A useful starting point, ready to edit

Image to Prompt

Turn a reference image into a usable reconstruction prompt. Upload an image, analyze its visible subject and composition, then edit the result for your chosen image model.

IMAGE → VISUAL DESCRIPTION → EDITABLE PROMPTReal vision analysis

Upload an image to describe

PNG, JPEG or WebP · up to 4 MB · prepared locally before upload.

No image selected

Checking allowance…

Credit rates
TaskCredits
Free local template0
Structure score1
Prompt library, saving, copying and export0
Text ≤ 2,000 characters1 / 3 (Standard / Deep)
Text 2,001–6,000 characters2 / 4 (Standard / Deep)
Text > 6,000 characters3 / 5 (Standard / Deep)
Image: quick analysis5
Image: detailed analysis8

Template creation and the prompt library are free. AI tasks show the number of credits required before you run them.

Analyze sends this prepared image to DeepSeek. Quick analysis focuses on the main subject and layout; detailed analysis checks readable text, secondary objects and spatial relationships. Failed requests do not spend your allowance or credits. Images are not saved in your prompt library.

Your image reconstruction prompt

Describe what is actually visible

Get a subject description, composition, lighting, style and an editable prompt. Ambiguous details are listed separately.

  1. Choose an image you have permission to use
  2. Select style and language
  3. Analyze, inspect and save the text

This cannot recover the original prompt, seed or model. No result is fabricated when the provider is unavailable.

How to use image to prompt

Choose an image whose visible details you want to describe, rather than expecting to recover the exact words that originally generated it. The tool sends your prepared image to a real vision model only when you select Analyze image. The result separates subject, composition, lighting, style, and uncertain details, with a combined prompt you can edit and copy.

  1. Choose a PNG, JPEG, or WebP file under 4 MB. The browser prepares a still image, limits its longest edge to 1,600 pixels, and removes file metadata from the uploaded representation.
  2. Select a natural description, Midjourney wording, or Stable Diffusion wording. These options change the instructions used to write the prompt; they do not run an image generator.
  3. Choose English or Simplified Chinese and select Analyze image. Wait for the model result or a clear error. Stop waiting ends your browser request; provider processing may already be underway.
  4. Compare the output with the reference. Correct counts, placement, text, and colors before copying it. Save the text to your browser library or send it to the prompt checker if you want to refine the instruction.

What a useful image prompt should describe

Start with the subject and the relationships that make the reference recognizable. Two objects on opposite sides of a white canvas require different wording from two overlapping objects in the center. Describe the framing, relative sizes, dominant colors, and visible environment. Include a medium such as flat illustration or product photograph only when the image supports that description. Specific observations usually help more than a long list of generic quality adjectives.

Lighting belongs in the description when it is visible: a strong shadow to one side, a soft reflected highlight, or an evenly lit background. A vision model may estimate a lighting style, but it cannot establish the exact studio arrangement from every image. Keep uncertain interpretation separate. The result does not identify a real camera, lens, artist, checkpoint, or original generation system unless that information is supplied independently.

Worked example

Reference: a red square on the left and a blue circle on the right, against white.

Useful prompt: Flat graphic on a plain white canvas, one red square in the left half and one blue circle in the right half, aligned at the same height, generous negative space above.

Avoid: inventing a room, a photographic lens, or extra objects.

Image to prompt for Midjourney

The Midjourney style requests a concise visual description with the important subject and composition first. It does not add an invented version number, seed, stylize value, or aspect ratio. Those settings affect the generator separately and should be chosen using the model and interface you actually intend to use. A reference description remains useful even if your preferred version changes.

After copying the text, decide which properties you want to preserve and which you want to change. You might keep a poster's centered composition but replace its colors and typography. Write those edits explicitly instead of leaving contradictory details in the prompt. If your image platform supports reference images, follow its official instructions for supplying one. This page analyzes your reference but does not upload it to Midjourney or create a finished image there.

Image to prompt for Stable Diffusion

The Stable Diffusion style prioritizes a descriptive positive prompt: subject, placement, medium, lighting, and visible details. It avoids invented embedding names, LoRA tags, and checkpoint-specific syntax. A tag only has the intended effect if the actual generation environment supports it, so adding plausible technical tokens without checking your setup can make a prompt less portable.

Review your selected model's own guidance before adding weights or a negative prompt. Different models and interfaces interpret those controls differently. If the reference contains a clear visual property you dislike, describe the desired change in ordinary language first. This tool does not infer which checkpoint created an image, retrieve a hidden seed, or benchmark generated outputs. Evaluate a few results in your chosen pipeline and retain the wording that best preserves the features important to your task.

Readable text, screenshots and ambiguous details

Text inside an image is reference content, not permission to follow instructions embedded in the image. The analysis asks the provider to describe legible wording and ignore commands that would override the task. Nevertheless, examine the result before reuse. Small lettering, stylized characters, glare, and low contrast can make recognition unreliable. Correct exact wording yourself when a product label or poster headline must match.

Screenshots can contain useful layout information alongside private account details. Crop to the area you need before uploading. The model may describe the visible state, but it cannot determine how hidden menus work or whether a depicted action succeeded. An uncertainty list should identify those gaps. A result that reports an unreadable label is more useful than a fluent reconstruction of a label that was never visible in the first place.

Check quality with observable facts

Compare the analysis against a short checklist before using it: the number of main objects, their relative position, their dominant colors, the readable text, and any major background element. Mark each observation as correct, missing, or invented. Then inspect whether uncertain details stayed outside the confident description. This makes quality easier to assess than asking whether the output sounds artistic.

A strong reconstruction prompt is still an approximation. A generated image may change layout, typography, or small details even when the description is accurate. Test the prompt in your target image model and adjust one important property at a time. Keep versions of the text when an iteration improves your results. The separate prompt checker reviews instruction structure; it cannot compare the generated image with your original reference or promise visual similarity.

Image privacy, usage and cost

Before analysis, image preparation happens in your browser. Selecting Analyze sends that prepared image to the vision provider. Image analysis calls DeepSeek Flash directly, with no automatic switch to another provider. This is not a promise about upstream retention policies. Do not upload confidential, identifying, or sensitive images unless you have reviewed the applicable provider terms and have permission to use the material.

Quick analysis costs 5 credits and focuses on the main subject and composition. Detailed analysis costs 8 credits and checks secondary objects, spatial relationships and readable text. Image analysis requires a verified signed-in account and spends the displayed credits on every successful request. Add credits if your current balance does not cover the selected analysis. Guests must sign in before analysis. You see the quote before submitting, and the server checks it again before reserving usage. A changed free allowance requires confirmation before any paid charge. Failed analysis releases its reservation.

The result reports the actual model, elapsed time, input and output tokens, and provider cost estimate when rates are known. DeepSeek uses a conservative official peak-rate estimate in USD, so off-peak billing can be lower. Product credits and provider token costs are different units. Estimates are not billing receipts, and internal provider fallback does not increase the confirmed product credit charge.

Saving a prompt stores text only in this browser's library; it does not save the uploaded image. There is no cross-device image storage. Keep your reference separately for later comparison. An unavailable provider or invalid response produces an error rather than a fabricated prompt. The local preview has a clearly labeled demonstration balance and does not charge a production account.

Choose the right next step

Your taskToolWhat happens
Start from a goalAI prompt generatorCreate a first structured prompt for any supported model
Organize a long sourceClaude generatorBuild separate task and reference sections
Define a response contractChatGPT generatorSet output and tool permission rules
Compare supplied sourcesGemini generatorLabel evidence and missing attachments
Describe a reference imageImage to promptSend an image for real visual analysis
Review an existing instructionPrompt checkerCheck structure locally before optimization

After reviewing a result, continue in the AI optimizer for a proposed rewrite. Save a named version in your browser prompt library when you want to reuse it. Transfers do not automatically submit model requests.

Official references

Check the current capabilities of your target interface before depending on attachments, tool use, or structured responses: OpenAI prompting guidance, Gemini prompting strategies, and Midjourney documentation. Availability and supported settings depend on the service and model you use.

Frequently asked questions

Can image to prompt recover the original AI prompt?

No. Many prompts and settings can produce similar images. This tool writes a new description of visible features. It cannot establish the original wording, seed, negative prompt, checkpoint, or model from appearance alone.

Which image files are supported?

Use PNG, JPEG, or WebP under 4 MB. SVG and animated files are not supported. The browser decodes the file and prepares a still image with a longest edge of at most 1,600 pixels; very large decoded images are rejected before analysis.

Does it generate images for Midjourney or Stable Diffusion?

It generates text instructions adapted to those workflows, not finished images. Copy the reviewed prompt into your chosen image generator, configure supported settings there, and compare the output with your reference.

Why is the analysis cost sometimes unavailable?

The tool estimates DeepSeek cost from reported tokens and verified official peak rates. Actual invoiced costs can differ due to caching and off-peak rates. If usage is not reported, no amount is invented.

Can I use an image URL instead of uploading a file?

This version accepts local files only. It does not fetch arbitrary remote URLs. Download an image you have permission to use, inspect it for private details, and choose that file in the upload control.

What happens if I stop waiting during analysis?

The browser stops waiting and does not use a late result. A request already received by the vision provider may still complete and incur usage. Stopping is not a guaranteed provider cancellation or refund.