def _quantize(t: Image, palette) -> Image:
with Image.new('P', (1, 1)) as palette_img:
p = [x for sub in palette for x in sub] + [0] * (768 - 3 * len(palette))
palette_img.putpalette(p)
palette_img.load()
im = t.im.convert('P', 0, palette_img.im)
return t._new(im).convert('RGB')
def _quantize(t: Image, palette) -> Image:
with Image.new('P', (1, 1)) as palette_img:
# Flatten 2d array to 1d, then pad with first color to 786 total values
p = [v for color in palette
for v in color] + list(palette[0]) * (256 - len(palette))
palette_img.putpalette(p)
palette_img.load()
im = t.im.convert(
'P', 0, palette_img.im
) # Quantize using internal PIL shit so it's not dithered
return t._new(im).convert('RGB')
def recolor(img: Image, alpha: bool, **options):
# Alpha channel take 1 color
# QUANTIZE with kmeans != 0
# FOR ALPHA IS MANDATORY 2 (3depends on a lib)
colors_count = options["colors"][
"colors"] if not alpha else options["colors"]["colors"] - 1
result = img.quantize(colors=colors_count, method=2, kmeans=0)
if options["dither"] and options["colors"]["mode"] is not "LA":
# utilizing underlying imaging core C library directly
# maybe implementing custom algorithm would be better
# also reorganizing the chain of operations for recoloring
# (by better understanding different parts at work and properly chaining them)
dithered = img._new(
img.im.convert(options["colors"]["mode"], Image.FLOYDSTEINBERG,
result.im))
dithered = dithered.quantize(colors=colors_count)
return dithered
result = result.convert(mode=options["colors"]["mode"],
palette=Image.ADAPTIVE,
colors=colors_count)
return result
