Pruna AI Turns MiniMax H3 Video Model into an Image Generator: 2.6 s Instead of 12.4 s at 4 MP

Pruna AI took the heavyweight MiniMax H3 video model and turned it into P-H3-Image, a fast image generator and editor with open weights. The problem is that video generators are powerful but slow and resource-hungry: getting a single image out of H3 takes hundreds of gigabytes of memory and tens of seconds of waiting. Now, according to Pruna's own measurements, a 4 MP frame (megapixels; roughly 2000×2000 pixels) on an NVIDIA B200 is ready in 2.6 seconds, versus 12.4 seconds for the original H3. For developers, this means the idea of "using a video model as the basis for images" can now run without an entire data center.

Why a video model works for images

There had been talk for a while that an image generator could be built on top of MiniMax, but MiniMax itself stopped at promises. The logic is straightforward: to render a plausible clip, a video model learns during training how an object looks from different angles and how lighting, perspective and motion change. That knowledge can be applied to a single frame as well.

But Pruna did not simply ask H3 to output one frame instead of a video. The architecture was substantially reworked:

• The 32-billion-parameter text encoder (the part that turns your prompt into a representation the model can work with) was replaced with a distilled one, i.e. "compressed" to 4 billion parameters;
• The main transformer was reduced from 33 to 20 billion parameters;
• The decoder was retrained to produce a single sharp image rather than a sequence of video frames;
• The diffusion schedule was reworked, and distillation cut generation from 50 to 8 steps; a step is one pass of the model over the image, so fewer steps mean less waiting;
• A prompt enhancer based on Qwen3.5-9B was added, which expands a short request into a more detailed one.

Speed and memory

On an NVIDIA B200, a 4 MP image is generated in 2.6 seconds versus 12.4 for the original H3. At 8 MP the gap is just as noticeable: 5.8 seconds instead of 30.6. That is roughly a fivefold speedup. An important caveat: the figures come from the developers themselves, there are no independent tests yet, and the B200 is a very expensive card, so on more modest hardware the times will be longer.

Memory requirements have changed too. The minimal set of quantized weights (that is, stored at reduced precision to take up less space) occupies 20 GB, whereas the original MiniMax H3 takes 143 GB. The W4A8 version, where weights are stored in 4 bits and activations in 8, needs about 21 GB of VRAM when generating at 8 MP. Quantization usually costs some share of quality, so before using it in a product it makes sense to compare results on your own prompts.

What it can do and how to try it

The model is stated to support resolutions from 2 to 16 MP, and the developers call 8 MP optimal. Its strong suit is working with references: you can pass up to eight source images and combine characters, objects and environment elements in a single composition. According to the authors, this ability is inherited from H3's training on video references. In practice it looks like this: you upload a hero character, a product and a background, and get a unified scene without manual compositing.

You can run the model through Diffusers. Native support without third-party nodes arrived in ComfyUI 0.44.0 on November 4. Pruna has published ready-made workflows and quantized weights, so you can get started without digging deep into code. It is worth reading the license terms on the model's Hugging Face page, especially if you plan commercial use.

This is not the team's first experiment of this kind: Pruna previously created Wan Image the same way, based on the Wan 2.1 video model. The practical takeaway: if you need an open image generator and editor with multiple references that fits in roughly 20 GB of VRAM, P-H3-Image is worth testing on your own tasks, keeping in mind that its speed and quality are so far confirmed only by the developers' own data.