What if one of the most powerful AI video models could become an incredibly fast image generator?
Not by simply asking it to generate a single frame.
By rebuilding its architecture specifically for image generation.
Meet P-H3-Image, a new open-weight AI image generation and editing model from Pruna AI, built on top of MiniMax H3 Video.
And the engineering behind it is fascinating.
Here's the idea.
When an AI model learns to generate videos, it also learns something about how the visual world works.
Objects from different angles. Changing perspectives. Lighting. Movement. Spatial relationships.
So why not take that knowledge and use it to generate better still images?
That's exactly what Pruna decided to explore.
But they didn't just modify the output format.
They substantially redesigned the original model:
Smaller text encoder: Distilled the original 32B encoder into a 4B version.
Optimized transformer: Reduced the main processing architecture from 33B to 20B parameters.
New image decoder: Retrained the decoder to produce sharp single images instead of video frames.
8-step generation: Reduced the original 50 diffusion steps to just eight.
Smarter prompting: Added an automatic prompt enhancer based on Qwen3.5-9B.
Quantized weights: Reduced the smallest model download from 143 GB to approximately 20 GB.
And the performance numbers are pretty wild.
According to Pruna's benchmarks on an NVIDIA B200:
At 4 megapixels:
Original MiniMax H3: 12.4 seconds
P-H3-Image: 2.6 seconds
At 8 megapixels:
Original MiniMax H3: 30.6 seconds
P-H3-Image: 5.8 seconds
That's roughly 5× faster, with a much smaller memory footprint.
Of course, these are developer-reported measurements, not independent benchmarks.
And here's what P-H3-Image can actually do:
Text-to-image: Generate images from 2 to 16 megapixels.
Image editing: Modify existing images using natural-language instructions.
Multi-reference composition: Combine up to eight source images in a single scene.
Photorealistic visuals: Generate portraits, editorial photography, cinematic scenes, and realistic environments.
Creative styles: Produce illustrations, anime frames, drawings, prints, and graphic designs.
Local inference: Run the quantized model through ComfyUI.
Pruna recommends 8 MP as the sweet spot for image quality, although lower resolutions can be faster.
And yes, you can run it without a B200.
The smallest W4A8 configuration requires approximately 21 GB of VRAM for 8 MP generation.
That's still a substantial hardware requirement, but considerably more accessible than the original video model.
And here's another interesting detail.
P-H3-Image now has native ComfyUI support.
Starting with ComfyUI 0.44.0, released November 4, 2026, you can use the model without installing third-party custom nodes.
Pruna has also published ready-made workflows, quantized weights, and a Diffusers pipeline.
There are limitations, of course. Hands, small background figures, and fine details can still produce artifacts, especially at lower resolutions. Dense text remains unreliable.
And the model is open-weight under the MiniMax H3 Community License, not an unrestricted open-source license.
But here's what I find most interesting.
We're starting to see AI models being repurposed across different creative disciplines.
A video model becomes an image generator.
The knowledge learned from generating motion becomes useful for understanding still images.
And instead of training everything from scratch, developers can build on existing models, specialize them, compress them, and make them dramatically more efficient.
Pruna already explored a similar approach with Wan Image, based on Wan 2.1.
Now they've done it again with MiniMax H3.
Maybe the next big image model won't actually start as an image model at all.
And that's a pretty interesting direction for generative AI.
Official technical announcement:
https://www.pruna.ai/blog/oss-p-h3-image-launch
Model weights and Diffusers:
https://huggingface.co/PrunaAI/P-H3-Image
ComfyUI weights and ready-made workflows:
https://huggingface.co/PrunaAI/P-H3-Image-Comfy
ComfyUI 0.44.0 release:
https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.44.0
#PH3Image #PrunaAI #MiniMaxH3 #ComfyUI #AIImageGeneration #GenerativeAI #OpenWeights
YouTube / ABVX