Animated, transparent GIFs not calculating diff correctly

[Artheau]

function _write_multiple_frames in GifImagePlugin.py has an error on line 445 (line 445 in present repo, line 443 in released version 5.4.1).
Specifically, when the disposal_method is set to '2', it should calculate the diff against the background color, but currently it is calculating the diff against the previous frame
previous = im_frames[-1],
and this produces substantial artifacts in the rendering whenever there are still portions of the image in adjacent frames.

[dhsdshdhk]

Can you make a pull request to fix that?

[sircinnamon]

--- a/src/PIL/GifImagePlugin.py
+++ b/src/PIL/GifImagePlugin.py
@@ -443,7 +443,10 @@ def _write_multiple_frames(im, fp, palette):
             if im_frames:
                 # delta frame
                 previous = im_frames[-1]
-                if _get_palette_bytes(im_frame) == \
+                if "disposal" in encoderinfo and encoderinfo["disposal"] == 2:
+                    # If diposing whole frame, treat full new frame as delta
+                    delta = im_frame
+                elif _get_palette_bytes(im_frame) == \
                    _get_palette_bytes(previous['im']):
                     delta = ImageChops.subtract_modulo(im_frame,
                                                        previous['im'])

Tried this out in a fork, but the travisCI build failed on the test_save_dispose test, it seems like it's causing it to drop a frame somewhere, although I honestly cannot tell why. Open to ideas.

[Artheau]

I haven't dev'd on this project before, so please save me from myself, but I think the following snippet works. I didn't fork and pull request because the code needs a little love. It, for instance, creates a new background image on every iteration of the loop. That should probably be moved out of the loop, but then it would not have access to im_frame.size. Also, the new image should probably be created with the same palette, but again, if it is created outside of the loop then it is not clear how this would be accomplished.

Importantly, this code does not just use the whole frame as the delta. It does continue to try to optimize the frames as much as possible by comparing against the background color.

In any case, I think this code works, but please help me optimize the number of calls to Image.new:

            if im_frames:
                # delta frame
                previous = im_frames[-1]
                if ("disposal" in im.encoderinfo and im.encoderinfo["disposal"] == 2):
                    base_image = Image.new("P", im_frame.size, 0)
                else:
                    base_image = previous["im"]

                if _get_palette_bytes(im_frame) == \
                   _get_palette_bytes(base_image):
                    delta = ImageChops.subtract_modulo(im_frame,
                                                       base_image)
                else:
                    delta = ImageChops.subtract_modulo(
                        im_frame.convert('RGB'), base_image.convert('RGB'))
                bbox = delta.getbbox()
                if not bbox and not ("disposal" in im.encoderinfo and im.encoderinfo["disposal"] == 2):
                    # This frame is identical to the previous frame
                    if duration:
                        previous['encoderinfo']['duration'] += \
                            encoderinfo['duration']
                    continue

[radarhere]

Is there sample code that demonstrates this problem?

[burnoo]

@radarhere

import requests
from PIL import Image
response = requests.get('https://66.media.tumblr.com/tumblr_lzkp40b0Sl1qhwcy0.gif')
im = Image.open(BytesIO(response.content))
im.save('1.png', 'PNG')
im.seek(1)
im.load()
im.save('2.png', 'PNG')

The 2.png has second frame pasted onto first.
Expected result is second frame pasted into background (transparent).

[sircinnamon]

@Artheau With further testing, this introduces a new problem - each frame generates a diff compared to a colour 0 image - but each frame calculates its own palette and so colour 0 is not the same (nor are and colours necessarily.

I've tested and it seems like if every colour that appears in any frame appears in the first frame, it all works great. Trying to work out a solution where a palette that includes every colour from every frame can be generated and used for every frame.