Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix Quantize Image on images with an alpha channel (#15630)

Christian Byrneaa4582f93a2e
diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py
index 763b8a52f..9b66ced0f 100644
--- a/comfy_extras/nodes_post_processing.py
+++ b/comfy_extras/nodes_post_processing.py
@@ -156,11 +156,12 @@ class Quantize(io.ComfyNode):
@classmethod
def execute(cls, image: torch.Tensor, colors: int, dither: str) -> io.NodeOutput:
- batch_size, height, width, _ = image.shape
- result = torch.zeros_like(image)
+ rgb = image[..., :3]
+ batch_size, height, width, _ = rgb.shape
+ result = torch.zeros_like(rgb)
for b in range(batch_size):
- im = Image.fromarray((image[b] * 255).to(torch.uint8).numpy(), mode='RGB')
+ im = Image.fromarray((rgb[b] * 255).to(torch.uint8).numpy(), mode='RGB')
pal_im = im.quantize(colors=colors) # Required as described in https://github.com/python-pillow/Pillow/issues/5836
@@ -175,6 +176,8 @@ class Quantize(io.ComfyNode):
quantized_array = torch.tensor(np.array(quantized_image.convert("RGB"))).float() / 255
result[b] = quantized_array
+ if image.shape[-1] == 4:
+ result = torch.cat((result, image[..., 3:]), dim=-1)
return io.NodeOutput(result)
class Sharpen(io.ComfyNode):
diff --git a/tests-unit/comfy_extras_test/image_quantize_alpha_test.py b/tests-unit/comfy_extras_test/image_quantize_alpha_test.py
new file mode 100644
index 000000000..114bf808e
--- /dev/null
+++ b/tests-unit/comfy_extras_test/image_quantize_alpha_test.py
@@ -0,0 +1,65 @@
+import pytest
+import torch
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+ cli_args.cpu = True
+
+from comfy_extras.nodes_post_processing import Quantize # noqa: E402
+
+DITHERS = ["none", "floyd-steinberg", "bayer-2", "bayer-4", "bayer-8", "bayer-16"]
+
+
+def image(channels, alpha=0.8, size=8):
+ torch.manual_seed(0)
+ t = torch.rand(1, size, size, channels)
+ if channels == 4:
+ t[..., 3] = alpha
+ return t
+
+
[email protected]("dither", DITHERS)
+def test_rgb_still_quantizes(dither):
+ src = image(3)
+
+ out = Quantize.execute(src, 4, dither).result[0]
+
+ assert out.shape == src.shape
+ assert len(torch.unique(out.reshape(-1, 3), dim=0)) <= 4
+
+
[email protected]("dither", DITHERS)
+def test_rgba_does_not_raise_and_keeps_alpha(dither):
+ src = image(4)
+
+ out = Quantize.execute(src, 4, dither).result[0]
+
+ assert out.shape == src.shape
+ assert torch.equal(out[..., 3], src[..., 3])
+
+
+def test_rgba_colour_channels_are_quantized():
+ src = image(4)
+
+ out = Quantize.execute(src, 4, "none").result[0]
+
+ assert len(torch.unique(out[..., :3].reshape(-1, 3), dim=0)) <= 4
+
+
+def test_varying_alpha_is_preserved_per_pixel():
+ src = image(4)
+ src[0, :, :, 3] = torch.linspace(0.0, 1.0, src.shape[2])
+
+ out = Quantize.execute(src, 8, "none").result[0]
+
+ assert torch.equal(out[..., 3], src[..., 3])
+
+
+def test_does_not_mutate_input():
+ src = image(4)
+ before = src.clone()
+
+ Quantize.execute(src, 4, "none")
+
+ assert torch.equal(src, before)