Image to prompt
Upload one picture and the tool writes a description of it that you can paste into an image model as a prompt. You decide which model the prompt is written for, how much detail it carries, whether a negative prompt comes with it, and which language it is written in.
Creative settings
Image to Prompt
The prompt structure will be adapted to the target workflow.
Advanced options
Generated result
Complete the settings and generate. Your result will appear here without leaving the page.
One picture in, one written description out
The tool reads an image and returns text. That text is a description detailed enough to be used as a prompt in a different image model, which is the whole point of the exercise: the picture is the starting point, the words are what you take away.
Nothing about this is a search for the original prompt. A prompt is not hidden inside an image, and no tool can recover the words somebody else typed. What this tool produces is a new description of what is visible, written in the vocabulary that image models respond to.
That distinction matters more than it sounds. If you feed it a photo you took yourself, the output describes your photo rather than reproducing whoever made it. If you feed it something you found, the output gives you a way to build something in the same direction without copying the file.
Five destinations, and what changes between them
The destination setting is not cosmetic. Different image models reward different writing, so the same picture produces a differently shaped description depending on which one you pick.
| Destination | What the description is shaped for |
|---|---|
| General image model | Plain sentences that any model can read. The safe choice when you have not decided where the prompt will go. |
| Nano Banana | Phrasing aimed at the Nano Banana family, which reads conversational description well and handles text inside the image. |
| GPT Image | Phrasing aimed at the GPT Image family, where instructions read as directions rather than as a scene description. |
| Midjourney | Shorter, weightier phrasing with the parameters that model expects. |
| Stable Diffusion / Flux | Comma-separated tags in an order that checkpoint models parse predictably. |
Pick the destination first if you already know where the prompt is going. Changing it later means running the picture again, because the shape of the text changes with it.
Three levels of detail, and when the shortest one wins
Detail level controls how much of the picture ends up in the text. It is not a quality setting, and longer is not better.
- Concise returns the subject and the broad treatment, which is what you want when you intend to add your own direction afterwards.
- Professional detail returns composition, lighting and materials alongside the subject, which suits a prompt you will hand to somebody else.
- Maximum detail returns as much as the tool can read, including incidental objects in the frame. Useful for reference work, and it also carries more of the original picture's accidents into your next generation.
The longest setting is not the most accurate one. Every extra clause is another thing a model will try to honour, including details you may not have wanted to keep.
Analysis focus is the field with no equivalent in the tools we measured
Every competing tool we looked at lets you choose a target model and an output format. None of them lets you say what to look at. This tool adds one free-text field for exactly that.
Write what matters here and the description leans that way. Ask for the lighting and the palette to be read closely and they are described in more depth than the objects on the table. Ask about a colour scheme and the text follows the colours.
It is the difference between a general description and an answer to a question you actually have. If you are rebuilding a look rather than a scene, this is the field that does the work.
The negative prompt is a switch, not a paragraph you write
Set the negative prompt option to yes and the tool returns two blocks of text instead of one: the description, and a list of what to keep out of the next generation.
On its own an image says very little about what should not be there. A described scene implies its own opposites, and this option is what turns that implication into words you can paste into the negative field.
Leave it off when the model you are targeting has no separate negative field. Turning it on costs nothing extra, but a second block of text is one more thing to carry.
Four output languages, Japanese among them
The description can be written in English, Simplified Chinese, Traditional Chinese or Japanese, independently of the language this page is in.
This is not a translation of a finished English prompt. The description is produced in the language you select, which matters for models that were trained with a particular language in the prompt.
It also means you can read the output in your own language while the picture and the target model stay the same.
How a run works, and what it costs
A run is one picture and one set of choices. There is nothing to install and no queue to wait in beyond the generation itself.
- Sign in. The tool runs against your account balance, so an account is required. There is no anonymous mode.
- Upload the picture you want described, by dragging it in or picking a file.
- Choose a destination model, a detail level, whether you want a negative prompt, and the language the description should be written in.
- Optionally write an analysis focus to steer what the description pays attention to.
- Press Generate. The run costs one point, charged per run rather than by the length of the answer.
Your last twelve runs are kept in this browser, so you can compare two attempts or recover a description you closed by accident.
What a written description can and cannot do
A description does not guarantee an identical image. This is the honest limit of the whole category, and it is worth stating plainly rather than burying it.
Two things stand between a description and a copy of your picture. The first is that a model generating from text starts from noise rather than from your file, so the result is a new image built to match the words. The second is that a description is lossy: whatever the tool did not mention is free to change, and every model fills silence differently.
What the description can do is carry the parts you cared about: the subject, the arrangement, the light, the palette, the treatment. Whether those survive the next generation depends on the model you send them to and on how much of the picture you asked to keep.
No tool in this category can promise a faithful reconstruction from text alone. Ours does not either.
Three prompts written for three different models
The descriptions below were published by another site alongside its own case pairs, and are reproduced here word for word so you can see what a usable prompt actually looks like. Each was written for a different target model, and the difference between them is the point.
A lone bicycle courier in a translucent raincoat waits at a crosswalk in a futuristic megacity at night. Towering dark buildings, cyan and amber neon reflected across rain-soaked asphalt, drifting street steam, and a distant elevated tram. Cinematic realistic concept art, street-level wide shot, strong depth, atmospheric haze.
A colossal bioluminescent blue whale glides through a sunrise sky above floating stone temples and waterfalls, tiny birds circling its fins and clouds drifting between the islands. Painterly epic fantasy concept art, whale sweeping diagonally across the upper frame, golden dawn backlight, luminous turquoise markings, serene and awe-inspiring.
Anime key visual of a young shrine guardian holding a white fox mask beside her face beneath a vermilion torii gate during snowfall. Deep red scarf, dark indigo winter robes, warm stone lanterns, clean expressive linework, refined cel shading, medium portrait, blue winter twilight, quiet and determined.
Read them side by side and the first two are continuous prose with the camera and the mood folded into the sentences. The third is written as a stack of discrete facts, because that is the shape a checkpoint model parses most reliably. The same picture described for the same destination comes back in that shape.
Where the seven comparisons on this page come from
The pairs below were published by two other image-to-prompt sites as their own demonstrations. We did not produce them, and they are not output from this tool.
Each pair shows the same idea twice: once as it was originally generated, and once regenerated from a written description of it. That is the task this tool performs, which is why the pairs are worth looking at, but the credit for the images belongs to the sites that made them.
Every caption names the source. The three from img2prompt.app are reproduced with the prompt that site published beside them; the four from img2prompt.art carry the description of what was held constant that the site itself wrote.
We publish them with attribution and claim no ownership. If you want to know how this tool behaves on your own pictures, the honest way to find out is to run it on one.
Examples
Before

After

A lone bicycle courier in a translucent raincoat waits at a crosswalk in a futuristic megacity at night. Towering dark buildings, cyan and amber neon reflected across rain-soaked asphalt, drifting street steam, and a distant elevated tram. Cinematic realistic concept art, street-level wide shot, strong depth, atmospheric haze.
The pair the source site files as Case 01, Cinematic City, written for Flux. The regenerated frame keeps the neon reflections, the wet asphalt and the night haze while moving the figure and redrawing the street, which is what a description carried faithfully looks like: the treatment survives, the exact arrangement does not.
Before

After

A colossal bioluminescent blue whale glides through a sunrise sky above floating stone temples and waterfalls, tiny birds circling its fins and clouds drifting between the islands. Painterly epic fantasy concept art, whale sweeping diagonally across the upper frame, golden dawn backlight, luminous turquoise markings, serene and awe-inspiring.
Case 02, Epic Fantasy, written for Midjourney. Both frames hold the same subject and the same golden backlight, and the composition moves: the whale crosses the frame at a different angle and the islands sit differently beneath it. A short prompt with weights behind it behaves like this.
Before

After

Anime key visual of a young shrine guardian holding a white fox mask beside her face beneath a vermilion torii gate during snowfall. Deep red scarf, dark indigo winter robes, warm stone lanterns, clean expressive linework, refined cel shading, medium portrait, blue winter twilight, quiet and determined.
Case 03, Anime Key Visual, written for Stable Diffusion. This is the pair that shows why the third prompt is built as a list rather than as prose: the scarf is named, the robes are named, the linework and the shading are named, and the regenerated frame keeps all of them while changing the camera.
Before

After

The source site describes this pair as keeping scale, depth and light intact across the regeneration. The skyline, the bridge and the height of the towers stay where they were; what moves is the time of day and the amount of haze, which makes it the narrowest of the four pairs and the closest to a literal redraw.
Before

After

The source describes this pair as preserving warm mood, framing and material cues. It is the most useful of the four for anyone working on interiors, because the furniture changes while the light and the palette do not, which is the line a description has to hold if it is going to be worth reusing.
Before

After

The source describes this pair as keeping palette, pose and painterly finish intact. It is also the clearest warning on this page: look at the faces. The colour scheme and the posture were carried, the person was not, and that is the normal outcome when a description replaces the file.
Before

After

The source describes this pair as retaining the subject, the mood and the soft morning light. It is the smallest subject on the page and the easiest one to read: the kitten is unmistakably the same idea, and it is a different photograph of a different kitten.
Questions
Does this tool keep the picture I upload?
The picture is sent for the run and the description comes back. There is no gallery of your uploads on the site and nothing of yours is published. Treat any upload as something you were willing to send, and avoid pictures you would not want processed at all.
Can it recreate an image exactly?
No, and no tool in this category can. Generation happens from text rather than from your file, and a description leaves things unsaid. What it carries is the subject, the arrangement, the light and the treatment you asked it to notice.
Which models can the description be written for?
Five destinations: a general image model, Nano Banana, GPT Image, Midjourney, and Stable Diffusion with Flux. The destination changes the shape of the text rather than only its vocabulary.
Do I need an account, and what does a run cost?
An account is required, because the run is charged to your balance. Each run costs one point, charged per run rather than by the length of the description. There is no free anonymous mode.
What does the analysis focus field actually change?
It decides what the description pays attention to. Whatever you name there is described in more depth and the rest of the picture is covered more briefly, which is useful when you are rebuilding a look rather than a scene.
Is the negative prompt a second run?
No. It is a switch on the same run. Turn it on and the answer carries a second block listing what to keep out of the next generation, which you paste into the negative field of the model you are using.
Which languages can the description come back in?
English, Simplified Chinese, Traditional Chinese and Japanese. The description is written in the language you choose rather than translated from English afterwards.
Are the example images on this page your own output?
No. They come from img2prompt.app and img2prompt.art, which published them as their own demonstrations, and every caption says which. They are here to show what the task looks like, not to advertise what this tool produces.
Explore next
AI image generator
Describe an image or upload a reference, choose a model, and read the quoted credit cost before generating.
AI video generator
Write a shot, animate a still, or drive a clip with a reference video, then read the quoted credit cost before the job starts.
From prompt to finished image
A worked walkthrough of one image task: choosing a model, writing the prompt, reading the credit quote and judging the result.