familiar shape. Over the past year, Chinese labs have settled into a pattern: publish weights under a permissive-sounding name, attach a community license with territorial carve-outs, and sell the carve-outs to whoever needs them. H3 has spawned derivatives from Nvidia, Pika, Creatify and now Pruna, and every one of them navigates the same terms.

For an individual developer in Ohio, that means the model is technically downloadable and legally awkward. For a company, it means a compliance review. For a lab in Singapore or Brazil, it means a genuinely free high-quality image model, which is a real redistribution of capability that does not get discussed much in the US press.

There is a second layer worth noting. The license travels with the weights, so a company that fine-tunes P-H3-Image for its own product is still bound by the exclusions. Derivative work does not launder the terms.

Why distill a video model at all

There are already excellent open image models. FLUX 3 Image Klein is 5.6 billion parameters and prices at eight tenths of a cent per image on hosted APIs. Why take a 143 GB video stack and spend engineering time shrinking it to 20 GB?

The answer is in what video training data teaches. A model trained on clips sees the same object under changing light and angle and learns a more complete internal representation of it than a model trained on isolated photos. That shows up in a specific place: multi-reference editing.

P-H3-Image accepts up to eight or nine reference images in one composition, depending on which document you read. Keeping a character consistent across a product shot, a poster and a UI mock is the task where image models usually fall apart, and it is exactly the task that benefits from a representation built on video.

Pruna's own framing is that a video model learns the physical world as a byproduct of keeping consecutive frames coherent. For a still, that knowledge is mostly wasted. For a set of stills that need to agree with each other, it is the point.

The quantized builds are three different products

ComfyUI users get three complete weight sets: bf16, 8-bit and 4-bit. They are not the same model at different file sizes.

At bf16 you get the intended output and need 56 GB of VRAM at 8 MP, which means an A100 or an H100 class card. At 4-bit you fit in 21 GB on a 4090, and you accept whatever the quantization costs in fine detail. The middle tier exists for people who have 40 GB and want to keep the quality rather than the resolution.

The practical advice that follows is to pick your target resolution first, then the quantization that fits, then decide whether the result is good enough. The reverse order, downloading the 4-bit build because it is smallest and then complaining that hands look wrong, is how people conclude a model is bad when they were running a compromise build at a resolution it was not tuned for.

What it is not good at

Pruna states the limits itself. Fine details, hands, and small or dense text can be unreliable. The company positions the model as strongest at photorealistic and documentary photography, film stills, anime, and graphics with short text. If your job is a poster with a paragraph of body copy, this is not the tool.

The benchmark situation is also thin. There is no independent evaluation yet. Every speed and quality number in the release is a vendor measurement, and the comparison set is MiniMax-H3 and Pruna's own earlier Wan Image work, not FLUX 3 or Nano Banana.

How to try it

The Diffusers pipeline is at PrunaAI/P-H3-Image on Hugging Face. There is a separate ComfyUI repository with three complete weight sets, bf16, 8-bit and 4-bit, and native ComfyUI support with no custom nodes. Pick one set, download all four files in it, and put each in the folder the repo names.

If you have a 24 GB card, start with the 4-bit set at 4 MP and the upsampler off. Compare it against whatever you already use on ten of your own prompts, not on the launch gallery. The interesting question is whether the multi-reference editing holds up on your actual assets, since raw speed stopped being a differentiator about a year ago.

Related articles