Analysis2026-11-13Updated 2026-11-13

P-H3-Image: What Pruna AI’s MiniMax H3 Image Model Can Do and How to Use It

TL;DRP-H3-Image is Pruna AI’s open-weight image generation and editing model based on MiniMax H3. It supports eight-step generation and image editing, with online and ComfyUI workflows. Check GPU requirements and the MiniMax H3 Community License before local or commercial use.

By Cameron Voss and West Mao

P-H3-Image is an image generation and editing model from Pruna AI, built on the MiniMax H3 video model. It turns text prompts into still images and can edit an existing image using additional visual references. Pruna released the model weights and ComfyUI workflows in November 2026, making it possible to try the model online or run it locally if your hardware and license permit.

The quick answer: Start with Pruna’s Hugging Face Space for a no-install test. For local use, update ComfyUI, download the P-H3-Image-Comfy files and open the supplied text-to-image or edit workflow. The smallest image-model set requires around 21 GB of GPU memory at 8 megapixels; the prompt upsampler is a separate download. Check the license before using it.

What is P-H3-Image?

MiniMax H3 was designed as a video model. Pruna adapted it for still images by changing its prompt processing, text encoder, sampling schedule, model backbone and decoder. The resulting P-H3-Image can generate photographs, illustrations and graphic-style images, or follow instructions to modify an input image.

Pruna says its model produces an image in eight sampling steps, compared with 50 for the original H3 in its comparison. On a single NVIDIA B200, the company reports 2.6 seconds for a 4 MP image versus 12.4 seconds for the original model. These are provider-reported measurements under specific conditions, not a promise of speed on a home GPU.

  • Tasks: Text-to-image generation and image editing

  • References: Up to eight input images according to the model card

  • Resolution: 2–16 MP supported; 8 MP recommended for generation

  • Sampling: Eight steps; the supplied workflow uses no CFG or negative prompt

  • Weights: Full-precision and quantized options, subject to the MiniMax H3 Community License

How to try H3-Image online

  1. Find PrunaAI/P-H3-Image on Hugging Face and open the linked Hugging Face Space from the model page.
  2. For text-to-image, enter a short description of the subject, setting and visual style. Choose an available aspect ratio and generate.
  3. For editing, upload an image if the Space offers an edit mode. Describe the desired result and specify what should remain unchanged.
  4. Inspect the result closely, especially faces, hands and text. Revise the prompt or generate another version if needed.

Hosted availability, queues, controls and pricing can change. Check the current interface before relying on it. The similarly named Pruna API products p-image and p-image-edit should not be assumed to be P-H3-Image.

How to run P-H3-Image locally in ComfyUI

The supplied ComfyUI workflows are a practical starting point if you have a compatible GPU and want local control. Pruna says the model uses native ComfyUI nodes. Its model card notes that support was on the latest development branch and would arrive in a subsequent stable release, so check your installed version if nodes are missing.

Before downloading: The smallest image-model, text-encoder and VAE set is approximately 20 GB and peaks at about 21 GB of VRAM at 8 MP. The recommended prompt upsampler adds a separate 19.3 GB download and has its own memory requirements. Allow extra disk space; requirements vary with settings and hardware.

  1. Update ComfyUI. Use a version that supports P-H3-Image; if the stable version lacks its nodes, follow the model card’s instructions for the development version.
  2. Find the model files. Search Hugging Face for PrunaAI/P-H3-Image-Comfy. Choose one diffusion model, one text encoder, one VAE and a prompt upsampler. The w4a8 model and encoder with the int8 VAE are the smallest documented set.
  3. Place the files in the matching folders. Put the diffusion model in ComfyUI/models/diffusion_models, the text encoder and prompt upsampler in ComfyUI/models/text_encoders, and the VAE in ComfyUI/models/vae.
  4. Open a supplied workflow. Load p-h3-image_text_to_image.json for generation or p-h3-image_edit.json for editing from the repository’s workflows folder. Select your downloaded files in the loader nodes.
  5. Enter your prompt and run. The supplied workflows use the prompt upsampler and an eight-step sampling schedule. For editing, load the source image and any optional references into the appropriate image inputs.

At 8 MP, Pruna lists roughly 21 GB peak VRAM for the smallest set, 31 GB for int8 and 56 GB for bf16. The 4-bit version may change small details or composition compared with higher precision. Quantized kernels are documented for NVIDIA Ampere or newer and AMD RDNA3/4; compatibility can depend on your setup.

Choose a size

For generation, the model card recommends about 8 MP: 2912 × 2912 for square images or 3872 × 2176 for 16:9. For edits, 4 MP or 8 MP are recommended; the provided edit workflow scales the primary image to 4 MP by default. Generating larger images can increase memory use and processing time.

Prompt examples for generation and editing

Start with the medium, subject, setting and light. The included prompt upsampler expands a short request, so you do not need a long list of quality adjectives.

A documentary photograph of a baker arranging fresh bread in a small neighborhood bakery, morning window light, natural colors.

For an edit, identify the image and describe the finished result. State the details you want preserved:

Edit <Picture 1>. The red jacket is now dark green. Keep the person's face, pose, background, and lighting unchanged.

Put short words to appear in the image inside quotation marks. Treat generated lettering as a draft: small text, long passages and code may contain mistakes. For product images or portraits, compare the output with the original before using it.

Strengths and limitations

Pruna highlights realistic photography, portraits, film-like scenes, illustrations and short text as strengths. Multi-image editing can be useful for combining visual references or changing a scene while keeping a recognizable subject.

Results still need review. Hands and small background figures can be distorted; dense text can become unreadable; and edits can redraw details you intended to preserve. Short prompts may perform poorly without the upsampler, while very long prompts or excessive style buzzwords can also weaken results. These are model-maker observations, not independent benchmark results.

License and commercial use

P-H3-Image uses the MiniMax H3 Community License. Pruna’s model card states that the license excludes the European Union, United Kingdom, Republic of Korea and United States unless a separate MiniMax license is obtained. It also says products built on the model must display “Powered by MiniMax H3.” Read the current license and component notices for your intended location and use before deploying or distributing it. Open weights do not mean unrestricted use.

Frequently asked questions

Is H3-Image the same as MiniMax H3?

No. P-H3-Image is Pruna AI’s image-focused model derived from MiniMax H3, which is a video model. The two have different workflows and resource requirements.

Can P-H3-Image edit an existing image?

Yes. Use the edit workflow, provide an image and describe the desired result. You can add reference images where the workflow supports them.

Can I run it on a 16 GB GPU?

Do not assume the supplied 8 MP workflow will fit. Pruna reports about 21 GB peak VRAM for its smallest image-model set at 8 MP. Lower resolutions or offloading may change requirements, but performance and compatibility depend on the setup.

Is it free for commercial use?

Availability of weights does not by itself grant unrestricted commercial rights. Review the MiniMax H3 Community License and its territorial restrictions, and obtain a separate license where required.