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 the RGB/YUV conversion nodes on images with an alpha channel (#15632)

Christian Byrne07dd46dc4abd
diff --git a/comfy_extras/nodes_morphology.py b/comfy_extras/nodes_morphology.py
index 0142040dd..cbdc4d013 100644
--- a/comfy_extras/nodes_morphology.py
+++ b/comfy_extras/nodes_morphology.py
@@ -73,6 +73,7 @@ class ImageRGBToYUV(io.ComfyNode):
@classmethod
def execute(cls, image) -> io.NodeOutput:
+ image = image[..., :3]
out = kornia.color.rgb_to_ycbcr(image.movedim(-1, 1)).movedim(1, -1)
return io.NodeOutput(out[..., 0:1].expand_as(image), out[..., 1:2].expand_as(image), out[..., 2:3].expand_as(image))
@@ -96,7 +97,7 @@ class ImageYUVToRGB(io.ComfyNode):
@classmethod
def execute(cls, Y, U, V) -> io.NodeOutput:
- image = torch.cat([torch.mean(Y, dim=-1, keepdim=True), torch.mean(U, dim=-1, keepdim=True), torch.mean(V, dim=-1, keepdim=True)], dim=-1)
+ image = torch.cat([torch.mean(Y[..., :3], dim=-1, keepdim=True), torch.mean(U[..., :3], dim=-1, keepdim=True), torch.mean(V[..., :3], dim=-1, keepdim=True)], dim=-1)
out = kornia.color.ycbcr_to_rgb(image.movedim(-1, 1)).movedim(1, -1)
return io.NodeOutput(out)
diff --git a/tests-unit/comfy_extras_test/rgb_yuv_alpha_test.py b/tests-unit/comfy_extras_test/rgb_yuv_alpha_test.py
new file mode 100644
index 000000000..fe72ab94d
--- /dev/null
+++ b/tests-unit/comfy_extras_test/rgb_yuv_alpha_test.py
@@ -0,0 +1,81 @@
+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_morphology import ImageRGBToYUV, ImageYUVToRGB # noqa: E402
+
+
+def image(channels, alpha=0.8, size=4):
+ torch.manual_seed(0)
+ t = torch.rand(1, size, size, channels)
+ if channels == 4:
+ t[..., 3] = alpha
+ return t
+
+
+def test_rgb_converts():
+ y, u, v = ImageRGBToYUV.execute(image(3)).result
+
+ assert y.shape[-1] == 3 and u.shape[-1] == 3 and v.shape[-1] == 3
+
+
+def test_rgba_does_not_raise():
+ y, u, v = ImageRGBToYUV.execute(image(4)).result
+
+ assert y.shape[-1] == 3 and u.shape[-1] == 3 and v.shape[-1] == 3
+
+
+def test_rgba_matches_rgb_conversion():
+ """Alpha must not leak into the luma/chroma values."""
+ rgba = image(4)
+ rgb = rgba[..., :3].clone()
+
+ from_rgba = ImageRGBToYUV.execute(rgba).result
+ from_rgb = ImageRGBToYUV.execute(rgb).result
+
+ for a, b in zip(from_rgba, from_rgb):
+ assert torch.equal(a, b)
+
+
+def test_alpha_value_does_not_change_the_conversion():
+ opaque = image(4, alpha=1.0)
+ transparent = image(4, alpha=0.0)
+
+ for a, b in zip(ImageRGBToYUV.execute(opaque).result, ImageRGBToYUV.execute(transparent).result):
+ assert torch.equal(a, b)
+
+
+def test_round_trip_recovers_rgb():
+ rgb = image(3)
+
+ y, u, v = ImageRGBToYUV.execute(rgb).result
+ out = ImageYUVToRGB.execute(y, u, v).result[0]
+
+ # float32 through the YCbCr matrix loses ~2e-4 either way; that is pre-existing.
+ assert torch.allclose(out, rgb, atol=1e-3)
+
+
+def test_yuv_to_rgb_ignores_alpha_on_its_inputs():
+ """A 4 channel Y/U/V input must not drag alpha into the channel mean."""
+ rgb = image(3)
+ y, u, v = ImageRGBToYUV.execute(rgb).result
+
+ def with_alpha(t):
+ return torch.cat((t, torch.full_like(t[..., :1], 0.3)), dim=-1)
+
+ expected = ImageYUVToRGB.execute(y, u, v).result[0]
+ out = ImageYUVToRGB.execute(with_alpha(y), with_alpha(u), with_alpha(v)).result[0]
+
+ assert torch.equal(out, expected)
+
+
+def test_does_not_mutate_input():
+ src = image(4)
+ before = src.clone()
+
+ ImageRGBToYUV.execute(src)
+
+ assert torch.equal(src, before)