P-H3-Image: Fast Image Generation Based on MiniMax H3

Pruna AI has released P-H3-Image, an open-source image generation and editing model built on the MiniMax H3 video architecture.

Model

Developers have significantly optimized the original video model's architecture. The 32-billion-parameter text encoder was replaced with a distilled 4-billion-parameter alternative, and transformer parameters were reduced from 33 billion to 20 billion. The decoder was retrained to output a single sharp image instead of a sequence of frames.

Generation speed was improved by redesigning the diffusion schedule: steps were reduced from 50 to 8 via distillation. An enhanced prompt improver based on the Qwen3.5-9B model was also integrated.

Performance

According to developers, generating a 4 MP image on an NVIDIA B200 takes 2.6 seconds, compared to 12.4 seconds for the original H3. For 8 MP resolution, the time is 5.8 seconds versus 30.6 seconds.

The minimum quantized weight size has been reduced to 20 GB, down from 143 GB in the original MiniMax. The W4A8 version requires approximately 21 GB of VRAM for high-resolution image generation.

Capabilities

The model supports resolutions from 2 to 16 MP, with 8 MP being optimal. Users can use up to eight source images to combine characters and objects into a single composition.

• Available via the Diffusers library
• Supported in ComfyUI 0.44.0 without third-party nodes
• Ready-made workflows and quantized weights published