fixed the bug concerning png-images with transparency palette

[d-schmidt]

#95 (comment)

[olt]

There is a + instead of *? This also removes the ability to store with limited palette size.

[d-schmidt]

you mean something like im.encoderinfo["transparency"][:2**bits]?
How would that work with a real image?

[olt]

For the bytes/string palette? Yes, but I would keep the alpha_bytes variable to make the code more readable.
For example, you need only 16 colors if you store a paletted image with 4 bits.
I would also swap the if statements and check for int instead of bytes (because transparency might be a string on Python 2).

PS: b'\xff' * 4 is b'\xff\xff\xff\xff'

[d-schmidt]

ah got it, reading http://www.libpng.org/pub/png/spec/1.2/PNG-Chunks.html#C.Anc-trns

[wiredfool]

I agree with @olt. Looks like encoderinfo['transparency'] can be an int representing # of bytes, the output type of ImageFile._safe_read (bytes or string), or an int16, or a 3-tuple of int16. Line 557 should be a * instead of +.

If we're round tripping the transparency with no change, then it looks like line 554 should be:

   chunk(fp, b"tRNS", im.encoderinfo["transparency"])

Since we get a bytes/string when we don't find a \0 in the tRNS chunk. To roundtrip, we'd have to emit it as we got it.

(note that I have not checked the spec to see how well this is reflected in the proper format of a png image)

[olt]

Ah, wait, now I see it. It doesn't work when multiple colors at the end are fully transparent.
^\xff+\x00+$ should be ^\xff+\x00\xff*$. Good catch.

[d-schmidt]

Couldn't we discard it all together and always use all given bytes?

[olt]

On decoding? Yes, probably. But I would keep support for encoding for backwards compatibility (if a user sets transparency).

[wiredfool]

This was probably originally added to make PNG transparency look like GIF transparency, which has just the one fully transparent color. I'd expect that it's expected behaviour from the users. I'd say keep it, but fix the regex as suggested by @olt.

[d-schmidt]

I've added the fixed pattern to the current open pull.

[d-schmidt]

contains new error, will send another pull