Image Prompts: Examples for Midjourney and DALL·E

Which parts make up an image prompt?

First make the image visible in one sentence. Then add only the details that affect the decision. The order below is not a mandatory formula, but it's a good checklist for seeing what you forgot.

Midjourney's official prompt guide recommends short, specific descriptions, exact numbers and writing the result you actually want. As detail decreases, variety increases and control decreases. This simple trade-off is often the answer to "why did the same prompt give a different result?"

Why are Midjourney and ChatGPT Images prompts not the same?

Midjourney works well with a short visual phrase and parameters written at the end. Parameters go at the end of the prompt, without punctuation. The current parameter list includes --ar for ratio, --no for unwanted elements, --raw for a less default style, --seed for repeat testing and --sref for style references.

ChatGPT Images, by contrast, lets you edit the result in stages within a conversation. Change instructions like "keep the cup on the left, change the background, don't touch the text" can continue in natural language. OpenAI's current ChatGPT Images documentation says it supports creating new images, uploading and modifying existing images, editing by selecting a region, and various ratios.

DALL·E 3 is still reachable via the separate DALL·E GPT, but OpenAI's API information lists it as a legacy model. Building a new workflow on DALL·E-3-specific "magic words" is therefore not a sound investment.

5 prompt examples for Midjourney V8.2

According to Midjourney's Version documentation, V8.2 has been the default since 24 July 2026. Parameter compatibility changes by version, so don't blindly copy every command from an old tutorial.

1. Product photo

   matte black reusable water bottle on a pale stone surface, three-quarter product view, soft side light, controlled shadow, clean commercial photography, empty space above for headline --ar 4:5 --raw

   For your own product, change the material, the colour, the camera angle and where the empty space sits. If the real logo on the product matters, placing it later in a design program is more reliable.

2. Blog cover

   editorial illustration of a founder organizing scattered notes into a clear system, white background, black geometric forms, one red accent, subtle paper texture, generous negative space, no text --ar 16:9 --s 150

   Write the empty space for the headline explicitly in the prompt. "No text" can be tried in the description and --no letters words logos as a parameter; still check the result by eye.

3. Baku street scene

   rainy evening on a quiet central Baku street, wet stone reflecting warm shop lights, one person walking away, eye-level documentary photography, restrained colors, natural atmosphere --ar 3:2 --raw

   The model can get a specific street and architectural details wrong. If there's a claim about a historic building, signage or a real location, don't publish it as documentary photography.

4. Minimal poster

   minimalist geometric poster about order emerging from chaos, thin black grid interrupted by one bold red line, off-white paper, Swiss editorial composition, no text, no logo --ar 4:5 --raw

   Here we translated the theme into a relationship of forms: a grid, a disruption and a red line. Writing the visible element instead of a broad word like "futuristic" makes the result easier to judge.

5. Light test in the same frame

   small home office with a wooden desk, open notebook and single plant, front-facing interior photograph, soft morning window light, neutral palette --ar 3:2 --seed 24817 --raw

   With the same --seed, change only "morning" to "late afternoon" and compare the light's effect. Midjourney presents seeds as a testing tool; it does not treat them as a way to keep a style or character permanent.

5 prompts for ChatGPT Images and DALL·E searchers

6. Social visual with short on-image text

   "Prepare a social media visual at 1080×1350. Use black typography on a white background and one red line. The text must be only 'A system doesn't reduce chance; it makes the error visible.' Don't change letters, don't write extra text or a logo; show a close-up so I can verify the text's accuracy separately before publishing."

7. Replacing a product background

   "In the product photo I uploaded, keep the container's shape, colour, label and all text unchanged. Replace only the background with a light grey studio backdrop, add soft light from the top left and a natural shadow under the product. The light falling on the label must not reduce the text's legibility."

8. Editorial illustration for an article

   "Create a 16:9 editorial illustration for the article 'Where should a small business start with AI?'. Show one person not among ten different tools but in front of a simple process map. Colours: white, black, grey and one red accent. Leave the right side empty for the article headline; don't create text in the image."

9. Editing a selected region

   "Edit only the bottom-right area I selected: replace the plastic cup with a black notebook. The person, table, light, camera angle, wall and colour balance must remain as they are. Match the new object's shadow to the existing light direction."

10. Three images in one visual system

    "From the reference I uploaded, take only the colour palette, margins, line weight and paper texture as a visual system. Create three separate topics in the same format: strategy, automation and measurement. Don't copy the reference's objects, text or logo. Each result should differ in composition while the system stays recognisable."

OpenAI's current Image Generation documentation explains creating and editing with GPT Image, multi-step changes, and controlling size, quality, format and compression. The documentation also openly notes that text accuracy, recurring-character consistency and precise composition can still fail.

How do reference images help keep style and objects stable?

One reference does not solve everything at once. In Midjourney, an Image Prompt affects content, composition and colour; a Style Reference is for carrying the overall visual feel without copying objects. The Style Reference documentation recommends writing a simple text prompt and reducing style words that conflict with the reference.

Style Reference works in V8.2, but Character Reference and the V7-era Omni Reference are not current V8.2 features. If the same character matters, check the model version and the compatibility table first. Merely copying --oref from an old prompt does not make it work in V8.2.

In ChatGPT Images, list in a sentence which parts of the reference should be preserved: the face, the clothing, the product shape, the camera, the palette or the background. "Make the same thing" is both imprecise and hides legal and plagiarism risk.

How is an image prompt tested and delivered?

1. Choose the place of use first. A blog cover, an ad, print and a presentation don't want the same ratio and level of detail.
2. Create one core composition. Don't move on to style exploration before the object, camera and empty space are settled.
3. Test one variable. If you change light, colour, angle and medium at once, you won't know which change affected the result.
4. Zoom in on text and anatomy. Fingers, eyes, reflections, logos, product labels and fine print need a separate look before publishing.
5. Keep the prompt and output together. Record the model, version, ratio, references, seed and the final human edits.
6. Export the file to fit the purpose. Check size and compression for the web, real pixels and colour requirements for print.

Before choosing a tool, read the creating images with AI guide for the full workflow, and the 10 rules of prompt writing for general prompt logic. An image generator can speed up design decisions; which visual fits the brand and the context is still decided by a human.

How are references, real people and copyright checked?

Before uploading photos, logos, illustrations and product packaging that aren't yours, check your usage rights. Midjourney's terms of service require you to hold the necessary rights to input material and note separate plan-related exceptions for commercial ownership. The legal outcome can vary by country, contract and usage; get a lawyer's opinion for commercial campaigns.

Don't use a real person's face in advertising or in a scene creating the impression of a real event without consent. OpenAI's usage policies prohibit misleadingly using a person's photorealistic likeness without consent regarding authenticity. The generate button working is not permission to publish.

Don't present the final file as fact. A visual distorting a product's colour, size and function in an ad; a non-event presented as a real photo in news; an unconsented likeness in a portrait — these are separate risks. Judge whether an AI label is needed together with the platform, advertising rules and the audience's likelihood of being misled.