added asymmetric x/y tiling to model SeamlessTile

SeamlessTile.py

@@ -2,6 +2,11 @@ import copy
 
 import PIL
 import torch
+from torch import Tensor
+from torch.nn import Conv2d
+from torch.nn import functional as F
+from torch.nn.modules.utils import _pair
+from typing import Optional
 
 
 class SeamlessTile:
@@ -10,8 +15,8 @@ class SeamlessTile:
         return {
             "required": {
                 "model": ("MODEL",),
-                "tiling": (["enable", "disable"],),
-                "copy_model": (["Modify in place", "Make a copy"],),
+                "tiling": (["enable", "x_only", "y_only", "disable"],),
+                "copy_model": (["Make a copy", "Modify in place"],),
             },
         }
 
@@ -25,10 +30,15 @@ class SeamlessTile:
             model_copy = model
         else:
             model_copy = copy.deepcopy(model)
+            
         if tiling == "enable":
-            model_copy.model.apply(make_circular)
+            make_circular_asymm(model_copy.model, True, True)
+        elif tiling == "x_only":
+            make_circular_asymm(model_copy.model, True, False)
+        elif tiling == "y_only":
+            make_circular_asymm(model_copy.model, False, True)
         else:
-            model_copy.model.apply(unmake_circular)
+            make_circular_asymm(model_copy.model, False, False)
         return (model_copy,)
 
 
@@ -37,11 +47,39 @@ def make_circular(m):
         m.padding_mode = "circular"
 
 
+# asymmetric tiling from https://github.com/tjm35/asymmetric-tiling-sd-webui/blob/main/scripts/asymmetric_tiling.py
+def make_circular_asymm(model, tileX: bool, tileY: bool):
+    for layer in [
+        layer for layer in model.modules() if isinstance(layer, torch.nn.Conv2d)
+    ]:
+        layer.padding_modeX = 'circular' if tileX else 'constant'
+        layer.padding_modeY = 'circular' if tileY else 'constant'
+        layer.paddingX = (layer._reversed_padding_repeated_twice[0], layer._reversed_padding_repeated_twice[1], 0, 0)
+        layer.paddingY = (0, 0, layer._reversed_padding_repeated_twice[2], layer._reversed_padding_repeated_twice[3])
+        layer._conv_forward = __replacementConv2DConvForward.__get__(layer, Conv2d)
+    return model
+
+
+def __replacementConv2DConvForward(self, input: Tensor, weight: Tensor, bias: Optional[Tensor]):
+    working = F.pad(input, self.paddingX, mode=self.padding_modeX)
+    working = F.pad(working, self.paddingY, mode=self.padding_modeY)
+    return F.conv2d(working, weight, bias, self.stride, _pair(0), self.dilation, self.groups)
+
+
 def unmake_circular(m):
     if isinstance(m, torch.nn.Conv2d):
         m.padding_mode = "zeros"
 
 
+def unmake_circular_asymm(model):
+    for layer in [
+        layer for layer in model.modules() if isinstance(layer, torch.nn.Conv2d)
+    ]:
+        layer.padding_mode = "zeros"
+        layer._conv_forward = Conv2d._conv_forward.__get__(layer, Conv2d)
+    return model
+
+
 class CircularVAEDecode:
     @classmethod
     def INPUT_TYPES(s):