This is MiniMax H3 Image, the image model. It is not MiniMax H3 video model, which is the video model. It is also not P-H3-Image photoreal model, which is a separate photoreal image model on PixelDojo. The names get mixed up. On the pricing page, the unlimited Relax line labeled "H3 Image" (1K, 0 credits) links to P-H3-Image, not to MiniMax H3 Image. MiniMax H3 Image uses 1.5 credits per image at 2K and 4 credits per image at 4K.

The longer, copy-ready version of this, with the 4K examples, is the MiniMax H3 Image prompting guide. This post is a shorter look at how to write those prompts.

{% embed https://www.youtube.com/watch?v=hE7kX1SFULw %}

MiniMax H3 Image prompt guide and same-prompt tests, from All Your Tech.

How to prompt MiniMax for photos that look taken

It follows a long, detailed prompt closely. A short line works, but a full paragraph is closer to the picture in your head. Write it like a brief for a photographer: what kind of image it is, who or what is in it, where each thing sits, and where the light comes from.

Quoted text comes out as typed if you put the words in quotes and name the spot. Text you only describe, like "a short line under each step," turns into made-up lettering.

It also edits a photo, or combines up to 8 images, and it can render up to 16 megapixels. A square image at 4K is 4096 by 4096, which is useful if you are going to crop or print a label.

On PixelDojo, a 2K image is 1.5 credits and a 4K image is 4 credits. Those are the public prices on the changelog and the guide.

MiniMax prompt habits that work

Start with what the image is. "A candid photograph," "a frame from a quiet 1990s independent film," "a risograph-printed poster," "a hand-painted gouache illustration." That first line does more than a list of style tags at the end.

Give each object its own sentence and its own place. Top left corner, bottom right, to the left of the blueprint, in that order from top to bottom. The guide's overhead desk shot works this way: a brass compass pinning the top left of a blueprint, a white coffee cup on the bottom right, a ruler on the diagonal, a notebook with a red elastic band to the left, and three pencils (yellow, blue, green) to the right. Each one landed where the prompt put it.

The guide's overhead desk shot. Each object landed where the prompt put it.

Quote every line of type. The guide's two-color bookshop poster quotes the shop name at the top, the line under it, and the address line at the bottom, then ends with "No other text." The print method ("risograph") and the two inks (teal and coral) set the look.

The two-color bookshop poster from the guide, with every line of type quoted in the prompt.

For a photo edit, say the change, then list what must stay. One of the guide examples takes a fishmonger portrait and asks for a yellow apron and light snow through the skylight, while keeping his face, the fish, the scale, and the counter the same. Leave the aspect ratio on Auto if you want to keep the shape of the original photo.

The same fishmonger after the edit: yellow apron and snow through the skylight, face and counter unchanged.

A MiniMax prompt you can copy

This is the candid portrait from the guide. It is a good template for any unposed person in a real place:

A candid photograph of a fishmonger in his fifties with a salt-and-pepper beard and a navy rubber apron, laughing as he weighs a silver sea bass on a brass scale. Crushed ice and small handwritten price tags fill the counter in front of him. Cold morning light falls from a market skylight above. Shot on a 50mm lens, natural skin texture, slight motion in his hands.

The candid fishmonger portrait from the MiniMax H3 Image prompting guide.

Two small choices keep it from feeling posed: "candid photograph" plus one action in progress, and a named light source plus a lens so it stays a photo instead of drifting into illustration.

If you want handwriting instead of a poster, treat it the same way as print. Quote each line, and say the pen and the paper. A recipe card in blue ballpoint on cream stock with a thin red border is much more likely to look like a card someone actually wrote than "a handwritten note" with no other detail.

When I would pick something other than MiniMax

Seedream 5.5 is on PixelDojo too, and I would keep both around. MiniMax H3 Image is the one I would reach for when I want a frame that feels photographed, or when I have a long, fussy brief with a lot of object placement. If I am trying to cut a sticker or a logo as a see-through PNG, I would go elsewhere. Different jobs.

I am not picking a winner against Nano Banana, GPT Image, or Seedream in this post. If you want same-prompt grids, we published those as separate guides. Start with the MiniMax Image prompting guide if you just want prompts that already have 4K frames next to them.

And if you came here looking for the video model, that is MiniMax H3 video model. If you want the other photoreal image model on PixelDojo, that is P-H3-Image photoreal model.

The short version: it makes a sharp 2K image in about 15 seconds, it spells the words you put in quotes, it can edit your own photos with up to 10 reference images, and it can hand you a PNG with nothing behind the subject. That last one is handy when you make stickers, logos, or product cutouts, because you can skip a second round in a background remover.

The live tool is on this page:

Seedream 5.5 on PixelDojo

If you want the copy-ready prompts and the notes next to each one, use the Seedream 5.5 prompting guide. This post is the shorter version of that.

How to prompt Seedream for exact quoted text and see-through PNGs

Quoted text is the part that stands out. Put each line in double quotes, say where it sits, and end with something like "the text must be exact, no other text anywhere." The plant-shop photo in the guide comes back with a painted wall sign, a chalkboard line, and a paper tag on a pot, all spelled as typed. A concert poster in the same guide carries four separate lines the same way.

A Seedream 5.5 concert poster from the PixelDojo guide, with four quoted lines spelled as typed.

See-through backgrounds are the other one. In the tool, pick See-through in the Background menu and say "transparent background" in the prompt. The guide uses that for a six-sticker sheet and a circular surf-school badge. Both come back as PNGs you can drop onto a slide or a product page without cutting anything out by hand.

A six-sticker sheet from the guide, returned as a see-through PNG.

It is also fine as a regular camera. The guide's portraits, product bottle with a label, library interior, and ramen shot are each one prompt, at 2K. If you describe the lens, the light, and one or two physical details (individual hairs, wool fibers, droplets on glass), it tends to keep them.

The guide's library interior, one of the regular-camera 2K examples.

Edits are another mode. You can attach up to 10 reference images. The prompt pattern in the guide is boring and effective: list what must stay first, then describe the change. In the harbor-wall portrait example, the prompt asks it to keep the same man, face, smile, and tight 85mm framing, then move him to a snowy street in a camel coat. The face held. The coat, the street, and the snow were new.

The harbor-wall portrait after the edit: same face and framing, new coat, street, and snow.

What Seedream costs on PixelDojo

These are the public credit prices on the model page and the guide, not a comparison against anyone else's per-image API.

• From a text prompt: 1 credit at 2K, 0.5 credits at 1K
• With reference images: 1.5 credits at 2K, 1 credit at 1K
• For context, Seedream 5 Pro is 2.75 credits at 2K on PixelDojo

Seedream 5.5 is included in every plan, from $10/month. It is also a model choice in Character Studio, Character Sheets, Consistent Characters, Reality Polisher, and Canvas, if you already work in those tools.

A Seedream prompt pattern you can steal

Here is the plant-shop one from the guide, unchanged. It is a good template for any scene that has more than one piece of type in it:

Photograph on a 35mm lens at eye level: a beautiful woman in her late twenties with long glossy chestnut hair, bright green eyes, flawless glowing skin and a radiant smile, in a fitted cream knit top and a mustard corduroy jacket, stands behind the counter of a small plant shop at golden hour, holding a terracotta pot with a fern. A hand painted wooden sign on the wall behind her reads "THE GREEN ROOM". A small chalkboard on the counter reads "REPOTTING EVERY SATURDAY". A paper tag tied to the pot reads "FERN 14". Low warm sun through the front window lighting her face, dust in the light, shelves of plants softly blurred, shallow depth of field, natural skin texture, the text must be exact, no other text anywhere.

The plant-shop photo from the guide, with all three signs spelled as typed.

Three habits make this kind of prompt work:

1. Give each piece of text its own sentence: what it is written on, where it sits, and the words in double quotes.
2. Name the lens and the light. "35mm at eye level" and "low warm sun through the front window" do more than a list of quality words.
3. If you are editing a photo, lock the face, the framing, and anything else that must not drift, then describe only the change.

If you are making a poster or an infographic, write it top to bottom. Headline, second line, the picture in the middle, then the lines at the bottom. Short labels of two or three words come out cleaner than long sentences.

Seedream compared with other image models this week

People have been posting same-prompt grids of Seedream 5.5 against MiniMax H3 Image, Nano Banana Pro, and GPT Image 2.6. PixelDojo now has Seedream 5.5 and MiniMax H3 Image in the same place, plus Seedream 5 on PixelDojo if you want to see the previous Seedream next to it. I am not going to pick a winner in this post. The prompting guide is a better place to look at the actual frames and decide for yourself.

If you want to try it, start on the Seedream 5.5 tool page, or skip straight to the Seedream prompting guide and copy one of the ten prompts.