Loading...

P-H3-Image

0/5000
Aspect Ratio
Images
Required credits
1

Sample Images

Forest under a green aurora at night

Explore image models

Make and edit still images with P-H3-Image

P-H3-Image is Pruna's still image model, built from the MiniMax-H3 video model so one frame can stand on its own. It is a different checkpoint from MiniMax-H3-Image and from the P-Image family. Use it when you want a photograph, a film frame, an editorial portrait, or a graphic with a short line of text. Start from a written scene in Text to Image, or switch to Image to Image when you already have a photo to change. The model card describes text to image and image editing in one model, with up to eight input images, and it asks you to judge a new picture at about 8 megapixels. This page is the workspace for that job. Write the scene, pick an aspect ratio, and generate.

Forest under a green aurora at night

How to create with P-H3-Image text to image

Use Text to Image for a new still, or Image to Image when the first picture already exists.

  1. Choose Text to Image or Image to Image

    Text to Image starts from your words and the aspect ratio you select. Image to Image starts from pictures you upload. For an edit, the first image is the one being changed. Further images are references you can name in the prompt as Picture 2, Picture 3, and so on, up to eight images in total. Leave the reference slots empty when you are only describing a new scene.

  2. Write one scene or one change

    Open with what the picture is. A street photograph, a frame from a film, a two color print, or a flat scan of a recipe card will steer the look more clearly than a pile of style adjectives. Then describe one moment: where the camera is, who or what is in frame, a few supporting objects, and light that has a source and a direction. For an edit, say what changes and what must stay. Point at the object and where it sits, such as the faded blue cap rather than "his hat." If you add an object, give it a size next to something already in the frame. Put every word that must appear in quotes, say where it sits, and end with "no other text." If you want no lettering, do not mention signs, labels, or words at all.

  3. Generate and check the still

    Select Generate. When a result appears, open it large enough to look at faces, hands, and any words you placed in quotes. Short quoted text is the kind of lettering this model is built to attempt. Keep long paragraphs and code in the caption you write beside the image. If a hand or a small figure in the background looks wrong, name that part and try the prompt again, or start from a clearer reference photo. Download the still you want to keep from the result area.

Image editing and new scenes with Pruna P-H3-Image

P-H3-Image image editing and a new scene use different inputs in this workspace.

Describe a photograph or a film still

Use Text to Image when the picture does not exist yet. The published examples lean on documentary and editorial photographs, portraits with visible skin, film frames, ink, charcoal, screen prints, and flat graphics. Say the medium in the first sentence, then one scene. A red fox in fresh snow at first light is a useful start once you also say it is a photograph and where the light comes from. Leave out words such as stunning, epic, cinematic, vibrant, hyper detailed, and masterpiece. The model card says those words push the picture toward a glossy, oversaturated finish. A very short line is a weak instruction until you add the subject, the place, and the light. A very long line, or a line full of praise, gets worse. Write the scene you could walk into, then stop.

Start from a prompt
Convertible on a coastal highway beside the ocean

Change one part of a photo

Switch to Image to Image and upload the photo you want to change as the first image. A short line such as "make the pickup truck red" often causes a full redraw, because the model then fills in the rest of the scene. Name the finished color or object, and list what has to survive: the faces, the pose, the background, and the light. New text goes in quotes, next to the words that must stay. On the model card, edits are aimed at about 4 or 8 megapixels. Without a new canvas, an edit follows the first picture and is brought up to at least about 4 megapixels. This workspace asks for an aspect ratio rather than a pixel box. Pick the ratio that matches the picture you want to keep, and keep the first upload as the image being edited.

Edit a photo
Person in a white dress standing beside a building

Build one scene from several pictures

You can add more references after the first image, up to eight in total. The model card treats the extra pictures as sources the prompt can name. Eight portraits and one sentence can become a dinner once the prompt says what to take from each person. A room of specific furniture works the same way: each object has its own photo, and the prompt says where it goes. Eight references already cover a small cast or a short product layout. If you only need a color change on one photo, extra references do not help. Upload the pictures that carry a face, a hat, or a chair the words cannot fully describe, and say which picture supplies each one.

Add references
Sunset above a sea of clouds

Settings that change the still

Four choices to make before you generate, taken from the current model card.

Name the medium first

The first sentence sets the medium more reliably than a style adjective added at the end. Try a photograph by a street photographer, a frame from a Japanese anime series, a two color risograph, or a flat scan of a handwritten recipe card. Then stay inside that medium for the rest of the prompt.

Quote only the short text

Spell every word that must appear, in quotes, and say where it sits. Short quoted strings are the text this model handles. Long passages can miss a letter, and small print turns into shapes that only look like words. If the sentence has to be exact, set the type beside the picture after you download it.

Match a published aspect ratio

The model card lists seven ratios for an about 8 megapixel text to image canvas: square, 4:3, 3:4, 3:2, 2:3, 16:9, and 9:16. The square default on the card is 2912 by 2912. A 16:9 frame at that area is 3872 by 2176. Sides are rounded to a multiple of 32. This workspace offers those same ratios. Pictures at about 2 and 4 megapixels can work, and the card says faces, hands, and small details break more often there than at about 8 megapixels. When a face or a short headline matters, choose the ratio you mean to keep instead of stretching a small preview later.

Look at hands, faces, and edges

After the image appears, check the hands, the small people in the background, and the crop at the edges. The model card says hands can still break and that small background figures can distort, especially on a smaller canvas. If you liked a rough idea, describe it again at the ratio you mean to keep. A new size can move the framing, because the noise field changes with the canvas.

Questions before you generate

What this model is, what you can upload, and what the workspace asks for.

No. P-H3-Image is Pruna's still image model, based on the MiniMax-H3 video model. MiniMax-H3-Image is MiniMax's own image model, and this site already has separate text to image and image to image pages for it. The shared H3 name does not move the step count, the megapixel target, or the reference limit from one checkpoint to the other. Stay on this page when you want P-H3-Image.

No. P-Image, P-Image Edit, and P-Image Upscale are a different Pruna image family. P-H3-Image is the still image model Pruna built from MiniMax-H3. This page uses P-H3-Image for both Text to Image and Image to Image.

Start a still with P-H3-Image

Write a scene for Text to Image, or upload the photo you want to change with Image to Image.