P-H3-Image: Pruna's Open Image Model From MiniMax H3

What is P-H3-Image?

P-H3-Image is an open image generation and editing model from Pruna AI, announced Nov 3, 2026. It is built from MiniMax-H3, a video model from MiniMax, and fine-tuned to output a single frame. Pruna calls it a video model turned into an efficient open image model.

@multimodalart, who shared Pruna's announcement, describes it as converting a video model back to an image model. The poster says the aesthetics feel fresh and that some outputs are ones they cannot get from other models. That is one person's impression, not a benchmark.

Pruna says it reworked every component of H3: the prompt upsampler, the text encoder, the diffusion schedule, the transformer and the decoder.

P-H3-Image release date and availability

P-H3-Image is available now. Pruna published open weights on Hugging Face on Nov 3, 2026, in two repositories: one with the Diffusers pipeline and one with ComfyUI files. A tech report and a hosted demo on Hugging Face Spaces are also referenced.

ComfyUI support is on the master branch and ships with the next stable release. Until then, users need the nightly build from GitHub. The ComfyUI desktop app and Comfy Cloud follow stable releases, so they get it later.

Pruna has not published a hosted API or per-image pricing for P-H3-Image. Its video model, P-Video 2 Pro, is already sold through several platforms, so hosted availability for the image model is plausible but not announced.

P-H3-Image key features and specs

P-H3-Image generates in 8 steps without classifier-free guidance, renders up to 8 megapixels per the model card (Pruna's announcement says 2 to 16 MP), and edits with up to 8 input images. It includes a prompt upsampler based on Qwen3.5-9B that is on by default.

A photo retoucher's desk with a high resolution portrait on screen and printed proofs

Supported 8 MP sizes include 2912 x 2912 (1:1), 3360 x 2496 (4:3), 3872 x 2176 (16:9) and 2176 x 3872 (9:16). Pruna says images at 2 and 4 MP can be good, but artifacts become more frequent in faces, hands and small details. For ads, that points to 8 MP as the working size.

Pruna says the model is strongest at photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. It also says the model is sensitive to prompting and renders full, specific prompts better than short requests, which is why the upsampler exists.

P-H3-Image hardware requirements and ComfyUI files

The Diffusers pipeline needs about 101 GB of GPU memory at 8 MP in bf16 with the upsampler, or 83 GB without. Pruna names one H200 or B200, or 80 GB with offloading. The ComfyUI files are far lighter.

The prompt upsampler is a separate download. Pruna says the int8 set stays very close to bf16, while w4a8 keeps quality but changes more details and sometimes the composition. On a 48 GB L40S, Pruna says w4a8 renders a 2048 x 2048 image in 6.3 s instead of 19 s.

Pruna also ships a 2.0 GB LoRA for MiniMax-H3 and four workflows: text-to-image, edit, and LoRA versions of each. They use only ComfyUI's own nodes, per the model card.

P-H3-Image vs MiniMax-H3 and other models

Against its parent, Pruna claims 5x faster generation and a model 7x smaller. The ComfyUI card puts the smallest P-H3-Image set at 20 GB against 143 GB of weights for MiniMax-H3. These are Pruna's figures, not independent tests.

H3 itself is a video model. Pruna's P-Video 2 Pro, a separate product built on H3, makes 5 to 15 second clips at 480p or 768p, according to each::labs. P-H3-Image is a different release aimed at stills.

No head-to-head comparison against other image models appears in the available material. Judge it on your own product shots and prompts before swapping it into a pipeline.

How to use P-H3-Image for AI video and static ads

For ad teams, the fit is static creative and video starting frames: product-in-scene shots, editorial looks, and graphic layouts with a short headline. Multi-image editing, where the first image is edited and others are references, suits swapping a product into a scene.

A product photo set with reference cards and a laptop showing a node-based workflow

Practical notes from the model card:

• Open prompts with what the image is, such as a photograph by a street photographer or a two-color risograph print.
• Spell out every word of any on-image text, or it can turn into gibberish.
• Use 8 MP for text-to-image and 4 or 8 MP for edits.
• Check the license before shipping paid ads, since terms are not stated here.

Teams that do not run GPUs can compare image models in the AdGraam AI Playground, or build static ads from a product URL with the Creative Generator. For process, see how to make static ads with AI. A stills model like this could also feed first frames into a Flow Builder workflow once hosted versions appear.

What to watch next

Watch for the license terms, hosted API availability and independent comparisons. Pruna has not said whether a hosted endpoint is planned, or at what price.

Also watch the ComfyUI stable release that includes native support, and community tests of the w4a8 files on consumer GPUs. Whether the 21 GB footprint holds up on real ad prompts is the question that decides if small teams can run it locally.