P-H3-Image: The MiniMax H3 Image Model, Explained
P-H3-Image is Pruna AI's open-weights image model built on MiniMax H3, announced in November 2026. One checkpoint does text to image and image editing with up to 8 input images, at about 8 megapixels in 8 steps. It is not on Nereo: its license excludes the United States, the EU, the UK and South Korea. The generator below runs Seedream 5.0 Pro for text to image and multi-reference editing. Nereo is an independent platform, not affiliated with Pruna AI or MiniMax.
Text to Image
What Is P-H3-Image?
P-H3-Image is an image generation and editing model from Pruna AI. It is fine-tuned from MiniMax H3, MiniMax's open video model, and adds Pruna's own text encoder, prompt upsampler, VAE and LoRA. Pruna released it as open weights on Hugging Face and in ComfyUI in early November 2026. It is not an official MiniMax release.
The model makes stills only. One checkpoint handles text to image and image editing with 1 to 8 input images. It renders at about 8 megapixels by default, such as 2912 x 2912 for a square image, in 8 sampling steps with no CFG or negative prompt.
It is released under the MiniMax H3 Community License. That license excludes the United States, the European Union, the United Kingdom and South Korea, so people and products there cannot use the model or its outputs without a separate license from MiniMax. Because of that, Nereo does not run P-H3-Image. This page explains the model and points to image models you can use on Nereo today.
P-H3-Image Specs
From Pruna AI's model card on Hugging Face and the ComfyUI guide, checked on November 11, 2026.
What P-H3-Image Does Well
8-Megapixel Stills in 8 Steps
Pruna tuned the model to finish in 8 steps at about 8 megapixels, so it produces large, detailed images quickly on capable hardware. Pruna's card says the model looks best at 8 MP and shows more artifacts at 2 or 4 MP.
| What it means | Example |
|---|---|
Large outputs suit print, product pages and crops. On Nereo, Nano Banana Pro renders at up to 4K, and Seedream 5.0 Pro at 1K or 2K. | ![]() |
Edits That Keep the Rest of the Picture
For edits, Picture 1 is the image you change. Pruna's prompting guide says to point to the exact thing, describe the finished result and say what must stay, for example a truck repainted red with the road, framing and light unchanged.
| What it means | Example |
|---|---|
The same prompting style works with Seedream 5.0 Pro image to image on Nereo: name one change and list what stays. | ![]() |
Up to 8 Reference Images
An edit can use 1 to 8 input images, so you can combine a product, a material and a color palette into one result.
| What it means | Example |
|---|---|
On Nereo, Seedream 5.0 Pro accepts up to 10 reference images for the same kind of composition. | ![]() |
Plain, Specific Prompts
Pruna recommends saying what the image is, describing one scene with its light, quoting every word of text and leaving out hype words such as epic or 8k, which push the model toward a glossy look.
| What it means | Example |
|---|---|
These habits help most image models. Try them in the generator above with Seedream 5.0 Pro. | ![]() |
How to Use P-H3-Image
Ways to run it where the license allows, and what to use on Nereo.
Check the License First
Read the MiniMax H3 Community License. It excludes the US, the EU, the UK and South Korea, and large commercial users need written permission from MiniMax.
Run It Locally
Use the Diffusers pipeline from Pruna's Hugging Face page on an 80 GB GPU, or the ComfyUI int8 files on about 21 GB of VRAM.
Or Use a Hosted API
WaveSpeedAI lists a P-H3-Image API from $0.02 per image at 1K. Check that your use fits the license.
On Nereo, Use Seedream 5.0 Pro
For text to image and multi-reference edits without a GPU or license check, use the generator at the top of this page.
Prompts Written in P-H3-Image's Style
Following Pruna's guide: name the image type, describe one scene, quote text exactly. They also work with Seedream 5.0 Pro on Nereo.
Street Photo
“A photograph by an experienced street photographer. A flower seller arranges tulips at a market stall, low morning sun from the left, shallow depth of field, people blurred in the background.”
Product Shot
“A studio product photograph of a matte green travel mug on pale sandstone, soft leaf shadow, side light from the right, clear space on the left third.”
Poster Text
“A minimalist concert poster with a single red guitar on cream paper. Large text "SUMMER NIGHTS" at the top and "July 12" below it, no other text.”
Edit
“Edit Picture 1. The pickup truck is now glossy red. Keep the road, the trees, the framing and the light exactly as they are.”
P-H3-Image vs Image Models on Nereo
| Feature | P-H3-Image | Seedream 5.0 Pro | Nano Banana Pro |
|---|---|---|---|
| On Nereo | No | Yes | Yes |
| Made by | Pruna AI (on MiniMax H3) | ByteDance | |
| Open weights | Yes | No | No |
| Reference images | Up to 8 | Up to 10 on Nereo | Up to 5 on Nereo |
| Default output | About 8 MP | 1K or 2K on Nereo | 1K, 2K or 4K on Nereo |
| Can be used in the US and EU | Not without a separate MiniMax license | Yes | Yes |
Related Models on Nereo

MiniMax H3 Image
MiniMax's own H3 image model, live on Nereo: 4K images and edits with up to 16 references.
Nano Banana Pro
Google's image model for edits and text in images.

MiniMax H3
The MiniMax video model P-H3-Image is built on.
Seedream 5.5
ByteDance's newest image model and how it compares.
P-H3-Image FAQ
What to know about Pruna's MiniMax H3 image model.
P-H3-Image is an open-weights image model from Pruna AI, fine-tuned from MiniMax H3. It does text to image and image editing in one checkpoint and renders at about 8 megapixels in 8 steps.
No. Pruna AI built it on MiniMax H3's open weights. MiniMax H3 itself is a video model, and MiniMax has not released an official H3 image model.
No. Its license excludes the United States, the EU, the UK and South Korea, so Nereo does not offer it. Seedream 5.0 Pro and Nano Banana Pro are available on Nereo for text to image and editing.
Only within the MiniMax H3 Community License. It excludes the US, the EU, the UK and South Korea, requires written permission from MiniMax for companies above 20 million US dollars in yearly revenue, and asks commercial products to show MiniMax H3 in their interface. Read the license before you use it.
The weights are free to download. You need your own GPU, about 80 GB for the full model or about 21 GB with the ComfyUI int8 files. Hosted APIs charge per image.
Pruna's model card lists 1 to 8 input images, with Picture 1 as the image being edited.
About 8 megapixels by default, for example 2912 x 2912 at 1:1 or 3872 x 2176 at 16:9. Pruna says it looks best at 8 MP.
WaveSpeedAI lists a hosted P-H3-Image API, priced from $0.02 per image at 1K to $0.12 at 4K as of November 11, 2026.
Say what the image is, describe one scene with its light, quote every word of text and avoid hype words like epic or 8k. For edits, name the change and list what must stay.
No. Nereo is an independent platform and is not affiliated with Pruna AI or MiniMax.



