PNG: a very small percentage of images decoded incorrectly

[Shnatsel]

The results from a test run on the manebooru.art database are in.

There are no decode failures other than ERROR [zune_bin] Could not complete workflow, reason png: Error decoding idat chunks Output limit exceeded, set limit was 1073741824 and output size is 107374185 which appears to be legitimate.

However, there are some discrepancies in decoding compared to imagemagick on a total of 70 images out of 600,000. This should hopefully be a representative sample: https://mega.nz/file/tslC1azb#CBg56O8T7uwC1uXBnbcQLjSaYiQALtaCz_b97hXDm3Y

Tested on commit 15e5d6c. The run on 600,000 images was done on an earlier commit, but I've re-tested everything on the specified commit, which is the latest as of this writing.

[etemesi254]

You have images with 1GB output size for dimensions?

That's a 16384 * 16384 RGBA image to say the least.

I should allow such settings to be tweaked in the command line though

[Shnatsel]

You have images with 1GB output size for dimensions?

One out of 600,000 images, yes.

[etemesi254]

A summary is that.

There was a bug on tRNS expansion, it's now corrected.

Most of these images are APNG (ffprobe will tell you imagemagick's identify won't), I am yet to add support for that.

You will get a warning on encountering APNG files informing you that you will only get the first frame (another cool thing was that we gained --probe abilities in the command line).

I don't think there are any further errors on PNG but I am not satisfied with some things i.e bit depth expansion, and some allocations we can avoid. I'm going to spend some time fixing those ones and will ping when I'm satisfied with speed and memory use.