About

What this site is

You have a reference image and you want something in the same vein — the same light, the same palette, the same framing. Writing that description by hand is slow, and most people are not good at naming colours or describing how light falls across a frame.

So this tool does the measuring for you. Drop an image in, pick the model you are prompting, and it writes a prompt from what is actually in the picture. The output is formatted differently for Midjourney, Stable Diffusion, Flux and DALL·E, because those four do not respond to the same shape of prompt.

Who runs it

Beixiangdexing, an independent developer. Not a company, not a team with a support desk — one person who got tired of typing out colour descriptions by hand and built the thing.

That is also why the site is free. It costs nothing to run beyond hosting and a domain, and advertising covers that. There is no premium tier, no credits to buy, and no plan to add one.

When this site started

The oldest date on this site is 16 September 2026. That is the earliest entry in the sitemap, and it is the date this project is counted from — everything here has been added one page at a time since.

How the numbers on these pages were produced

Every figure on this site comes from one of two places, and both are public. The first is the official documentation of the model a page is about: where a page quotes a field name, a size or a limit, it says which document that came from. The second is this site's own code. The measuring and the prompt-writing both run in your browser, and the file that does it is served openly at /script.js, so you can read it and run it on your own image.

You do not have to take any of it on trust. Drop an image into the tool and the Read from your image panel prints every value that was measured, so any phrase in a generated prompt can be traced back to the number behind it. Nothing on these pages is an estimate.

One worked example, so the chain is checkable end to end. The first of the three test images this site keeps is night-fireworks.png, 640×360. Dropped into the tool it is drawn onto a 192×192 canvas — 36,864 cells, the constant var W = 192, H = 192; on line 49 of /script.js. Each cell's luma is computed on line 65, lum[i] = 0.2126 * r + 0.7152 * g + 0.0722 * b;, and averaged on line 86, var meanL = sumL / n;, which returns meanL = 34.51. That one number is then compared against the exposure cut-offs on lines 189–192 of the same file: 34.51 is not above 175, not above 130 and not above 80, so it falls to the last branch and the phrase written into the prompt is dark, low-key lighting. No model is involved at any step — the phrase is the last line of a four-way comparison over one average. The panel above shows this value as 34.5 / 255, because it prints one decimal.

Three things this site will not do

These are deliberate choices, not gaps waiting to be filled:

• It will not ask you to sign up. There is no account system. If a tool needs your email address to describe a photograph, that tells you something about how it is funded.
• It will not limit you. No daily credits, no five-images-a-day, no counter. Those limits exist on most image to prompt tools because each run costs their server money. Here the work happens on your device, so there is nothing to ration.
• It will not upload your image. Everything is measured in your browser. This is the whole reason the other two are possible.

What it is honest about

The tool measures colour, light, contrast, detail and framing, and writes those up accurately. It does not recognise faces, objects, brands or scenes, and it will not guess at them — a prompt that confidently describes a subject that is not in your picture is worse than one that stays quiet about it.

Every value it measured is shown to you under the prompt, so you can see exactly which numbers produced which phrase. If you want to know more about that boundary, the FAQ covers it, and so does How it works.

This site runs the other direction — image in, prompt out. For the reverse, see what actually decides the result in prompt to image.

Getting in touch

Bugs, wrong output, or a format that will not load — all worth reporting. The contact page has the address and what to expect.