Pixel comparison

Tools

Pixel comparison

The AMS sample paper testmath.tex, typeset twice from one run of LuaTeX: as a pageless PDF, and in the browser by ReflowTEX, with its column exactly as wide as the PDF’s. Both are drawn at 4 pixels per point, and the two pictures are matched line by line. Where each glyph is, the render tests compare without pixels; Accuracy describes both.

The browser is photographed with its text in TEX’s black rather than the site’s dark grey, and without the smoothing macOS adds to thicken letters, so that it draws glyphs the way a PDF renderer does. Then LuaTeX typeset the paper again with its column 85pt wider, and the same browser page, built from the 345pt run, was set at 430pt: there every paragraph is broken in the browser for a width the page was never compiled at.

In the pictures below, white means the two agree and red marks pixels that differ, the more they differ the darker, on a scale that keeps a pixel partly off paler than one wholly off, so a glyph that is not where it should be stands out in full red. The pink along edges is rasterisation: the PDF renderer and the browser smooth the edges of glyphs differently, and the PDF renderer puts each glyph on a whole pixel row. The bar above the picture has a cell for every 500pt of the paper, red where rows are out of place rather than just drawn differently. Switch between PDF and Browser to see the two at the same place.

Drawn finer shows the same comparison with both the PDF and the browser drawn at 16 pixels per point and averaged down to 4, the same way on both sides. That takes out the PDF renderer’s snapping of each glyph to a pixel row, and fewer pixels differ strongly. But at that scale the browser draws its glyphs a little differently again, so more rows differ slightly; the table has both.