@@ -71,6 +71,11 @@ def define_schema(cls):
     def execute(cls, upscale_model, image) -> io.NodeOutput:
         device = upscale_model.patcher.load_device
 
+        alpha = None
+        if image.shape[-1] == 4:
+            alpha = image[..., 3:4]
+            image = image[..., :3]
+
         memory_required = (512 * 512 * 3) * image.element_size() * max(upscale_model.scale, 1.0) * 384.0 #The 384.0 is an estimate of how much some of these models take, TODO: make it more accurate
         memory_required += image.nelement() * image.element_size()
         model_management.load_models_gpu([upscale_model.patcher], memory_required=memory_required, force_full_load=True)
@@ -96,6 +101,11 @@ def execute(cls, upscale_model, image) -> io.NodeOutput:
                     raise e
 
         s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0).to(comfy.model_management.intermediate_dtype())
+
+        if alpha is not None:
+            alpha = comfy.utils.common_upscale(alpha.movedim(-1, -3).to(s), s.shape[2], s.shape[1], "bilinear", "disabled")
+            s = torch.cat((s, alpha.movedim(-3, -1)), dim=-1)
+
         return io.NodeOutput(s)
 
     upscale = execute  # TODO: remove
 
@@ -0,0 +1,58 @@
+import torch
+
+from comfy.cli_args import args as cli_args
+
+if not torch.cuda.is_available():
+    cli_args.cpu = True
+
+import comfy.model_management  # noqa: E402
+from comfy_extras.nodes_upscale_model import ImageUpscaleWithModel  # noqa: E402
+
+
+class StubUpscaleModel:
+    """Stands in for a spandrel ImageModelDescriptor that only accepts RGB."""
+
+    scale = 2
+
+    class patcher:
+        load_device = torch.device("cpu")
+
+    def __init__(self):
+        self.seen_channels = None
+
+    def __call__(self, image):
+        self.seen_channels = image.shape[1]
+        assert image.shape[1] == 3, "model was handed a non-RGB tensor"
+        return torch.nn.functional.interpolate(image, scale_factor=self.scale, mode="nearest")
+
+
+def rgba_image():
+    """16x16 RGBA where the top half is transparent and the bottom half opaque."""
+    image = torch.zeros(1, 16, 16, 4)
+    image[..., :3] = 0.5
+    image[0, 8:, :, 3] = 1.0
+    return image
+
+
+def upscale(image, monkeypatch):
+    monkeypatch.setattr(comfy.model_management, "load_models_gpu", lambda *args, **kwargs: None)
+    model = StubUpscaleModel()
+    return model, ImageUpscaleWithModel.execute(model, image)
+
+
+def test_rgba_input_does_not_crash_and_alpha_is_preserved(monkeypatch):
+    model, out = upscale(rgba_image(), monkeypatch)
+    image = out.result[0]
+
+    assert model.seen_channels == 3
+    assert image.shape == (1, 32, 32, 4)
+    assert image[0, :14, :, 3].max() < 0.1
+    assert image[0, 18:, :, 3].min() > 0.9
+
+
+def test_rgb_input_is_untouched(monkeypatch):
+    model, out = upscale(torch.zeros(1, 16, 16, 3), monkeypatch)
+    image = out.result[0]
+
+    assert model.seen_channels == 3
+    assert image.shape == (1, 32, 32, 3)