How to Reverse-Engineer Any AI Image into a Prompt

Every image prompt is really six decisions wearing a trench coat: subject, style, lighting, composition, camera and colour. Once you can see those six axes in a finished image, writing the prompt that recreates it stops being guesswork.

Start with the subject, stripped to a noun phrase

"A woman" is the subject; "a woman in a red coat waiting at a tram stop" is the subject plus composition plus colour, and mixing them together is why prompts drift. Separate them first, then reassemble deliberately.

Style is the axis people underestimate

"Cinematic" does almost nothing on its own, because every model has a different idea of what that means. Naming a medium (35mm film, gouache, 3D render) and a concrete treatment (high contrast, muted palette, hard shadows) moves the result far more than any number of mood adjectives.

Lighting and camera decide whether it reads as a photograph

Overcast diffuse light, golden hour backlight and a bare on-camera flash produce three completely different images of the same subject. Focal length does the same for composition: 24mm and 85mm frame a face in ways no amount of prompt tuning will reconcile.

The usual mistake here is describing the mood the light creates rather than the light itself. "Moody" is a result; "single hard source from camera left, deep shadow on the far cheek" is an instruction, and it produces the mood as a side effect. The same applies to depth of field: say the aperture behaviour you want — background dissolved, or everything from the near edge to the wall in focus — rather than the word "professional".

Check your reading against the picture

Once you have named all six axes, the test is cheap: generate from your description and put the two images side by side. Nine times out of ten the gap is a single axis you read wrong — usually the medium, occasionally the focal length. Fix that one and stop; changing three things at once tells you nothing about which one mattered.

Do it by hand once, then let the tool do it

Doing this manually a few times is the fastest way to learn it. When you want the answer in seconds instead, reverse-engineer a prompt from an image — it returns the same six axes as editable fields, plus the prompt written out for Midjourney, Stable Diffusion and natural-language models. Going the other direction, from an idea to a prompt, is what the image prompt generator is for.