P-H3-Image Prompt and Edit Guide

P-H3-Image is sensitive to how you write. Pruna AI's model card gives four rules for text-to-image prompts and one form for edit prompts. Following them helps, though artifacts can still appear. This guide explains each rule, shows example prompts we wrote to follow them, and covers references, sizes and the API. You can test every example in the online generator.

Four rules for a P-H3-Image prompt

1. Say what the image is in the first sentence

Start by naming the kind of picture, not the subject. "A photograph by an experienced street photographer", "a two-color risograph print" or "a still from a 1970s film" tells the model which look to aim for, and everything after it is read in that frame.

2. Describe one scene

Give the model a single, concrete moment:

• a viewpoint (eye level, from above, across a table),
• one main subject doing something,
• a few supporting elements,
• light with a source and a direction (late sun from the left, a desk lamp behind her).

Prompts that list many subjects or several moments tend to come out muddled.

3. Spell out every word of text in quotes

Write the exact words in quotes, keep each string short and say where it sits. End with "no other text". If you don't want any text in the image, don't mention signs, labels or posters at all, because naming them invites lettering.

4. Leave out hype words

Stunning, epic, cinematic, vibrant, hyper-detailed, 8k and masterpiece push the image toward a glossy, oversaturated, airbrushed look. Describe what is in the picture instead.

Example prompts

These are our own examples, written to the rules above.

Documentary photo

A photograph by an experienced documentary photographer of a fisherman mending a net on a concrete pier in the early morning. Seen from a low angle, he sits on an upturned crate with the net across his knees while two gulls wait on a bollard behind him. Soft light from the low sun on the right, mist over the water.

Screen-printed poster with text

A two-color screen-printed poster in orange and dark blue showing a single bicycle leaning against a lamppost at night. The words "NIGHT RIDE" sit in large block letters across the top and "Saturday 9 pm" in small letters at the bottom right, no other text.

Ink drawing

A loose ink and charcoal drawing of an old woman reading a newspaper at a café table, seen from across the room. A cup and a folded umbrella sit beside her, and window light falls from the left onto the page.

Film frame

A still from a 1970s film: a man in a brown suit waits alone at a bus stop on an empty suburban road at dusk, seen from across the street. A single street light above him has just come on, and the sky behind the houses is pale yellow.

Editing an image

Edits use a fixed form from the model card:

Edit <Picture 1>. <The change, described as the finished result>. Keep <everything that must stay> exactly as it is.

Four points make it work:

1. Point to the exact thing. "The pickup truck", "the boy holding the ball", not "the car" when there are three.
2. Describe the result, not the action. Write "The truck is painted red" rather than "make the truck red".
3. Say what stays. Without a "Keep … exactly as it is" sentence, the model tends to redraw the whole image. Name the face, pose, background and light if they matter.
4. Quote new text exactly. If the edit adds or changes words on a sign, give them in quotes.

Examples:

Edit <Picture 1>. The man is wearing a red knitted beanie. Keep his face, his jacket, the street and the light exactly as they are.

Edit <Picture 1>. The shop sign above the door reads "OPEN LATE", no other text. Keep the storefront, the people and the weather exactly as they are.

Using several images as references

An edit can take 1 to 8 input images on this site. Picture 1 is always the image being edited; the rest are references, named Picture 2, Picture 3 and so on in the order you upload them. Say in the prompt what each reference is for:

Edit <Picture 1>. The woman on the left is wearing the green coat from <Picture 2>, and the dog at her feet is the dog from <Picture 3>. Keep the park, the bench and the afternoon light exactly as they are.

The model card shows eight portrait paintings combined into one dinner party from a single sentence, so references can also be people placed into a new scene. Keep one clear job per reference, and upload them in the order you name them.

Choosing a size

The model's native size is 8 MP for text to image and 4 or 8 MP for edits. At 2 and 4 MP faces, hands and small details break more often, and at 1 MP skin looks smooth and waxy. On this site:

• 1K or 1.5K for quick drafts of composition and wording.
• 3K (about 9.4 MP) is closest to the native 8 MP and is the best choice for a final image.
• 4K (about 16.8 MP) for the largest output.

For edits, "Same as image 1" keeps the shape of your picture. The credit cost for each size appears under the Generate button before you run it.

Limits

• Very long prompts and hype words make results worse.
• Hands can break, and small people in the background can be distorted.
• Small or dense text, such as body copy or long code, turns into pseudo-text. Keep quoted strings short.
• The model makes images only, no video or audio.

Using it by API or with diffusers

Hosted API. Developers can call the model through WaveSpeed AI. The model ID is wavespeed-ai/p-h3-image, with a text-to-image endpoint and an edit endpoint. Request fields and authentication are in the WaveSpeed documentation.

Self-hosting. The weights and a diffusers pipeline are in the PrunaAI/P-H3-Image repo on Hugging Face. The pipeline needs diffusers installed from GitHub. The basic idea:

from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained("PrunaAI/P-H3-Image")
image = pipe(prompt="A two-color risograph print of a lighthouse at dawn.", height=3360, width=2496)

Pass image to edit a picture, or a list of up to 8 images to use references. Follow the model card for loading, device and memory settings: in bf16 the model needs about 83 GB of GPU memory at 8 MP without the upsampler and about 101 GB with it. For a smaller local setup, see our ComfyUI guide.

The model is released under the MiniMax H3 Community License, which excludes some territories unless you have a separate license from MiniMax and requires "Powered by MiniMax H3" on products built on it. Read the license before you ship anything; this guide is not legal advice.