P-H3-Image (H3 Image): Specs, Setup, Price and License

P-H3-Image is an open-weight image model from Pruna AI, built on MiniMax's H3 video model and released on November 3, 2026. It makes still images from text and edits them with up to 8 input images, renders 8-megapixel pictures in 8 steps without CFG, and ships with a built-in prompt upsampler. The smallest ComfyUI set is a 20 GB download that peaks at about 21 GB of VRAM, and WaveSpeed hosts it from $0.02 per image. The catch is the license: like MiniMax H3 itself, it does not cover the United States, the European Union, the United Kingdom or South Korea, and that includes using or displaying its images there.

Verification note: We read Pruna's model cards for the diffusers and ComfyUI versions, the license file, Pruna's launch post and WaveSpeed's model pages on November 12, 2026. We did not run P-H3-Image or show any of its images on this page. Its license does not allow using or displaying outputs outside its territory, and this page is published worldwide. Kensa runs MiniMax H3 for video, not P-H3-Image.

P-H3-Image specs

The 8-megapixel sizes Pruna recommends are 2912 × 2912 (1:1), 3360 × 2496 (4:3), 3552 × 2368 (3:2) and 3872 × 2176 (16:9), plus their portrait versions. Pruna says 2 and 4 MP images work too, but faces, hands and small details break more often below 8 MP.

How a video model became an image model

MiniMax H3 is a video model; P-H3-Image is Pruna's rebuild of it for single frames. According to Pruna's launch post:

• Fewer steps. Sampling went from 50 steps to 8 through DMD distillation.
• A smaller text encoder. Distilled from 32B to 4B parameters, which cuts its download from 67 GB to 9 GB.
• A lighter transformer. The timestep block, 40% of H3's transformer, was swapped for a smaller equivalent, taking the backbone from 33B to 20B parameters. (The diffusers model card still lists a 33.1B-parameter transformer at 100 GB in bf16, so the two official pages don't agree on this number.)
• A new decoder. Retrained to decode one frame into one sharp image.
• A prompt upsampler. A Qwen3.5-9B-based model rewrites a one-line request into the long, specific prompt H3 renders best.

Pruna's launch post gives these speeds:

Pruna's diffusers model card measures it differently, at about 9 seconds per 8 MP image on a B200, plus 6 to 11 seconds when the prompt upsampler runs. On an L40S, the upsampler takes about 7 seconds per prompt.

Is P-H3-Image free?

The weights are free to download from Hugging Face, so running it on your own GPU costs only the hardware. Pruna also runs a free demo in a Hugging Face Space on ZeroGPU, where Hugging Face's usage limits apply.

Whether you may use it is a separate question. P-H3-Image is a derivative of MiniMax H3 and comes under the MiniMax H3 Community License, which:

• Excludes four territories. The license applies worldwide except the European Union, the United Kingdom, South Korea and the United States. It says you may not use, reproduce, modify, distribute or display the model or its outputs outside that area. People in the excluded territories can ask MiniMax for a separate license.
• Requires a notice. Products built on it must say "Powered by MiniMax H3", and commercial products must show "MiniMax H3" in their interface.
• Caps revenue. Commercial products with more than $20 million in yearly revenue need written permission from MiniMax first.
• Requires safeguards from anyone who lets other people generate with it, including a way to report misuse.
• Leaves outputs to you. MiniMax claims no rights over the images you generate, but forbids using them to improve other AI models.

The Qwen-based parts of the text encoder and the prompt upsampler are under Apache 2.0, and the ComfyUI nodes are MIT. This is our reading of the license text, not legal advice. The same territory limits apply to MiniMax H3's own open weights; see Is MiniMax H3 free?.

How to run P-H3-Image in ComfyUI

ComfyUI supports P-H3-Image natively, without custom nodes, but at launch the support was only on ComfyUI's master branch. Update to the latest version from GitHub first; the desktop app and Comfy Cloud get it with the next stable release.

Then download one file per component from PrunaAI/P-H3-Image-Comfy into ComfyUI/models. Precisions can be mixed:

The prompt upsampler is extra: 19.3 GB in bf16 (about 21 GB of VRAM, recommended), 9.7 GB in int8 or 6.7 GB in 4-bit. ComfyUI moves it out of memory before the image model runs. The smallest set:

hf download PrunaAI/P-H3-Image-Comfy \
  --include "diffusion_models/p-h3-image_w4a8.safetensors" \
            "text_encoders/p-h3-image_text_encoder_w4a8.safetensors" \
            "text_encoders/p-h3-image_prompt_upsampler_bf16.safetensors" \
            "vae/p-h3-image_vae_int8_convrot.safetensors" \
  --local-dir ComfyUI/models

Open p-h3-image_text_to_image.json or p-h3-image_edit.json from the repository's workflows/ folder and type a short request. Pruna says the int8 set stays very close to bf16, while the 4-bit set keeps the quality but can change small details and sometimes the composition. The quantized files run on NVIDIA Ampere or newer and AMD RDNA3/4 GPUs. If you already run MiniMax H3 in ComfyUI, a 2 GB LoRA turns your H3 model into P-H3-Image without the second large download.

How to run it with diffusers

The diffusers version needs a GPU with about 101 GB of memory at 8 MP (an H200 or B200), or one 80 GB GPU such as an H100 with automatic CPU offloading, which peaks at about 76 GB:

# pip install git+https://github.com/huggingface/diffusers
# pip install -U transformers accelerate torch pillow
import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("PrunaAI/P-H3-Image")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

image = pipe(
    prompt="A 1970s film photograph of a night market in the rain, lit by paper lanterns.",
    height=3360, width=2496,  # 8 MP, 3:4
    output="images",
)[0]
image.save("market.png")

Passing image= (one image or a list of up to 8) turns the same pipeline into an editor. The first picture is the one edited; the others are references you can name as Picture 2, Picture 3 and so on.

P-H3-Image API price

On November 12, 2026, WaveSpeed was the only hosted API we found for P-H3-Image. Pruna's Hugging Face page says no Hugging Face inference provider serves it yet.

Edits cost $0.005 more per extra reference image, and the LoRA versions (up to 3 LoRAs) add $0.015 per image. Its default tier is 1K, well below the 8 MP Pruna recommends, so expect more broken hands and faces unless you pay for 3K. WaveSpeed's pages leave commercial rights to the model's license, so the territory limits above still apply.

Prompting tips from Pruna

• Say what the image is in the first sentence, such as "a photograph by a street photographer" or "a two-color risograph print". Pruna says this sets the look better than style words.
• Describe one scene: a viewpoint, one main subject, a few supporting details, and a light source with a direction.
• Put every word of text in quotes, keep each string short and say where it goes. Unquoted text turns into fake lettering.
• Leave out hype words like stunning, epic, cinematic, 8k or masterpiece, which push it toward a glossy, airbrushed look.
• For edits, say what must stay the same. Without that, it tends to redraw the whole picture. The upsampler writes this for you.

Pruna's listed weak spots: hands, small people in the background, small or dense text, and lower resolutions.

P-H3-Image vs MiniMax H3 vs Z-Image Turbo

If the license rules P-H3-Image out for you, Z-Image Turbo is the closest open alternative: also 8 steps, much smaller, and Apache 2.0.

Where Kensa fits

Kensa does not run P-H3-Image. It runs the model P-H3-Image comes from, MiniMax H3, for video, through a hosted API. You don't need a GPU or the open weights, and there are 480P, 768P and 2K tiers. Upload a picture on Image to Video, choose MiniMax H3, and describe the motion. A 5-second 480P clip costs 8 credits; see MiniMax H3 on Kensa for every length and tier.

For the still image itself, Kensa's AI Image Generator runs Z-Image Turbo at 1 credit per image and sends the result straight to Image to Video. Kensa shows the credit price before you generate and returns the credits if a generation fails. Eligible new accounts that sign up with Google get 15 free credits, valid for 72 hours.

The smallest image-to-video clip on each model on Kensa, November 12, 2026:

FAQ

What is P-H3-Image?

An open-weight text-to-image and image-editing model from Pruna AI, built on MiniMax's H3 video model and released on November 3, 2026. It renders 8-megapixel images in 8 steps without CFG, edits with up to 8 input images, and includes a prompt upsampler based on Qwen3.5-9B.

Is P-H3-Image made by MiniMax?

No. Pruna AI made it from MiniMax's open MiniMax-H3 weights, so it falls under MiniMax's H3 Community License. MiniMax H3 itself is a video model.

Is "H3 image" the same as P-H3-Image?

Usually, yes. MiniMax H3 is a video model, and P-H3-Image is the image model people mean when they search for "H3 image" or "MiniMax H3 image". MiniMax H3 can also start a video from an image you upload, which is image-to-video, not image generation.

Is P-H3-Image free?

The weights are a free download, and Pruna runs a free Hugging Face demo. Hosted on WaveSpeed it costs from $0.02 per image at 1K. Its license excludes the US, EU, UK and South Korea unless you get a separate license from MiniMax.

Can I use P-H3-Image in the United States?

Not under the community license. It excludes the US, the EU, the UK and South Korea, including using or displaying the model's outputs there. MiniMax says people in those places can contact it about a separate license.

How much VRAM does P-H3-Image need?

About 21 GB at 8 MP with the 4-bit ComfyUI files, 31 GB with int8 and 56 GB with bf16. The diffusers bf16 pipeline needs about 101 GB with the prompt upsampler, or about 76 GB on an 80 GB GPU with offloading.

Can P-H3-Image edit images?

Yes. Give it one image to edit, or up to 8 in total: the first is edited and the others are references, such as a hat or a background to bring in. Pruna recommends 4 or 8 MP for edits.

Can P-H3-Image make videos?

No. It makes single images only. For video, MiniMax H3 is the model to use, and Kensa runs it for image-to-video and text-to-video.

Does Kensa offer P-H3-Image?

No. Kensa runs MiniMax H3 for video and Z-Image Turbo for images in its AI Image Generator.

Sources

• PrunaAI/P-H3-Image on Hugging Face: specs, diffusers setup, resolutions, prompting and limitations, read November 12, 2026
• PrunaAI/P-H3-Image-Comfy on Hugging Face: ComfyUI files, sizes, VRAM and workflows, read November 12, 2026
• P-H3-Image license (MiniMax H3 Community License Agreement): territory, notice and commercial terms, read November 12, 2026
• Pruna AI launch post: how the model was built and speed figures, dated November 3, 2026
• WaveSpeed P-H3-Image text-to-image and edit: prices per resolution, checked November 12, 2026
• MiniMaxAI/MiniMax-H3 on Hugging Face: the base video model
• Kensa credits: kensa.cc/pricing, from Kensa's billing configuration, November 12, 2026