Pruna AI Releases P-H3-Image, an Image Model Derived from MiniMax H3 That Runs in 21 GB of VRAM

Pruna AI, a specialist in model compression, released P-H3-Image on November 3, 2026, an open-weights image generation and editing model. Its key feature: it was not trained from scratch, but derived from MiniMax H3, an open video model released this past summer. The announced result: images generated in 8 steps instead of 50, with a 4-bit version that fits in roughly 21 GB of VRAM. However, its license, inherited from MiniMax, leaves Europe and the United States on the sidelines.

Why Start from a Video Model?

Pruna’s argument is straightforward: to maintain consistency from frame to frame, a video model must learn how an object looks from multiple angles, how light interacts with materials, and how bodies move. That makes it an excellent foundation for an image model. While MiniMax H3 can already produce still images, they tend to turn out blurry, and the model is too heavy and slow for most users.

Pruna claims to have revamped every component:

• The prompt: a small model based on Qwen3.5-9B rewrites a one-line request into a detailed prompt (enabled by default), as H3 is highly sensitive to phrasing;
• The text encoder: scaled down from 32 to 4 billion parameters through distillation (9 GB instead of 67);
• Diffusion: a schedule tailored for fine details alongside distillation that slashes the process from 50 to 8 steps, with no CFG guidance needed;
• The decoder: a new decoder that turns a single frame into a crisp image, whereas H3’s decoder—built to decode multiple frames—yields blurry still images.

The model also edits images and blends multiple reference images into a single scene, a capability inherited from H3’s video training.

System Requirements

This is the core appeal of the release. The full diffusers pipeline in bf16 requires roughly 101 GB of GPU memory at 8 megapixels, demanding an H200 or a B200. However, the ComfyUI files are available in three precisions: bf16 (55 GB, about 56 GB of VRAM), 8-bit (30 GB), and 4-bit (20 GB, about 21 GB of VRAM). ComfyUI natively supports it without custom nodes, currently via its development release. Pruna cautions that while the 4-bit version “preserves quality,” it may alter subtle details or even overall composition.

Regarding performance, all figures come directly from Pruna: 2.6 seconds for a 4-megapixel image on a B200, down from 12.4 seconds for H3, and 6.3 seconds on a 48 GB L40S card in 4-bit precision. The developer acknowledges several limitations: hands and small background figures can sometimes appear distorted, especially below 8 megapixels, and small or dense text remains unreliable (only short text enclosed in quotation marks works consistently). Furthermore, its documentation shows discrepancies: the blog post cites 2 to 16 megapixels and up to 9 input images, while the Hugging Face model card specifies 8 megapixels and up to 8 images.

A License That Excludes Europe and the US

The weights are not under an open-source license. Just like Pika H3 Flex, another MiniMax H3 derivative, they fall under the MiniMax H3 Community License. According to Pruna’s documentation, this license excludes the European Union, the United Kingdom, the Republic of Korea, and the United States unless a separate license is obtained from MiniMax, and mandates displaying “Powered by MiniMax H3” on products using it. Only the text encoder (specifically its Qwen3-VL-4B parts) and the prompt rewriter are released under Apache 2.0, while the ComfyUI nodes are licensed under MIT.