oxipng corrupts certain images

[CounterPillow]

I used the following script to run oxipng on a large number of PNG files from my mpv screenshot directory (~650 files), comparing the output of oxipng with the original file using graphicsmagick's compare command. The oxipng version used was 3f42eae.

#!/usr/bin/env bash
OXI=~/Projekte/oxipng/target/release/oxipng
NUM=0
if [ ! -d "/tmp/oxifaults" ]; then
    mkdir "/tmp/oxifaults"
fi
while IFS= read -rd '' f; do
    rm -f "/tmp/oxitest.png"
    $OXI -F --no-clobber --out "/tmp/oxitest.png" "$f" >> /dev/null 2>&1
    gm compare -maximum-error 0 -metric MSE "$f" "/tmp/oxitest.png" >> /dev/null 2>&1
    if [ $? -ne 0 ]; then
        cp "/tmp/oxitest.png" "/tmp/oxifaults/$NUM-oxi.png"
        echo "$f -> $NUM-oxi.png"
        ((NUM++))
    fi
done < <(find "$1" -type f -iname '*.png' -print0)

A number of images where the output did not correspond to the input were detected.

1. Original -> Oxipng
2. Original -> Oxipng
3. Original -> Oxipng
4. Original -> Oxipng
5. Original -> Oxipng
6. Original -> Oxipng
7. Original -> Oxipng
8. Original -> Oxipng
9. Original -> Oxipng
10. Original -> Oxipng

(For reasons beyond my comprehension, KonoSuba screenshots seemed to be more likely to get corrupted than others; I'm going to assume oxipng has bad taste in anime.)

As one can see, subtitles, for some reason, often trigger this.

rustc version is rustc 1.9.0-nightly (998a6720b 2016-03-07).

[CounterPillow]

As @ion1 suggested, I attempted to flip the image so the subtitles are on top, and as it turns out this affects whether the images are corrupted or not.

gm convert 9.png -flip 9_f.png
oxipng -F 9_f.png

Note that no corruption occurs.

EDIT: Cropping from top or bottom also changes the outcome.

[shssoichiro]

Can you test again against f1e6506? I modified much of the filtering code and can't reproduce the issue anymore (I was able to reproduce on 3f42eae).

[CounterPillow]

The images which resulted in corruption before no longer appear to get corrupted.

[shssoichiro]

Great. I'll add a test to make sure the behavior doesn't return, then I'll count this as closed.