GifImagePlugin normalize_palette destroys bounding box of frame

[sircinnamon]

What did you do?

Ran single colour frames through GifImagePlugin _normalize_palette

What did you expect to happen?

Image object palette changes, nothing else

What actually happened?

Image.getbbox() went from 100x100 to None

What are your OS, Python and Pillow versions?

• OS: Archlinux
• Python: 3.4.9 (in venv)
• Pillow: 5.4.1

To reproduce:
(This is a modification of the test_save_dispose test from the test suite - I think it's preventing a valid fix for something else but unsure)

from PIL import Image
out = "realtest.gif"
im_list = [
	Image.new('L', (100, 100), 255),
	Image.new('L', (100, 100), 127),
	Image.new('L', (100, 100), 0),
]
for method in range(0, 4):
	print("========= {} ========".format(method))
	im_list[0].save(
		str(method)+out,
		save_all=True,
		append_images=im_list[1:],
		disposal=method
	)
	img = Image.open(str(method)+out)
	for _ in range(2):
		img.seek(img.tell() + 1)
		print("{} == {}".format(img.disposal_method, method))
# check per frame disposal
im_list[0].save(
	out,
	save_all=True,
	append_images=im_list[1:],
	disposal=tuple(range(len(im_list)))
	)

img = Image.open(out)

for i in range(2):
	img.seek(img.tell() + 1)
	print("{} == {}".format(img.disposal_method, i+1))

Also, modify _write_multiple_frames in GifImagePlugin.py to contain a log of the image bounding box before and after the '_normalize_palette` call (around line 407):

for imSequence in itertools.chain([im],
                                      im.encoderinfo.get("append_images", [])):
        for im_frame in ImageSequence.Iterator(imSequence):
            # a copy is required here since seek can still mutate the image
            print("NEWFRAME")
            print(im_frame.getbbox())
            im_frame = _normalize_mode(im_frame.copy())
            print("MODE")
            print(im_frame.getbbox())
            im_frame = _normalize_palette(im_frame, palette, im.encoderinfo)
            print("PALETTE")
            print(im_frame.getbbox())

            encoderinfo = im.encoderinfo.copy()
            if isinstance(duration, (list, tuple)):
                encoderinfo['duration'] = duration[frame_count]
            if isinstance(disposal, (list, tuple)):
                encoderinfo["disposal"] = disposal[frame_count]
            frame_count += 1

Each time I see:

NEWFRAME
(0, 0, 100, 100)
MODE
(0, 0, 100, 100)
PALETTE
None

Is it possible that the Image.remap_palette being done by _normalize_palette is putting the only colour as colour 0, which causes the Image.getbbox to see the whole image as colour 0 and therefore not at all filled?

The frame having no bounding box results in the gif optimisation essentially ignoring the frame and extending the previous one.

I ran into this while trying to work on #3665

[radarhere]

This is an unusual issue, because it's not really a discussion about Pillow's API, or an issue with an example of a broken handling of an image, but about Pillow's internal operations.

I don't think it matters that im_frame.getbbox() is empty - delta.getbbox() is what is used to determine whether or not this frame is different to the previous one.

[radarhere]

Yes though, you are correct in the behaviour of getbbox() -

Pillow/src/PIL/Image.py

Lines 1233 to 1236 in e2ac1d1

	def getbbox(self):
	"""
	Calculates the bounding box of the non-zero regions in the
	image.

And I agree that calling remap_palette() on a single color image with the single item list of that used color does change the color to zero.

Pillow/src/PIL/GifImagePlugin.py

Line 400 in e2ac1d1

return im.remap_palette(used_palette_colors, source_palette)

from PIL import Image
zeros = 0
nonzeros = 0
for c in range(0, 256):
    if Image.new("P", (10, 10), 0).remap_palette([0]).load()[0, 0] == 0:
        zeros += 1
    else:
        nonzeros +=1
print(zeros, nonzeros)

gives

256 0

Comment if you have any further questions.

Also for the record, #3665 has been resolved.