Image Generation Prompts: 8 Examples for Product, Ads and Social Media

The first problem with a product photo is rarely resolution. The image looks good, but the packaging has changed. Or the product is right, and there is no room left for the campaign headline. A good image-generation prompt describes what must stay the same as clearly as what you want.

This guide contains eight prompts for product photos, ad creative, blog covers and social content. Each example gives the required input, a copyable command and the specific thing to check in the result. Start with the example closest to your need, fill the brackets and delete the details you do not use.

How do you write an image-generation prompt?

An image-generation prompt tells the model the subject, the intended use, the composition and the details that must not change. Instead of asking for a “professional look”, explain where the photo will be used and which detail must be readable.

Before you start, complete one sentence: “In this image, the viewer will notice [one detail] first.” If you cannot finish it, do not add more decor to the scene; narrow the brief first.

These seven fields turn a prompt from a wish list into a usable brief:

OpenAI’s imagegen guide recommends stating each reference’s role and naming the features to preserve during an edit. If the second image is only a colour reference, say so; otherwise you have asked the model to merge two products.

Create an image of [subject] for [intended use].
Scene: [location or background].
Composition: [framing, subject position, space for text].
Light: [direction and character].
Details to preserve: [what must not change from the reference].
Text in the image: [exact copy or "no text"].

Set the aspect ratio and file size in the tool when those controls exist. Writing “4K” in the prompt does not mean the file will be delivered in 4K. Likewise, “cinematic” on its own is not a lighting or lens decision.

The loop from brief to output

Image generation is not a one-shot command but a short review loop. Write the brief, build the prompt, generate and judge one problem only. If the result misses, fix one variable instead of changing everything at once, so you can tell which change worked.

Review in this order: product shape, text readability, light direction, framing and only then overall aesthetics. If you start with aesthetics, you often notice the changed product too late.

Eight practical examples

Each example solves one job. Replace the bracketed parts with your own product and delete the sentences you do not need. Do not expect the same prompt to produce the same result across models.

1. A clean product photo without changing the product

In a catalogue, the customer wants to see the product itself. Get the product shape right before adding decor. Input: a sharp photo with a single product where possible. Use: product detail page. Ratio: 1:1.

Turn the flask in Image 1 into a clean photo for a product detail page.
The flask should stand upright on a light grey, seamless background, with a soft contact shadow beneath its base. The camera should look at the product from an angle close to eye level. Keep the whole product in frame.
Light should feel like a large soft source from the left; let a narrow, natural highlight appear on the metal surface.
Keep the body proportions, cap mechanism, surface colour and markings identical to the reference. Do not add accessories, slogans or decor.

Check: compare the cap, edges and logo with the original. If the label breaks, keeping the real product cut-out and editing only the background can be more controllable than regenerating. To fix: “Soften only the shadow under the product. Do not change the product’s shape, colour or framing.”

2. Ad creative that leaves room for the headline

If a headline will sit on the image later, reserve that space before generation. Empty space should not be a design mistake to rescue afterwards. Input: product photo. Use: social ad. Ratio: 4:5.

Create a vertical campaign photo for the desk lamp in Image 1.
Place the lamp in the lower-right area on a light oak desk. The background should be a warm off-white wall. Keep the upper-left area plain and low-contrast for a headline to be added later.
Let the lamp create a narrow pool of light on the desk. Keep the room’s ambient light low, but do not leave the product in darkness.
Keep the lamp’s arm joints, base shape and colour as in the reference. Do not add text. Do not place other electronics in the frame.

Check: place the actual headline and look at it on mobile. Confirm the empty area and the product do not overlap. Image generation alone does not prove ad performance; the live result is measured separately.

3. Change the background and keep the material

When the background changes, the product’s light may not match the new setting. The surroundings can look natural while the product looks cut out and pasted in. Input: a photo of a ceramic plate or similar product. Ratio: 4:5.

Change only the setting around the ceramic plate in Image 1.
Place the plate on a linen-covered table in a kitchen with morning light. A blurred window frame should appear in the background. Keep the plate at the same angle and size.
Preserve the rim thickness, glaze pattern and colour transitions from the reference. Add a soft contact shadow that matches the new light. Do not put food on the plate; do not redesign how the product is used.

Check: a clean background matters, but so does the plate sitting on the table. If it floats, has a double edge or loses its pattern, spend the next fix on that.

4. A natural expert portrait from a reference

Over-retouching an expert portrait can reduce recognisability. Skin texture and real facial proportions matter more than a glowing background. Input: a portrait you have the right to use, with a clear face. Ratio: 4:5.

Create a natural portrait edit that places the person in Image 1 in a simple work setting.
Preserve their facial features, hairline, eye colour and apparent age. The person should look at the camera with a relaxed expression; keep shoulders and head fully in frame.
Use a light-coloured wall behind them and a single bookshelf out of focus. Soft window light should illuminate one side of the face. Keep natural skin texture; do not turn the face into a smooth plastic surface.
Do not change clothing or accessories. Do not add text.

Check: do not judge recognisability from a small preview. Inspect eyes, teeth, glasses frames and hair edges at full size. An identity-preservation instruction is not a guarantee.

5. One idea per blog cover

A “digital marketing” cover does not need a laptop, rocket, target and rising chart in one frame. Choosing one idea from the article makes the cover clearer. Input: none. Use: content strategy article. Ratio: 16:9.

Create an editorial cover illustration for a blog article about search intent.
On an off-white table, three different paper paths lead to the same door. One path starts at a question mark, one at a small product drawing, one at an open notebook. The door should be a simple drawing; the scene should not look like a real architectural plan.
Use cut-paper textures and thin pencil lines. Limit the palette to dark blue, pale blue and off-white. Leave space in the upper left for a headline. Do not add text, numbers or fake data charts.

Intended output: an illustration that carries the article’s idea. It is not a data visualisation. If real proportions or results are being explained, build the chart from the data.

6. A controlled variation for the same campaign

If light, angle, product and colour change together across two images, it becomes hard to tell which one affected the result. Start the design comparison with one variable. Input: approved campaign image. Ratio: same as the original.

Create an alternative from the campaign image I uploaded.
Change only the background colour from warm cream to very light blue. Keep the product’s colour, position, scale, light, shadow and camera angle the same. Do not change the reserved headline area. Do not add new objects or text.

Check: compare the original and the alternative side by side. If the product itself changed, this is not a controlled variation. When testing performance, account for differences in audience, offer and placement too.

7. A consistent illustration series for a carousel

Describing each card independently can pull the visual language apart. Approve the first card, then use it as a style reference for the rest. Input: none for the first card; the approved first card for later ones. Ratio: 4:5.

Create an illustration for the first card of a carousel about a marketing team’s content planning process.
A single adult at a desk chooses one of three blank notes in front of them. Use hand-drawn thin outlines and limited colour fills. Keep a light background, dark blue lines and one coral accent.
Keep the illustration in the lower half; leave the upper half empty for a headline added later. Do not write letters, icons or numbers on the notes.

For the next card: “Use Image 1 only as a style and character reference. The same person places the chosen note on the board. Keep the line weight, clothing, colours and the empty space at the top.” Characters and colours should match across the series; generate each card separately rather than asking the model for four tiny cards inside one image and enlarging them later.

8. A start frame built for video generation

If the image will be animated later, leave room for the camera. A start frame with the object pressed against the edge gives a wide movement less space and needs more reinterpretation. Input: product reference. Use: image to video. Ratio: 9:16.

Create a start frame from the tote bag in Image 1 for a short video later.
Place the bag on a light wooden chair. Show the handles fully and leave space around and above it for camera movement. Let window light from the left reveal the fabric texture.
Keep the setting a simple workshop corner. Preserve the bag’s stitching, print, colour and body proportions. Do not add people or motion blur.

Once the frame is approved, continue with the image-to-video example in the video guide. Fix the still first; trying to correct it inside the video costs extra attempts.

When the result is not what you wanted

Fixing one problem per attempt makes it easier to see which change worked. The table below shows the first intervention for common problems:

Store the approved image, the prompt and the settings together. That record is the most reliable way to reproduce the same visual language and to understand what changed later.

Examples published by the sources

The images discussed in this section are not outputs of the original prompts above. They are examples published by their sources, separated to show which design decision to examine. Instead of copying third-party images onto our server, we link to the source; licensing and attribution terms can differ per asset.

Google example: describing a product’s material with light

In the black ceramic mug example in Google’s Gemini image-generation guide, the detail to study is how the dark product separates from the background and how the surface reads. Rather than copying the scene for a flask or ceramic product, apply that separation to your own product. The image is a source example; it was not generated or altered for this page.

Community example: separating the portrait and the text area

In Nicolechan’s portrait quote card, the layout is built so the person and the long copy do not compete. When making your own expert card, use a real, verifiable quotation. If the text must be exact, generate the portrait separately and add the copy in a design tool. This is a community example, not an independent model comparison.

Plan your first attempt

Fill in the fields below before generating. The goal is not a perfect brief but knowing what to check in the first attempt:

• Use — Where will I use the image?

• First detail — Which detail should the viewer notice first?

• Reference — Which reference will I use, and what is its job?

• Preserve — Which three details must not change?

• Text — Will I add the copy during generation or afterwards?

• Check — Which single problem will I check in the first attempt?

Frequently asked questions about image prompts