Text Behind Image

What this text behind image tool does

This tool builds the popular text-behind-subject effect out of three stacked layers on a single canvas. Layer one is your original photo, untouched. Layer two is your text. Layer three is a cut-out of the main subject, saved as a transparent PNG and drawn pixel-for-pixel over the same coordinates as layer one. Because the top layer carries the exact same pixels the photo already had wherever it keeps something, the subject appears to stand in front of the words while the rest of the scene stays behind them. Nothing is erased or masked away, and the illusion comes entirely from the drawing order. The cut-out comes from an ISNet salient-object segmentation model that runs through onnxruntime-web on your own machine. The model weights are roughly 40MB and download once, after which the browser serves them from cache. There is no upload step, no account, no queue, and no per-image cost, which is why the page has no watermark or usage cap. After the model returns its mask, a second pass erodes the mask boundary by one pixel and unmixes leftover background colour out of the semi-transparent edge pixels. That step matters more here than in a plain background remover: the subject sits on top of your text, so an edge pixel still tinted with the old background colour would read as a coloured fringe drawn across the letters. The one-pixel erosion costs nothing in this composite because the pixel underneath is the identical pixel from the untouched photo on layer one. Images larger than 2560 pixels on their longest edge are scaled down to that size first, because the edge cleanup is a per-pixel pass with a neighbourhood probe and a 24 megapixel phone photo would stall the browser for many seconds. The export is always a PNG so that partial text opacity and the soft subject edges survive intact.