How to prompt GPT Image

A vertical social ad. Scene: deep navy studio background with a soft radial glow. Subject: a matte-white wireless earbuds case, open, floating at a slight angle, crisp reflection below. Headline text: "HEAR EVERYTHING." in a bold geometric sans-serif, white, centered in the top third. Smaller text below the product: "Wireless. 40-hour battery." in light gray. Intended use: paid social ad. Constraints: exact text as written, no extra characters, no other text anywhere.

Quoted copy plus a font description plus "no extra characters" is the verbatim-text recipe. Labeled segments (Scene / Subject / Text / Constraints) keep a dense brief unambiguous — the model follows the structure, not just the words.

A clean vertical infographic titled "HOW COLD BREW IS MADE" in bold dark-blue letters at the top. Four numbered sections stacked below, each with a simple flat icon and one short label: 1. "Coarse grind", 2. "Steep 18 hours", 3. "Filter twice", 4. "Serve over ice". White background, generous white space, consistent 2-color palette of dark blue and warm orange, flat vector style. All text exactly as written, readable and horizontal.

GPT Image handles structured layouts — numbered sections, icon plus label pairs, consistent spacing — that keyword-style models scramble. Dense text layouts need quality high; every label is quoted so nothing gets paraphrased.

A real photograph, morning window light. Scene: a pale oak café table near a window. Subject: an iced matcha latte in a tall ribbed glass, distinct green-to-white gradient, condensation beading on the glass, a metal straw. Key details: close-up at eye level, shallow depth of field, blurred café interior behind, natural soft shadows. Intended use: menu photography. Constraints: no people, no text, no logos.

"A real photograph" plus camera-adjacent language (eye level, shallow depth of field) engages the realism mode more reliably than stacking "photorealistic 8K". The constraint line cleanly excludes the usual drift: stray text and invented branding.

A desktop dashboard UI for a habit-tracking app, presented as a straight-on screenshot. Layout: left sidebar with the app name "Streaks" and five nav items ("Today", "Habits", "Stats", "Friends", "Settings"), main area with a weekly grid of colored completion dots and a large card reading "12-day streak". Light theme, soft gray background, one accent color: teal. Realistic interface spacing and hierarchy, real UI text only, no lorem ipsum, no concept-art flourishes.

Describe the product as if it already shipped — real nav labels, real hierarchy — and ban concept-art language explicitly. Quoting every UI string keeps labels crisp and prevents invented menu items.

A photoreal packshot of a cylindrical tea tin on a white sweep. Front label layout, top to bottom: brand name "KOYAMA" (spell it exactly: K-O-Y-A-M-A) in a serif, dark green; below it "Roasted Hojicha" in smaller caps; a thin gold rule; then "Loose leaf - 80g" at the base. Soft even studio lighting, slight top-down angle. Constraints: all label text exact and correctly spelled, no additional words on the label.

Letter-by-letter spelling is the documented fix for brand names the model wants to "correct". Describing the label as an ordered layout (top to bottom) gets typography hierarchy right on the first pass.

A die-cut sticker illustration of a smiling cartoon avocado wearing sunglasses, thick white sticker border, bold flat colors with simple cel shading, slight drop shadow inside the sticker edge. Transparent background, nothing outside the sticker outline.

GPT Image 1.5 exposes a background control — set it to transparent and export PNG for a drop-in asset. The "nothing outside the outline" constraint keeps the alpha clean of stray marks.

Change only the wall color behind the sofa to sage green. Keep everything else exactly the same: the sofa, cushions, rug, plant, window light, shadows, and framing must be identical to the original photo.

The change-plus-preserve contract is the core editing pattern: one sentence for the change, one for the lock list. Anything you leave unlisted the model may reinterpret — the preserve list is not politeness, it is the spec.

Replace the man's gray t-shirt with a navy chunky-knit crewneck sweater. Do not change his face, facial features, skin tone, hair, body shape, pose, or identity in any way. Keep the background, camera angle and lighting identical, and make the sweater fabric fold and drape realistically with the existing light.

Identity-sensitive edits need the explicit lock phrasing — face, features, skin tone, pose, identity — repeated on every iteration to prevent drift. Asking for realistic fabric behavior under the existing light is what stops the garment looking pasted on.

Image 1: a product photo of a green glass skincare bottle. Image 2: a bathroom shelf scene with morning light. Place the bottle from Image 1 on the shelf in Image 2, centered on the middle shelf, matching the scene's lighting, adding a soft natural shadow under the bottle. Keep the bottle's label, proportions and color exactly as in Image 1; keep the scene otherwise unchanged.

With multiple input images, refer to each by index and state the interaction explicitly — which element travels, where it lands, whose lighting wins. Locking the label and proportions keeps the product legally accurate.

Make the lighting slightly warmer, like late golden hour. Change nothing else — composition, subject, colors of the objects, and all text stay exactly as they are.

The documented refinement loop: get a clean base, then one small change per pass instead of piling adjustments into a single prompt. Each micro-edit re-states the freeze so successive passes don't accumulate drift.