Prompting for images

A practical method for AI image prompts on Sunra: a reliable structure, weak-to-strong rewrites, the style and lighting vocabulary models understand, references, and fixes.

A good image prompt reads like a brief you'd hand a photographer or illustrator: what's in frame, the style, how it's composed, and how it's lit. Cover those four and you'll get a usable image far more often than with a one-word prompt.

The Sunra image generator's prompt bar — set the model, aspect ratio, resolution, quality, and count; the credit cost shows on Generate.

A reliable structure

Write in this order — it mirrors how a shot is actually planned:

Subject — what's in frame, specific. "A ceramic coffee cup on a linen napkin."
Style — the medium and treatment. "Editorial product photo," "flat vector illustration," "3D render," "watercolor."
Composition — framing and angle. "Close-up, top-down, centered, shallow depth of field."
Light & mood — "Soft morning light," "neon night," "studio softbox, high-key."

Editorial product photo of a matte-black wireless earbud case on a wet stone surface, top-down, shallow depth of field, soft diffused studio light, minimalist, cool tones.

From weak to strong

The same idea, sharpened by adding subject specificity, then style, then light:

Prompt	Result
"a coffee cup"	A generic cup, random style and lighting
"a ceramic coffee cup on a linen napkin"	Right subject, but flat and styleless
"editorial photo of a ceramic coffee cup on a linen napkin, close-up"	On-brief composition
"editorial photo of a ceramic coffee cup on a linen napkin, close-up, soft morning window light, shallow depth of field"	The shot you actually wanted

Each added clause removes a decision the model would otherwise make for you.

Note

There's an upper limit on prompt length (it varies by model), and Sunra won't trim an over-long prompt — it fails instead of running. Put the essentials first; if you're piling on clause after clause, you're past the point of diminishing returns anyway.

Vocabulary the models understand

Reach for concrete terms instead of vague adjectives — models map these to real visual patterns:

Medium — photo, illustration, 3D render, oil painting, line art, isometric, claymation.
Shot & lens — close-up, wide shot, macro, top-down, eye-level, 35mm, bokeh, fisheye.
Light — golden hour, backlit, rim light, softbox, hard shadow, high-key, low-key.
Mood / palette — muted pastels, high-contrast, monochrome, warm tones, cinematic.

Tip

One precise term beats three fuzzy ones. "Backlit at golden hour" tells the model far more than "nice lighting."

Say what you want, not what you don't

Models handle positive descriptions far better than negations. Ask for "an empty, minimalist desk," not "a desk with nothing on it." If you'll add text or a logo on top later, prompt for negative space — "lots of empty sky above" — rather than describing what shouldn't be there.

Work from reference images

Attach up to 10 reference images to guide style, composition, or a specific subject. While writing the prompt, type @ to point at a specific attached image:

Put @Image1 on the table in @Image2, matching the lighting of @Image2.

It's the cleanest way to combine several references into one shot — see Annotate edits & reference images for the full reference and editing workflow.

Match the prompt to the model

The same prompt carries across models, but each rewards a slightly different emphasis:

GPT Image 2 — spell out any in-image text exactly, in quotes; it renders words more reliably than the rest.
Seedream 5.0 — lean into photoreal detail (skin, gaze, depth); it reasons about realism well and also handles text in images.
Midjourney V8 — give it mood and style words; it leans cinematic and painterly by default.
Nano Banana Pro — for precision edits, describe just the change ("make the jacket red"); it edits locally without a mask, and is also strong at in-image text.

Tip

Rendering words inside an image is the hardest thing for most models. If your design needs legible text — a sign, a label, a poster — reach for a text-strong model and put the exact words in quotes: a neon sign reading "OPEN 24 HOURS". See GPT Image 2 text rendering.

Not sure which to use? Choosing a model breaks down all of them.

Warning

Coming from Discord Midjourney? Don't type parameter flags like --ar, --v, or --style into the prompt — Sunra parses them as words and the model rejects the whole generation. Use the aspect-ratio, version, and style controls in the prompt bar instead.

Fixing common problems

Problem	Try this
Garbled text in the image	Switch to GPT Image 2; put the exact words in quotes
Wrong subject emphasis	Put the subject first; cut background clutter from the prompt
Inconsistent character across images	Use a reference image and a consistency-strong model like Nano Banana Pro
Almost right, one detail off	Don't re-roll — edit the result or annotate-and-edit just that area
Style keeps drifting	Name the medium explicitly and provide a reference image

Iterate deliberately

Change one variable at a time — model, then light, then composition — so you learn what each move does. When a batch is close, switch to editing rather than rewriting the whole prompt: fix a word with text edit, change the camera with multi-angle, or refine the background instead of starting over.

Ready to carry these instincts into motion? Read Prompting for video.