Pruna releases P-H3-Image with regional license limits

Pruna AI released P-H3-Image on November 3, publishing weights for still-image generation and editing adapted from MiniMax-H3. It reduced sampling from 50 steps to eight and provided separate Diffusers and ComfyUI distributions.

The accompanying MiniMax H3 Community License excludes South Korea, the United States, the European Union and the United Kingdom from its default territory. It invites prospective deployers there to contact MiniMax for separate permission. Publicly downloadable files consequently retain geographic conditions on use.

Pruna reports 2.6 seconds for a four-megapixel image on one NVIDIA B200, against H3's 12.4 seconds. This is its own bf16 comparison of eight versus 50 steps. It also describes retraining the video decoder to reconstruct a sharp single frame.

What the 21 GB figure leaves out

Adding the public file sizes gives approximately 20.1 GB for the smallest image set: 4-bit model and text encoder plus an 8-bit decoder. The recommended 19.3 GB prompt upsampler brings that download to approximately 39.4 GB. These calculated decimal storage totals differ from GPU memory. The upsampler also has approximately 9.7 GB int8 and 6.7 GB 4-bit files; the 39.4 GB total uses recommended bf16.

The ComfyUI card lists peak image-generation memory at eight megapixels as 21 GB for the compact set and 56 GB for bf16. The recommended upsampler peaks at about 21 GB itself and runs first, then leaves GPU memory. Its prompt processing adds about seven seconds on an L40S. Expanded prompts are reused until the request or reference images change.

ComfyUI core shipped the integration in stable release v0.44.0 on November 4. Its release notes directly include the P-H3-Image pull request, advancing beyond the model card’s development-branch and next-release guidance. That core release alone does not establish when desktop-app or Comfy Cloud users receive it.

For editing, the first picture is the canvas and subsequent pictures provide references. The Diffusers card specifies eight input images in total, a default generation size of 2912 by 2912 pixels, and four or eight megapixels for edits. Without explicit dimensions, an edit follows the first picture, scaled to at least four megapixels.

The upsampler expands short requests into descriptions of what to change and retain. The card acknowledges distorted hands and small background figures, unreliable dense text, and more facial and detail artifacts at lower resolutions.

ComfyUI merged its native-support pull request on November 3 at 15:38:13 UTC. Pruna's public file list contains text-to-image and editing workflows, plus two examples applying an additional trained adapter to existing H3 weights. This report examined files and documentation, without running or benchmarking P-H3-Image.

Pruna's NOTICE identifies the image backbone, decoder, text connector and text encoder as H3-based modifications. The encoder's Qwen3-VL-4B portions and the Qwen3.5-9B-based prompt upsampler carry Apache 2.0 terms. These component families share a repository while retaining different permissions. The notice points to the separate license files covering each part.

The H3 agreement also covers distilled derivatives. Commercial products and services above $20 million in annual revenue require prior written authorization, and commercial interfaces must display “MiniMax H3.” Territorial permission and these commercial conditions appear together in the license supplied with the weights.