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.

Don't blend the alpha channel in the Blend Images node (#15625)

Christian Byrnef9c706f31748
diff --git a/comfy_extras/nodes_post_processing.py b/comfy_extras/nodes_post_processing.py
index 9b66ced0f..27f51c44f 100644
--- a/comfy_extras/nodes_post_processing.py
+++ b/comfy_extras/nodes_post_processing.py
@@ -47,6 +47,8 @@ class Blend(io.ComfyNode):
blended_image = cls.blend_mode(image1, image2, blend_mode)
blended_image = image1 * (1 - blend_factor) + blended_image * blend_factor
blended_image = torch.clamp(blended_image, 0, 1)
+ if image1.shape[-1] == 4: # alpha stores transparency, not color
+ blended_image[..., 3] = image1[..., 3]
return io.NodeOutput(blended_image)
@classmethod
diff --git a/tests-unit/comfy_extras_test/image_blend_alpha_test.py b/tests-unit/comfy_extras_test/image_blend_alpha_test.py
new file mode 100644
index 000000000..6178f82a5
--- /dev/null
+++ b/tests-unit/comfy_extras_test/image_blend_alpha_test.py
@@ -0,0 +1,84 @@
+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 Blend # noqa: E402
+
+MODES = ["normal", "multiply", "screen", "overlay", "soft_light", "difference"]
+
+
+def image(value, alpha=None, size=4):
+ channels = 3 if alpha is None else 4
+ t = torch.full((1, size, size, channels), value)
+ if alpha is not None:
+ t[..., 3] = alpha
+ return t
+
+
[email protected]("mode", MODES)
+def test_rgb_blend_is_unchanged(mode):
+ """3 channel images must keep going through the untouched code path."""
+ image1, image2 = image(0.8), image(0.3)
+
+ out = Blend.execute(image1, image2, 0.5, mode).result[0]
+
+ assert out.shape == image1.shape
+ assert torch.all(out >= 0.0) and torch.all(out <= 1.0)
+
+
[email protected]("mode", MODES)
+def test_rgba_keeps_image1_alpha(mode):
+ image1, image2 = image(0.8, alpha=0.6), image(0.3, alpha=0.1)
+
+ out = Blend.execute(image1, image2, 0.5, mode).result[0]
+
+ assert out.shape[-1] == 4
+ assert torch.equal(out[..., 3], image1[..., 3])
+
+
[email protected]("mode", MODES)
+def test_two_opaque_images_stay_opaque(mode):
+ """Regression: difference mode used to compute 1.0 - 1.0 and erase the image."""
+ image1, image2 = image(0.8, alpha=1.0), image(0.3, alpha=1.0)
+
+ out = Blend.execute(image1, image2, 1.0, mode).result[0]
+
+ assert torch.all(out[..., 3] == 1.0)
+
+
[email protected]("mode", MODES)
+def test_rgb_channels_still_blend_on_rgba(mode):
+ """Preserving alpha must not stop the colour channels from blending.
+
+ Mid-tones on purpose: pure white over pure black is a fixed point for
+ several of the modes, so it would not prove anything.
+ """
+ image1, image2 = image(0.6, alpha=1.0), image(0.25, alpha=1.0)
+
+ out = Blend.execute(image1, image2, 1.0, mode).result[0]
+
+ assert not torch.equal(out[..., :3], image1[..., :3])
+
+
+def test_does_not_mutate_inputs():
+ image1, image2 = image(0.8, alpha=0.6), image(0.3, alpha=0.1)
+ before1, before2 = image1.clone(), image2.clone()
+
+ Blend.execute(image1, image2, 0.5, "difference")
+
+ assert torch.equal(image1, before1)
+ assert torch.equal(image2, before2)
+
+
+def test_mismatched_channel_counts_still_supported():
+ """image_alpha_fix (CORE-103) pads the RGB input; that must keep working."""
+ rgba, rgb = image(0.8, alpha=0.5), image(0.3)
+
+ out = Blend.execute(rgba, rgb, 0.5, "normal").result[0]
+
+ assert out.shape[-1] == 4
+ assert torch.equal(out[..., 3], rgba[..., 3])