P-H3-Image open weights: Pruna’s MiniMax-H3 image model.

On 3 November 2026, Pruna released P-H3-Image, an open-weight model for text-to-image generation and image editing built on MiniMax’s open MiniMax-H3 video model. Pruna says it reworked H3’s prompt path, text encoder, diffusion schedule, transformer and decoder so the result is much smaller and faster than H3 while keeping a photoreal and cinematic look. Here is what Pruna and the Hugging Face model card confirm about P-H3-Image, how to try it, and what the licence allows.

What is P-H3-Image?

P-H3-Image is Pruna’s open-weight image model derived from MiniMax-H3. According to Pruna’s launch post and the Hugging Face model card↗ (opens in a new tab), it does text-to-image and image editing, samples in 8 steps without classifier-free guidance, and targets high-resolution stills rather than video.

Pruna says the model can render images from about 2 to 16 megapixels, with 8 megapixels recommended for quality. The Hugging Face card lists 8 MP canvas sizes such as 2912×2912 for 1:1 and 3872×2176 for 16:9. For edits, the model card says you can pass 1 to 8 input images: Picture 1 is edited and the rest are references.

Pruna’s reason for starting from a video model is consistency: a video model already has to learn how objects look from different angles and how light falls on materials. It says MiniMax-H3 could emit single frames, but those frames were soft and the full stack was too large and slow for most local setups.

What’s new versus MiniMax-H3

Pruna says it changed every major stage of the H3 stack for still images:

• Prompt upsampler. A small upsampler based on Qwen3.5-9B expands a short request into a prompt H3 understands; Pruna says it keeps prompts shorter and images more natural than glossy enhancers.
• Text encoder. H3’s 32B text encoder is distilled to a 4B encoder that Pruna says weighs about 9 GB instead of 67 GB while keeping similar image quality.
• Diffusion. The video-tuned schedule is adapted for still detail, and Pruna says DMD distillation cuts sampling from 50 steps to 8 — about 5× faster on its B200 comparisons.
• Transformer and decoder. Pruna’s launch blog says a smaller timestep block drops the backbone from about 33B to 20B parameters, though the Hugging Face model card lists a 33.1B-parameter transformer; separately, a new single-frame decoder replaces H3’s video decoder, which produced blurry stills.
• Quantization. Components are quantized to 8-bit and 4-bit; Pruna says the smallest pack is about a 20 GB download versus about 143 GB for H3.

Speed and memory Pruna reports

• On one B200 in bf16, Pruna reports 2.6 s for a 4 MP image with P-H3-Image versus 12.4 s for MiniMax-H3, and 5.8 s versus 30.6 s at 8 MP.
• Download size: MiniMax-H3 143 GB, P-H3-Image bf16 55 GB, 4-bit pack 20 GB (Pruna’s figures).
• Peak VRAM at 8 MP: about 56 GB bf16 and about 21 GB for the 4-bit ComfyUI set, per Pruna.
• The Hugging Face Diffusers card lists roughly 101 GB at 8 MP with the prompt upsampler on (about 83 GB without), or about 76 GB peak on an 80 GB GPU with auto offload.

Who P-H3-Image is for

Pruna positions P-H3-Image for people who want open weights for photoreal and editorial stills, film-like frames, anime TV frames, drawings and short on-image text, and who can run either a large GPU with Diffusers or a smaller card with quantized ComfyUI files. It is a poor fit if you need video, dense body copy in the image, or if you are in a region the MiniMax H3 Community License excludes without a separate MiniMax licence.

How to try P-H3-Image

1. Read the licence first. P-H3-Image inherits the MiniMax H3 Community License; Pruna says it excludes the European Union, the United Kingdom, the Republic of Korea and the United States unless you have a separate licence from MiniMax. The LICENSE↗ (opens in a new tab) says commercial products or services must prominently display “MiniMax H3” in their user interface (Section IV.2), while a “Powered by MiniMax H3” notice is only encouraged (III.3a); companies above $20M in yearly revenue need separate written permission from MiniMax (IV.1); and outputs can’t be used to improve other AI models (V.3). Qwen-based parts of the text encoder and prompt upsampler are under Apache 2.0 per Pruna.
2. For Diffusers, follow the install steps on the model card and load `PrunaAI/P-H3-Image` with `ModularPipeline`. Generate at an 8 MP size from the card’s table (default 2912×2912). Pass `image` or a list of images to edit.
3. On smaller GPUs, use PrunaAI/P-H3-Image-Comfy↗ (opens in a new tab). Pruna says ComfyUI supports the model natively with bf16, 8-bit and 4-bit sets — though per the P-H3-Image-Comfy card, that support is currently on ComfyUI’s master/nightly branch and arrives in the next stable release — and that the 4-bit set can run 8 MP images in about 21 GB of VRAM.
4. Leave the prompt upsampler on unless you write full prompts yourself. Pruna advises naming what the image *is*, describing one scene with light, quoting short on-image text, and skipping hype words like “8k” or “masterpiece”.

What people are building with it

P-H3-Image just shipped, so The Model Atlas does not yet list builds that use this exact checkpoint. Creators who already ship stills and edits with other recent image and video stacks will be the natural audience as local ComfyUI and Diffusers workflows land. Check the blog and the models index as we catalogue related open image tools.

Limitations Pruna calls out

• Very short, overly long or hype-heavy prompts can weaken results.
• Hands, small background figures and fine detail can distort, especially below 8 MP.
• Small or dense text (body copy, code) is unreliable; short quoted strings work better.
• Single images only — no video or audio.
• The 4-bit set can change small details and sometimes composition versus bf16, Pruna says.