Pruna's P-H3-Image Makes 8-Megapixel Images in 8 Steps, but Its License Excludes the US, EU, UK and South Korea

Pruna AI's P-H3-Image is an open image model that takes MiniMax's H3 video model and turns it into a fast image generator and editor. The Hugging Face repository was created on 27 October, and the ComfyUI version had about 704,000 downloads and 259 likes on 9 November, according to Hugging Face's listing. Most write-ups cover the speed. The license matters more for anyone who wants to use it.

What the model card says it does

What it's good at, and where it breaks

• Strongest at photorealism, film and editorial looks, and graphic styles with short quoted text. Styles Pruna highlights include risograph prints, 1960s colour photographs and TV-frame anime.
• Needs detailed prompts. A built-in Qwen3.5-9B upsampler rewrites a one-line request on every run. Pruna says hype words such as "epic", "cinematic" and "8k" push results toward a glossy, oversaturated look.
• Text is unreliable beyond short strings. Small or dense text turns into pseudo-text, and longer passages sometimes get single letters wrong.
• Quality depends on resolution. At 2 and 4 MP, faces, hands and small details break more often; Pruna recommends 8 MP.
• Hands can break, and small background figures can come out distorted.

The license is the catch

P-H3-Image is released under the MiniMax H3 Community License Agreement, which is dated 2 August 2026. The agreement defines its "Applicable Territory" as worldwide, excluding the European Union, the United Kingdom, the Republic of Korea and the United States. It says you may not use, modify, distribute or display the model or its outputs outside that territory, and the Pruna card says products built on it must show "Powered by MiniMax H3". The prompt upsampler and parts of the text encoder, which are Qwen-based, are under Apache 2.0.

That is stricter than the usual "open weights" label suggests. A designer in the US, a studio in Germany or an agency in London would need a separate license from MiniMax before using it, and the agreement's wording covers outputs as well as the model. The weekly roundup from Kombitz noted the same exclusions. We read the license text directly, but it's a legal document, so check it with counsel before using the model commercially.

What this means if you generate images for work

• Check where you operate before you download. The territory list covers the largest markets for most commercial teams.
• If you're allowed to use it, the 4-bit ComfyUI files are the realistic route. The full model wants a 100 GB-class GPU.
• For anything with text in the image, plan for short quoted strings only, and check every letter.
• Hosted tools that state their commercial terms plainly are easier to clear with a legal team than a model with a territory carve-out.