Image to Prompt Generator

A prompt is the only control you have over a text-to-image model, and most people lose it twice: once when they cannot describe what they want, and again when a picture they like gives them no way back to the words that made it. Reverse-engineering is the practical answer — you have a reference image, a colour mood, a lighting setup or a composition that already works, and you want a prompt that lands in the same place. Doing that by eye is slow and unreliable, because the things that actually steer a model are boring and measurable: how saturated the palette is, whether the contrast is hard or soft, whether the light reads warm or cool, and how wide the frame sits.

This tool measures those things directly from your pixels. The image is decoded onto a canvas, sampled down, and every pixel is counted into colour buckets to find the dominant palette — with each colour converted into a name a model understands, so the prompt says deep azure instead of handing over a hex code. Mean brightness and the spread around it produce brightness and contrast, the balance between red and blue gives colour temperature, and a gradient scan scores edge density to separate clean minimal artwork from intricate, texture-heavy photography. Width and height are reduced to a true ratio, and if the file carries an alpha channel the prompt notes that the subject is already isolated.

Where the prompt goes matters as much as what it says, so the output is written per model rather than once for all of them. Midjourney gets a descriptive sentence with trailing parameters for aspect ratio, version, style and stylize, plus a –no list. Stable Diffusion gets comma-separated tags with quality modifiers in front and a separate negative prompt covering blur, artefacts and watermarks. DALL-E 3 gets a natural-language paragraph, because it largely ignores tag lists. Flux gets prose with no parameters at all. Use it to rebuild a look you liked, to keep a batch of campaign images consistent, or to hand a client a prompt that matches the mockup they already approved.

One boundary is worth stating plainly. This tool does not recognise subjects: it cannot tell a bicycle from a bicycle rack, because that needs a vision model on a server and would mean uploading your picture. What it does is translate measurable facts into the words models respond to, which is the part people are worst at guessing. You supply the subject in one line, and everything after it is read from the actual pixels. Decoding, measuring and writing all happen in your browser, so nothing is uploaded and there is no account, no API key and no daily limit.