How it works
The tool estimates relative—not measured—depth, removes isolated depth speckles with a 5 × 5 median filter, reduces only the lighting-analysis field to at most 256 pixels on its long side, and heavily smooths that field before deriving broad surface direction. A fixed low-left simulated key applies restrained warm light to key-facing regions and a cooler tone to key-away regions in linear-light RGB; intensity scales that numerical lighting change. At its working size, every output pixel changes source RGB only at that same coordinate: no face, eye, arm, body, text, object, or background pixel moves, copies elsewhere, spatially resamples, inpaints, or generatively repaints. Framing stays fixed, working-frame alpha stays byte-exact, and the result is a lossless PNG. It does not recover real scene geometry, replace existing illumination, or synthesize cast shadows, rim light, reflections, or hidden surfaces. Larger uploads can be proportionally resized first: the standard free limit is 800 pixels on the long side, and the depth service defaults to 2048 pixels for remaining larger inputs. Flat or mistaken depth, glass, mirrors, reflections, smoke, fine hair or foliage, and strong existing directional light can put the tone on the wrong surface or make the simulated direction conflict with the photograph; lower intensity helps but cannot correct the source lighting.
