Pruna Open-Sources P-H3-Image: MiniMax-H3 Video Model Turned Into an Image Generation and Editing Model, 8 Steps, 5x Faster, 7x Smaller

Background and industry pain points

Open image models have improved fast, but quality and runnability rarely come together: top models can weigh over 100 GB and need 50 sampling steps, which is hard for local users and small teams. Pruna AI, a company focused on model compression and inference speed, previously built Wan Image on top of Wan 2.1. Its idea here is that a video model must learn how objects look from many angles and how light falls on materials to keep frames consistent, which makes it a strong base for an image model. MiniMax-H3, open-sourced by MiniMax this summer, can already output single frames, but Pruna's blog says those images come out soft and blurry, and the model is large and slow.

Architecture highlights and mechanisms

Pruna reworked every component of H3: a small prompt upsampler trained on Qwen3.5-9B rewrites short requests into prompts H3 handles well; the 32B text encoder is distilled to 4B; the video-tuned diffusion schedule is adapted to bring out detail, and DMD distillation cuts steps from 50 to 8; the timestep block, about 40% of the transformer, is replaced with a much smaller one, taking the backbone from 33B to 20B parameters; a new single-frame decoder replaces the video decoder that produced blur; finally the components are quantized to 8-bit and 4-bit.

Benchmarks and measured comparisons

No third-party quality leaderboard results were published; the figures below are Pruna's own speed and memory comparisons. On one B200, a 4 MP image drops from 12.4 s to 2.6 s and 8 MP from 30.6 s to 5.8 s; on an L40S (48 GB), 4 MP goes from 19 s (bf16, offloaded) to 6.3 s (4-bit). Download size falls from 143 GB to 55 GB (bf16) or 20 GB (4-bit), with peak VRAM at 8 MP of about 21 GB in 4-bit. Pruna also lists limits: hands, small background figures and fine details may distort below 8 MP, and small or dense text such as body copy or code is unreliable.

Developer adoption and getting started

Weights are on Hugging Face at PrunaAI/P-H3-Image as a Modular Diffusers pipeline with the prompt upsampler on by default; ComfyUI users can download bf16, 8-bit or 4-bit sets from P-H3-Image-Comfy with no custom nodes. Pruna recommends 8 MP for best quality and 4 MP for speed. When prompting, start by stating what the image is, describe one scene, quote any text that must appear and keep it short, and skip hype words like "8k" or "masterpiece". Read the license before commercial use: the EU, UK, South Korea and US are excluded by default.