P-H3-Image

P-H3-Image is an open-weights image generation and editing model released by Pruna AI on 2026-11-03, rebuilt from MiniMax-H3. Pruna reworked the prompt upsampler, text encoder, diffusion schedule, transformer, and decoder so the model is about 7× smaller (~20 GB vs ~143 GB) and about 5× faster (~2.6s vs ~12.4s for a 4MP image on one B200 at 8 steps).

It targets 2–16MP outputs and edits with up to 8–9 reference images, with strengths in photorealism, editorial/documentary looks, film stills, and short quoted text. Weights: PrunaAI/P-H3-Image (plus ComfyUI pack and HF Space).

Problem

Running a raw video model as an image generator is often soft and too heavy. P-H3-Image keeps H3’s appearance priors while shipping a practical open image stack for Diffusers/ComfyUI.

Differentiation

• Versus MiniMax-H3: same lineage, much smaller/faster for stills and edits.
• Versus typical T2I: multi-reference editing and video-backbone consistency cues.

Limits

• MiniMax H3 community license—check commercial terms.
• Early reports note occasional gradient banding.
• Image gen/edit focus, not a video model replacement.

Who it’s for

Creators and tool teams wanting open, self-hostable image gen/edit with video-model-like consistency.

Get started

1. Read the tech report/blog.
2. Load PrunaAI/P-H3-Image via Diffusers or use the ComfyUI pack.
3. Try the HF Space before wiring into production pipelines.