Pruna Cuts MiniMax-H3 Into P-H3-Image, an 8-Step Open Image Model That Still Needs a Separate MiniMax License in the US and EU

Pruna AI has released P-H3-Image, an open-weights text-to-image and multi-image editing model distilled from MiniMax’s H3 video system into a package the team says is seven times smaller and five times faster for stills. The November 3, 2026 engineering post and the Hugging Face card for PrunaAI/P-H3-Image describe an 8-megapixel-oriented pipeline that runs in eight diffusion steps without classifier-free guidance, edits with up to eight input images on the model card (the blog says up to nine), and ships a Qwen3.5-9B-based prompt upsampler that is on by default.

Out of the box, H3 can emit single frames, but Pruna says those frames are soft and the video stack is too large for most local image workflows—about 143 GB for the smallest H3 set versus a 20 GB 4-bit P-H3 download. On one NVIDIA B200, Pruna reports 2.6 seconds for a 4 MP image versus 12.4 seconds for H3 at 50 steps, and 5.8 seconds versus 30.6 seconds at 8 MP, both in bf16 without further compilation. Peak VRAM for the 4-bit ComfyUI set at 8 MP is listed around 21 GB; the bf16 Diffusers path with the upsampler needs roughly 100–101 GB at 8 MP, or about 76–83 GB with auto offload on an 80 GB GPU.

Every stage of H3, rebuilt for a still

Pruna walks the pipeline in generation order. A small prompt upsampler, trained from Qwen3.5-9B, expands a short request into a prompt H3-style models render cleanly, favoring natural length over glossy enhancer prose. H3’s 32B text encoder is distilled to a 4B encoder (about 9 GB instead of 67 GB) that is trained to match the large encoder’s signal. The diffusion schedule is retuned away from video’s high-noise settling, and advanced DMD distillation cuts the step count from 50 to 8 while keeping the look. The transformer block that consumed the timestep—about 40% of H3’s transformer—is replaced with a smaller equivalent, dropping the backbone from 33B to 20B parameters without a full retrain. A new decoder maps one latent frame to one sharp image instead of decoding video chunks. Finally, 8-bit and 4-bit quantizations shrink the download; the card lists a 33.1B-parameter transformer plus the 4B encoder, VAE, and upsampler.

Recommended canvas sizes center on 8 MP (default 2912×2912), with tabulated 8 MP and 4 MP sizes for common aspect ratios. Edits treat the first image as the canvas and later images as named references (“Picture 2”, “Picture 3”). Without explicit height and width, an edit scales Picture 1 up to at least 4 MP. The upsampler rewrites edit requests into a keep/change brief; Pruna shows that short typed edits often redraw the whole scene, while keep-lists hold identity.

Where it is strong, and where it breaks

Pruna positions quality at documentary and editorial photography, skin-texture portraits, film stills, anime frames, drawings, prints, and graphic design with short quoted text. Limitations are listed without hedging: hype words and very long or very short prompts degrade results; hands and small background figures distort, especially below 8 MP; dense body copy and code become pseudo-text; and the model does not generate video or audio. Hand-written prompting advice matches the upsampler’s bias—name what the image is in the first sentence, describe one scene and one light source, quote every string that must appear, and skip “8k” / “masterpiece” filler.

The license is not Apache, and geography matters

Weights build on MiniMax-H3 under the MiniMax H3 Community License. Among other terms, that license excludes the European Union, the United Kingdom, the Republic of Korea, and the United States unless the user has a separate MiniMax grant, and products must show “Powered by MiniMax H3”. The text encoder’s Qwen3-VL-4B parts and the Qwen3.5-9B upsampler are Apache-2.0; ComfyUI nodes are MIT. Anyone in a restricted region, or shipping a product that embeds the weights, needs to read NOTICE and the MiniMax terms before download—open weights here does not mean unconstrained redistribution. Diffusers lives at PrunaAI/P-H3-Image; quantized ComfyUI packs at PrunaAI/P-H3-Image-Comfy. For teams that can clear the license, P-H3-Image is a concrete path from an open video backbone to a stills stack that fits a workstation rather than a video cluster.