Summary:
Pruna's image model is available as weights for Diffusers and ComfyUI. Developer timing results show speed gains, while territorial licensing and configuration differences require careful checks.Pruna AI has released P-H3-Image, a downloadable image-generation and editing model derived from MiniMax-H3. Its November 3 announcement focuses on reducing the cost of turning a video model into a still-image workflow. For US users, however, the first deployment question is licensing: the underlying community agreement excludes the United States unless separate authorization is obtained.
Open weights come with territorial limits
The repository's MiniMax H3 Community License excludes the US, European Union, United Kingdom and Republic of Korea from its applicable territory. It invites interested users in those regions to seek a separate license. A public download link therefore does not establish permission to deploy there.
The agreement also sets additional commercial conditions, including prior written authorization above its $20 million annual-revenue threshold. Teams should review the actual license for their intended product and location before budgeting implementation.
Two download routes with different hardware needs
The Diffusers model card describes an eight-step pipeline for text-to-image generation and editing with up to eight input images. Its prompt upsampler expands short requests before rendering. The documented bf16 configuration needs about 101 GB of GPU memory at 8 MP with that upsampler, or 83 GB without it.
The separate ComfyUI repository offers bf16, int8 and w4a8 files. Its table reports 56 GB, 31 GB and 21 GB of peak VRAM respectively at 8 MP. Those are configuration-specific figures, not a promise that every pipeline fits a 24 GB graphics card. The smallest model, encoder and VAE set totals 20 GB on disk; the prompt upsampler is listed separately.
What the speed comparison measures
Pruna reports the following timings on one NVIDIA B200 using bf16. Lower is better; the fastest value in each row is bold.
| Image size | MiniMax-H3 | P-H3-Image |
|---|---|---|
| 4 MP | 12.4 seconds | 2.6 seconds |
| 8 MP | 30.6 seconds | 5.8 seconds |
The comparison uses 50 steps for MiniMax-H3 and eight for P-H3-Image, without further optimizations. It measures the reported generation-time benefit of that change, not equivalent image quality. Pruna's L40S results additionally change precision and offloading, so they should not be folded into the same hardware comparison.
Check the workflow, not just the download size
The ComfyUI card places support on the project's master branch and says stable desktop and cloud releases follow later. Confirm support in the exact installed version rather than assuming every ComfyUI distribution already includes it.
The Diffusers card recommends 8 MP and warns that lower resolutions make faces, hands and small details less reliable. Its documented editing limit is eight input images, while parts of the launch post mention nine. Use the documented pipeline limit when planning a workflow and verify behavior before relying on the larger claim.
For teams able to satisfy the license, the useful next step is a small evaluation using their own approved image-editing tasks, with fixed hardware, resolution and precision. For teams in excluded territories, resolving authorization comes first.
Sources and image credit
- Pruna launch announcement, November 3, 2026: release and developer-reported timing data.
- Diffusers model card and ComfyUI model card: access, memory, workflow and limitations, checked November 8.
- MiniMax H3 Community License: territorial and commercial conditions.
- Original PromptsZone chart from Pruna's reported data. The unaltered official Pruna provider logo is a fallback for the model logo, supplied by its press and media logo kit. The chart is not model-generated imagery.
