Object Remover

Painting with the brush does not erase anything directly. It builds a mask: a record of which pixels you have marked as unknown. The original colours in that region are then discarded and reconstructed from what surrounds them.

The reconstruction uses Telea's fast marching method. Rather than filling the region all at once, it works inwards from the boundary, always processing the unknown pixel closest to known image data next. Each pixel it reaches is estimated as a weighted average of the known pixels within the inpaint radius. The weights favour neighbours that are closer, that lie along the direction the image gradient is already travelling, and that sit near the advancing front, which is what allows edges and colour transitions to continue into the gap rather than simply smearing an average across it. As each pixel is filled it is treated as known, so the front advances until the whole masked region is complete.

The inpaint radius sets how far that neighbourhood search extends. A larger radius blends more surrounding pixels into each estimate. Finally a short post-processing pass softens the seam where the filled region meets the untouched image.

Load a 1600 by 1067 pixel photograph of a lawn with a litter bin standing in front of the grass. The panel reports the image dimensions once it loads.

Set the brush to around 40 pixels, large enough to cover the bin in a few strokes without excessive precision. Paint over the whole bin, going a little beyond its outline and including the shadow it casts, since a shadow left behind gives the removal away. The counter reports the painted area, in this case around 12,000 pixels, roughly seven tenths of one per cent of the image.

Leave the inpaint radius at 10 and press Remove Object. The fill runs in the page and completes in a second or two for a mask of this size; larger masks take longer. Switch to the compare view to check the result against the original.

If the patch looks slightly blurred where grass texture should continue, undo, raise the radius towards 15 to draw on a wider neighbourhood, and try again. If instead the patch has pulled in colour from a nearby path, the mask probably extends too close to that edge, so repaint it slightly tighter.

This is a reconstruction technique, not a generative one. It has no knowledge of what should be behind the object; it only continues what surrounds the gap. That makes it excellent on uniform or gently varying backgrounds such as sky, water, a plain wall, sand or grass, and on small objects like a stray cable, a blemish, a date stamp or a distant figure.

It struggles in three situations. Strongly textured backgrounds, such as foliage, gravel or brickwork, come back smoother than their surroundings, because an average of neighbouring pixels cannot invent detail. Straight lines and structural edges crossing the masked region, like a window frame or a horizon meeting a roofline, tend to bend or break. And large regions are hardest of all: the further the front travels from the boundary, the less information there is, so the middle of a big patch drifts towards a soft blur.

Practical tips: remove several small objects in separate passes rather than one enormous mask; paint just beyond the object edge and include its shadow and any reflection; and if raising the radius only makes the patch softer, the region is probably too large or too textured for this approach.