News

Pruna AI converted video model MiniMax-H3 into image model P-H3-Image

Compiled by Sergey KostenchukPublished 2026-11-04Updated 2026-11-04

2026-11-04 Multimodal HuggingFace

Expanded analysis for this story

Open the longform version with context, source trail, and what changed.

Hero image of the Pruna AI P-H3-Image launch blog post — P-H3-Image: Turning a SOTA Video Model into an Efficient Open Image Model Source

🎨 Pruna AI converted video model MiniMax-H3 into image model P-H3-Image

Open-weights model for image generation and editing: transformer compressed from 33B to 20B, encoder from 32B to 4B, diffusion from 50 to 8 steps via DMD. Renders up to 8 MP without CFG, editing based on one photo or up to 8 references.

🌍 Component-wise distillation yielded a 7x smaller size (20 GB vs. 143 GB) and 5x speed (2.6s vs. 12.4s for 4 MP on B200). What previously required H200 now runs in ~21 GB VRAM on Ampere+ and RDNA3/4.

👤 Weights on Hugging Face, native support in ComfyUI without custom nodes. Pros — photorealism, documentary photography, movie frames; cons — hands, fine details below 8 MP, dense text. The license restricts commercial use in the US, EU, UK, and South Korea without a separate agreement.

Source 1: https://www.pruna.ai/blog/oss-p-h3-image-launch Source 2: https://huggingface.co/PrunaAI/P-H3-Image

Sources

Story

On November 3, 2026, Pruna AI released P-H3-Image, an open-weights image generation and editing model based on the MiniMax-H3 video model: the transformer was reduced from 33B to 20B parameters, the encoder from 32B to 4B, and diffusion from 50 to 8 steps. P-H3-Image demonstrates a recipe for converting a heavy video model into an efficient image model: component-wise distillation (encoder 32B→4B, backbone 33B→20B, 50→8 diffusion steps, retrained decoder) results in a 7x smaller size (20 GB vs. 143 GB for MiniMax-H3) and a 5x speed increase (2.6s vs. 12.4s for 4 MP on B200) while maintaining q