Every PNG optimiser filters by default. On seven of eight animations that made the file bigger.

The mechanism is a disagreement between two cost functions. Minimum sum of absolute differences scores byte magnitude, DEFLATE pays for repetition, and on these frames the two point in opposite directions ^[14]. A run of identical pixels left at filter None is a run of identical bytes, which is the single thing LZ77 is best at; under Sub the same run becomes zeros, which also compresses, but the value at each boundary depends on whatever came before it, so an image made of flat regions with hard edges ends up with many small varying values instead of a few long identical runs ^[15]. The heuristic prefers the small varying values because that is the quantity it is scoring ^[14].

Its revealed preference is stark. Across the set it left between 0.0% and 0.2% of rows at filter None, on files where holding every row at None was 2.4x smaller ^[16]. Two rows in a thousand is not a rule that needs recalibrating; it is a rule that essentially never selects the winner, because the winner almost never has the lowest sum of absolute differences ^[16].

The route to that conclusion, as described in the dev.to write-up, is the transferable part. The first explanation, written into a source comment before anything was measured, was that flat art has repeating scanlines and that mixing filters per row rewrites identical rows into different bytes and destroys the long matches ^[7]. The worst file in the set has no repeated scanlines at all, and ranked across all 28 pairs of files, more repeated rows predicted a worse adaptive result in 6 of them, where a rule carrying no information lands near 14 ^[8]. That means the ordering ran the other way in 22 of the 28 pairs ^[9], so the data was mildly against the story rather than merely silent on it. Palette size managed 10 of 28, with the best and worst files sitting at 190 and 183 colours ^[10]. How often the heuristic switched filter type between adjacent rows landed on 14 of 28, exactly chance ^[11].

One detail deserves more attention than it usually gets. The forced-filter table came out wrong the first time: all five filters reported 1.43x, identical to the decimal ^[20]. Five different transforms agreeing that precisely is a measurement of the harness rather than of the images, and it is the cheapest bug signal available in work like this.

The probe runs once per animation rather than once per frame, on the reasoning that the frames of one animation are the same kind of picture, so re-confirming the winner on each of them doubles the compression work to learn nothing ^[18]. It also preserves the gain that is real: the single continuous-tone file still comes out at 0.83x because the probe finds it ^[19], and on that file every differencing filter wins ^[13]. The textbook advice is right about a kind of image, and seven of these eight were the other kind ^[13]. What eight files can establish is that the losing kind is common enough to dominate a small sample of real animations, not what share of animation payloads it represents.