MiniMax H3 Image Prompts: 12 Copy-Ready, 6 Tested
Twelve copy-ready MiniMax H3 Image prompts grouped by job, from product shots and typography to 21:9 panoramas and reference edits. We ran six on each::labs and wrote down what the model did with every line.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

A model that renders sixteen megapixels shows you every lazy word in your prompt, at full size. These MiniMax H3 Image prompts are the fix: twelve copy-ready prompts grouped by job, six of which we ran ourselves on each::labs this week, with notes on what the model did with them.
For the model itself, read our first look, MiniMax H3 Image: We Tested Native 4K on each::labs. This post is the practical half. Open MiniMax H3 Image Text to Image in another tab, paste, and adjust.
The Building Blocks of MiniMax H3 Image Prompts
MiniMax's own model card gives one prompting instruction that matters more than the rest: write a detailed English description of the whole image. Not keywords. A brief for a photographer who can't ask questions. It's a 33.1B diffusion transformer distilled to six steps with no classifier-free guidance, so there's no guidance scale to crank and no negative prompt. Everything you want goes into the prompt, in positive terms.
Five blocks cover almost every brief. Start with the subject, named precisely ("a matte black ceramic pour-over dripper", not "a coffee thing"). Then the setting and surface it sits on. Then light, with a direction and a quality ("single softbox from the upper left, long soft shadow to the right"). Then lens and finish: focal length, aperture, grain, color grade. Finally any text, inside straight quotes, with its position and typeface described.
Completeness beats order. In our runs the model filled the gaps with a confident decision of its own: leave out the product design and it invents one. Great for exploring. When you're matching a brief, close the gaps.

The Settings That Actually Change the Picture
All four MiniMax H3 Image pages on each::labs share a short input list, and three settings shape the result more than any adjective.
Size. The presets run from 1024x1024 up to 4096x4096, 5472x3072 and 3072x5472, with 2736x1536 and 1536x2736 in between. Then there's auto, which lets the model pick the shape at its 4K target area. Sneaky useful: when we set auto and wrote "Ultra-wide cinematic panorama in 21:9" in the prompt, the image came back at 6272x2688, the exact 21:9 resolution listed on the model card. Your words can choose the canvas.
Prompt enhancement. Text to image runs a prompt expander by default; disable_prompt_enhanced turns it off, but only with a fixed pixel size, and then only about the first 4,096 prompt tokens count. Keep it on for short ideas, off for full briefs you want read literally. Image to image always enhances.
Standard or Highspeed. MiniMax H3 Image Highspeed Text to Image is the drafting tool. In our product test below it returned in 6.9 seconds of model time against 10.7 for Standard at the same 2736x1536 size. Beyond that, n returns 1 to 10 images in one request, and output is JPEG, PNG or WEBP (no transparent backgrounds).

Product Shots: Light First, Props Never
Lead with the object and the surface, then spend your words on light. Our first test, run on Standard and Highspeed at 2736x1536:
Studio product photograph of a matte black ceramic pour-over coffee dripper on a slab of pale travertine stone. Single softbox from the upper left, long soft shadow falling to the right, a thin wisp of steam rising from the cone. Background is a seamless warm sand backdrop with a gentle falloff to darker tone at the top. Shot on a 100mm macro lens at f/8, everything tack sharp, visible glaze texture and tiny pores in the stone. Clean commercial catalog look, no props, no text.
Both versions got the light right: shadow to the right, steam rising, travertine pores you could count. Where they split was the product. Standard placed the dripper off-center and gave it a design we never asked for, ribbed sides with a triangular cut-out window. Highspeed centered it and produced a fluted, origami-style cone. If the shape matters, describe the shape, or edit from your real product photo (see below).
Macro product photograph of a small amber glass dropper bottle with a black rubber bulb and no label, standing on a wet slab of dark slate. Three water droplets on the glass catch a hard rim light from behind, a soft fill from the front left keeps the amber color rich. Shallow depth of field, the background dissolves into deep green out-of-focus leaves. 90mm lens, f/4, crisp highlights, editorial skincare look.
Posters and Typography: Quote Every Word
Put every word in straight quotes, say where it sits, and describe the lettering in type terms. Our test, at 1536x2736:
Vertical concert poster for a jazz night. Top third: the words "BLUE HOUR" in huge condensed sans-serif letters, cream on deep navy. Middle: a flat, screen-printed illustration of a saxophonist in silhouette against a large orange setting sun, two-color halftone texture. Bottom: a single line of small text reading "Friday 21 November, Doors 8 PM" centered in cream. Limited palette of navy, orange and cream, slight paper grain, mid-century Swiss poster layout with generous margins.
Both lines came back spelled exactly right, comma included, in the right places and the right colors. The sun became a halftone sunburst, better than we asked for. The one instruction it overrode was "generous margins": the headline ran almost edge to edge. If whitespace is part of your layout, give it a number ("the headline occupies the central 70 percent of the width").
Flat lay photograph of a glass jar of dark honey with a kraft paper label tied on with twine. The label reads "WILDFLOWER" in a hand-drawn serif on the first line and "Raw Honey, 340 g" in small sans-serif on the second line. Wooden spoon dripping honey beside the jar, linen cloth, soft overcast daylight from above, warm neutral palette, sharp focus on the label lettering.

Portraits That Keep Their Skin
Skip "beautiful". Ask for an age, a job, a texture and a light source. This one ran at 1536x2736:
Environmental portrait of a woman in her sixties who runs a small bicycle repair shop, standing in her doorway with grease on her hands and a folded cloth over one shoulder. Late afternoon side light from the street, warm on her face, the shop interior falling into shadow behind her with wheels hanging from the ceiling. Natural skin texture with fine lines and freckles, gray hair tied back loosely, calm direct gaze into the lens. 85mm lens, shallow depth of field, documentary color grading, slightly muted greens.
This was the strongest of our runs. Freckles, fine lines and greasy knuckles all landed, the wheels hung where we put them, and the greens in the shop really were muted. It also dressed her in a denim work jacket and a leather apron we never mentioned. Name the wardrobe when it matters.
Tight beauty close-up of a man in his twenties with deep brown skin and a short beard, eyes closed, a single line of gold glitter along one cheekbone. Hard light from directly above, deep shadows under the brow, black background. Visible pores and fine beard hair, no retouching look, 105mm lens, editorial magazine finish.
Wide 21:9 Scenes: Fill the Width
A 6272-pixel-wide frame is empty space unless you plan it left to right: what lives at each end and what the light does across it. Set size to auto and name the ratio.
Ultra-wide cinematic panorama in 21:9: a lone red tram crossing a long steel bridge over a misty river valley at dawn, terraced hills on both sides, the first sun hitting the mountain peaks on the far right while the left of the frame stays in cool blue shadow. Tiny figures waiting on a platform at the left end of the bridge. Anamorphic look, layered atmospheric haze, rich detail across the entire width of the frame.
That's the run that came back at 6272x2688. The plan landed almost literally: cool shadow left, tiny waiting figures at the left end of the bridge, sunlit peaks right. Spatial instructions are a strength.
Ultra-wide 21:9 night landscape: a white domed observatory on a desert ridge at the far left, its slit open, the Milky Way arching across the whole sky to the right. Faint red glow of a dirt road winding through the center, sand dunes in silver moonlight on the right edge. Long exposure astrophotography look, deep blue sky, crisp stars, no light pollution.

Edits and References: Name Your Pictures
MiniMax H3 Image Image to Image takes 1 to 16 reference images, and the model card refers to them in the prompt as <Picture 1>, <Picture 2> and so on, in upload order. Use those tags. Then say two things: what changes, and what must stay. We fed our Standard product shot back in:
Take the black ceramic coffee dripper from <Picture 1> and keep its exact shape, cut-out window, ribbing and matte glaze. Place it on a glass carafe on a worn oak kitchen counter next to a window, soft morning light from the right, a linen towel and a small bag of coffee beans slightly out of focus in the background. Lifestyle catalog photo, 50mm lens, natural colors.
The dripper survived intact, window and ribbing included, sitting on a carafe of coffee with the window light from the right. That's the product workflow: one approved hero, moved through scenes by edits. Text edits were just as clean:
Edit <Picture 1>: change the headline to "NIGHT SHIFT" in the same condensed lettering, change the date line to "Saturday 6 December, Doors 9 PM", and change the orange sun to a pale teal moon. Keep the saxophonist silhouette, the navy background, the halftone texture and the layout exactly as they are.
Both text lines swapped exactly, the sun became a teal moon, and the silhouette came back in the same pose. One poster, a season of events. We didn't run this multi-reference template, so treat it as a starting point:
The person from <Picture 1> wearing the jacket from <Picture 2>, standing in the street from <Picture 3> at dusk. Keep the face and hairstyle from <Picture 1> and the exact color and stitching of the jacket from <Picture 2>. Street lamps just switching on, light drizzle, 35mm lens, candid fashion photography.
When you're iterating on edits, run them on MiniMax H3 Image Highspeed Image to Image first, then render the keeper on Standard.
From Still to Motion With MiniMax H3
A strong still is also a strong first frame. MiniMax H3 image-to-video (the video model) takes a first frame plus an optional last one, adapts the video's aspect ratio to it and renders 4 to 15 seconds at 768P or 2K. So leave room for the motion you plan. A twelfth prompt, written as an opening frame:
Wide 16:9 photograph of a paper lantern festival on a river at blue hour, hundreds of glowing lanterns drifting downstream from the lower left, an arched stone bridge on the right, the upper half of the frame left as calm deep blue sky with the first stars. Soft reflections on the water, 24mm lens, cinematic color.
Then hand it to the video model with a short motion line such as "the lanterns drift slowly toward the bridge, the camera rises gently into the sky". To wire that pipeline once and reuse it, build it in each::labs Flows. For more on the video side, our post on MiniMax H3 and its reference inputs covers the family.
How to Iterate Without Starting Over
Draft on Highspeed with n at four. Note what the model decided on its own, write the keepers into the prompt as facts, then render once on Standard.
Fix mistakes with positive instructions: "a bare seamless backdrop" beats "no clutter". Fix placement with plain coordinates: top third, far left. And when a subject must stay identical across shots, stop describing it and reference it with an edit.
Every model in the text-to-image catalog and the image-to-image catalog runs through the same each::labs prediction API, so one prompt can test several models.
Frequently Asked Questions
What makes good MiniMax H3 Image prompts?
Full sentences that describe the whole frame; in our tests the unstated details got the model's own choice. Name the subject precisely, give light a direction, state the lens, and quote any text you want rendered.
Does MiniMax H3 Image support negative prompts?
Not according to the model card. Phrase exclusions as positive descriptions: "a bare seamless backdrop" instead of "no props", "unlabeled glass bottle" instead of "no logo".
How do I get a 21:9 image from MiniMax H3 Image on each::labs?
Set size to auto and say "21:9" or "ultra-wide panorama" in the prompt. The fixed presets top out at 16:9, but in our test auto returned a 6272x2688 image, the 21:9 size MiniMax lists for the model.
Should I turn off prompt enhancement?
Only when your prompt is already a complete brief. Turning the expander off (text to image only, with a fixed size) makes the model read your words more literally.
For the full model card, including the supported resolutions and the optional prompt expander, see MiniMax-H3-Image on Hugging Face.



