You found a picture with exactly the light, the colour grade or the composition you want. Describing it from scratch takes a while, and the words you pick may not be the ones an image model responds to. Image to prompt does the describing for you: upload the picture, and you get back a prompt that recreates it on the model you choose.
This guide covers when to capture the whole picture, the style only or the subject only, how to edit the result into something reusable, and the limits worth knowing.
How image to prompt works
A vision model looks at the picture and reads what is in it: the subject, the setting, the framing and camera angle, the lens and depth of field, the direction and colour of the light, the medium and the mood. It then writes those observations as a prompt in the syntax your target model follows.
That last part matters. The same picture needs a different prompt depending on where you paste it:
- Midjourney gets comma-separated phrases ending with --ar and --stylize.
- Nano Banana, GPT Image and FLUX get full natural-language sentences.
- Stable Diffusion gets weighted tags and a negative prompt.
- Seedream gets one or two concise sentences.
- Any model gets a plain paragraph that works in most tools.
The image to prompt tool also reads the shape of your picture and uses the nearest standard aspect ratio, so a vertical phone photo comes back as a vertical prompt.
Three ways to capture a picture
The whole image
Choose this when you want a close copy on another model, or a full prompt you will edit by hand. You get the subject and what they wear or hold, the setting, the composition and camera angle, the light, the palette and the style. Any text visible in the picture is copied in double quotes.
Here is the kind of prompt it produces for a product photo, written for Nano Banana:
Style only
Choose this when you like the look but not the content: a lighting setup, a film grade, an illustration technique. The subject is replaced by the placeholder [SUBJECT], and any specific place by [SETTING], so the prompt becomes a template.
Fill in the brackets with something new ("an old fisherman mending a net", "a harbour at dawn") and you have a fresh picture in the same visual language. This is the most useful mode for building a consistent set of images for a brand or a social feed.
Subject only
Choose this when the thing in the picture matters and the scene does not: a product, a character, a piece of furniture. The prompt describes shape, materials, colours, clothing and pose closely, and leaves [SETTING] for you to fill.
Swap [SETTING] for "a sunlit bathroom shelf" or "a dark wooden bookcase at night" and you can stage the same object anywhere.
Asking for one change
Often you want the picture, but slightly different: night instead of day, winter instead of summer, a teapot instead of a mug. Type the change in the optional note before you generate, and the prompt is written with the change already applied. This is cleaner than editing the prompt afterwards, because the light, colour and mood are rewritten to match the change instead of contradicting it.
Keep the note to one change. Two or three small ones work, but a long list turns the result into a different picture.
Making the prompt reusable
A prompt that recreates one image is useful once. A prompt with clear slots is useful every week. After you get the result:
- Find the parts you will change. Usually the subject, the setting and sometimes the colour palette.
- Turn them into placeholders. Square brackets work in any tool and make the slots obvious: [PRODUCT], [BACKGROUND COLOUR].
- Keep what makes the look. Light, lens, grade and composition are the parts you liked. Leave them exactly as written.
- Keep the original with it. Store the prompt in your notes next to the picture it came from, so you can check later whether a new result still matches the look.
Good pictures to start from
- Clear, well-lit images give the most specific prompts. A blurry, dark photo produces a vague description.
- One main subject is easier to describe than a crowded scene. For a busy picture, crop to the part you care about first.
- Your own photos and renders are the safest source. If you use someone else's picture as a reference, use the prompt to learn the look, not to copy their work.
Limits worth knowing
- It describes, it does not copy. The model reads the picture and writes words, so expect the same idea and look, not a duplicate. For exact likeness, edit the original photo with Nano Banana instead: see our Nano Banana photo prompt guide for the identity line that keeps faces stable.
- People are described, never identified. You get age range, hair, clothing, pose and expression, never a name.
- It will not describe sexual images, children in suggestive situations or graphic violence.
- It does not invent the opposite of what it sees. A night scene will not come back with golden-hour sunlight.
Run the prompt where you wrote it
Once you have the prompt, copy it into Midjourney, Gemini or ChatGPT, or run it in the studio on Nano Banana, GPT Image or FLUX and compare the result with your original side by side. If you would rather start from an idea than a picture, the image prompt generator writes the same kind of prompt from a sentence. For finished examples with our own renders, browse the image prompt library.
