Point at it. That is the interface.
Unprompter is one panel and four ways to feed it. Everything else happens between your browser and the model you chose.
Step one
Four ways in, because images do not all sit in an <img> tag.
Right-click
Any image on any page: Unprompt this image in the context menu. The panel opens with it staged.
Hover badge
Hover an image bigger than 128px and a small Unprompt button appears in its corner. Turn it off in the options if you would rather not see it.
Screen region
Alt+Shift+U and drag a box. Works on canvases, map tiles, video frames, PDFs and CSS backgrounds — anything you can see.
Drop or paste
Drag a file into the panel, paste from the clipboard, or pick one from disk. Local files never touch a page at all.
A whole page
Unprompt every image on this page finds every image over 96px, including CSS backgrounds, and reads the ones you tick — one after another so your provider does not rate-limit you.
From history
Everything you have read stays in the panel, searchable across every field and every generated prompt. Click one to bring it back.
Step two
One reading, and the parts of it we can measure rather than guess.
The image is resized to a 1024px long edge and re-encoded — which is also where EXIF dies — then sent once to your model with a request for a structured reading, not a caption.
What the model is asked for
Subject and details · scene, foreground, background and props · framing, angle and compositional rule · shot type, lens, aperture and focus · key light, direction, quality and time of day · colour grade, contrast and saturation · surface, medium, grain and finish · emotion, atmosphere and energy · genre, era, influences and render · any lettering in the frame · and a negative list of what must not appear.
What is measured instead
The colour palette is read off the pixels in your browser: six dominant colours, named and hexed, with the warm/cool, saturation and key reading derived from them. The aspect ratio comes from the real dimensions. Models are confident and often wrong about both, so the measurement wins and the model's own read is kept alongside it.
Step three
Nine formats, one bill.
Switching between Midjourney and SDXL, dialling faithfulness from loose to verbatim, or tapping a variation all rewrite the prompt from the reading you already have. The provider is called once per image, full stop.