Fill notes

An eraser that produces a good result and an eraser that produces a bad one are the same program running the same code on different pixels. Which you get is decided almost entirely before you press anything — by what is underneath the thing you painted over, and by how far past its edge you painted. Somebody who knows that spends thirty seconds adjusting a region and gets a removal nobody notices. Somebody who does not runs the fill four times, gets four smears, and concludes the tool is weak.

So these are not tutorials and there is no numbered walkthrough in any of them; the working pages carry that. They are explanations of the engines and of the shapes of picture each one cannot handle, written so that you can look at a photograph and predict the outcome before committing to it. Where an honest answer is that a removal will not hold up, they say so and name what to do instead.

• The mechanismHow the fill decides what goes thereWhat a patch search is doing when it votes on a pixel, what a learned engine does instead, and why the two fail in opposite directions.
• A specific jobDate stamps on scanned printsWhy an imprint from a film compact is orange, why the glow around each digit outlives a tight region, and when picking by colour beats painting by hand.
• The limitWhy straight lines breakThe difference between a surface whose details are interchangeable and one whose details are positioned, and four ways to work around the second.
• The boundaryRemoving a watermark — and whether you shouldSeparate legitimate cleanup from ownership marks, and what a fill can and cannot rebuild under a semi-transparent mark.
• The other jobErasing objects versus cutting out a subjectWhen you want the frame kept and a region filled, and when you actually want the surroundings made transparent.
• RedactionCover versus fillWhy hiding a name or a number has to look like hiding it — an opaque cover, never a plausible fill — and how to check one.
• Before you erasePhotos you should not editPassport, visa and exam photographs and identity documents, and what to do instead of retouching them.

Start with the mechanism piece if the question in your head is why a result looked smeared, because the answer is nearly always in how the search runs out of candidates. Start with the straight-lines piece if the question is why one region came out perfectly and the one beside it did not. The date-stamp piece is narrower than the other two and is worth reading only if you have a box of scanned prints with orange digits in the corner — in which case it is the only one of the three that matters.

• There is no separate note on the learned engine. What it is, what it costs to download, what its fixed 512-pixel input does to a large region and where it still fails are in the mechanism piece, alongside the two engines it is being compared with — which is the only place a comparison means anything.
• There is no note on cutting a subject out, replacing a background, retouching a face or straightening a horizon. None of those operations exist here and none is planned, so a guide to them would be a page about somewhere else.
• There is no note on batch work. A region is drawn per photograph and that is a design decision rather than a missing feature, which makes it a paragraph on the about page rather than an essay.