Use center-last order for merging in dynamic nodes

README.md

@@ -1,18 +1,42 @@
 
 # SimpleTiles
-
+## DynamicTileSplit / DynamicTileMerge
-## TileSplit
+![](dynamic.png)
 Splits image into tiles. Overlap value decides how much overlap there is between tiles on y axis, x axis is calculated to have the same ratio to image height as y axis.
 
-## TileMerge
+Automatically splits image into tiles based on image size and tile size. Tiles can be different ratio than images.
+
+DynamicTileSplit outputs a `tile_calc` object. The object contains info about size and overlap and should be passed to DynamicTileMerge.
+
+**Overlap** value decides how much overlap there is between tiles on y axis, x axis is calculated to have the same ratio to image height as y axis. Should be set to same value as used in TileSplit.
+
+**Blend** value decides how many pixels the blending is done over. Should be less than overlap value. Blending is done linearly from 0 to 1 over the blend distance.
+
+
+## Legacy
+DynamicTileSplit and DynamicTileMerge are the new versions of TileSplit and TileMerge. They are more flexible and easier to use.
+
+Legacy nodes don't work well if image ratio and tile ratio is different. 
+
+Use TileCalc to calculate the final image size, pipe the final size to TileMerge and ImageScale.
+
+
+### TileSplit (Legacy)
+Splits image into tiles. Overlap value decides how much overlap there is between tiles on y axis, x axis is calculated to have the same ratio to image height as y axis.
+
+### TileMerge (Legacy)
 Merge tiles into image. 
 
 **Overlap** value decides how much overlap there is between tiles on y axis, x axis is calculated to have the same ratio to image height as y axis. Should be set to same value as used in TileSplit.
 
 **Blend** value decides how many pixels the blending is done over. Should be less than overlap value. Blending is done linearly from 0 to 1 over the blend distance.
 
-## TileCalc
+### TileCalc (Legacy)
 
 Util to calculate final image size based on tile sizes and overlaps. 
 
 
+
+## Example Ipadapter
+![](ipadapter.png)
+

__init__.py

@@ -1,3 +1,3 @@
 from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
 
-__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

dynamic.py

@@ -0,0 +1,203 @@
+import sys
+import os
+
+import torch
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy"))
+
+
+def order_by_center_last(tiles, image_width, image_height, tile_width, tile_height):
+    # for 3x3: custom_order = [0, 2, 6, 8, 1, 3, 5, 7, 4] # First 4 corners, then the sides, then the center
+    # order the tiles so they are add based on absolute distance from the center of the tile to the center of the image
+    # this is done so that the center of the image is the last tile to be added, so that the center of the image is the most refined
+
+    # get the center of the image
+    center_x = image_width // 2
+    center_y = image_height // 2
+
+    # sort the tiles by distance from the center
+    tiles = sorted(
+        tiles,
+        key=lambda tile: abs(tile[0] + tile_width // 2 - center_x)
+        + abs(tile[1] + tile_height // 2 - center_y),
+    )
+
+    # reverse the order so that the center is last
+    tiles = tiles[::-1]
+
+    return tiles
+
+
+def generate_tiles(
+    image_width, image_height, tile_width, tile_height, overlap, offset=0
+):
+    tiles = []
+
+    y = 0
+    while y < image_height:
+        if y == 0:
+            next_y = y + tile_height - overlap + offset
+        else:
+            next_y = y + tile_height - overlap
+
+        if y + tile_height >= image_height:
+            y = max(image_height - tile_height, 0)
+            next_y = image_height
+
+        x = 0
+        while x < image_width:
+            if x == 0:
+                next_x = x + tile_width - overlap + offset
+            else:
+                next_x = x + tile_width - overlap
+            if x + tile_width >= image_width:
+                x = max(image_width - tile_width, 0)
+                next_x = image_width
+
+            tiles.append((x, y))
+
+            if next_x > image_width:
+                break
+            x = next_x
+
+        if next_y > image_height:
+            break
+        y = next_y
+
+    return order_by_center_last(tiles, image_width, image_height, tile_width, tile_height)
+
+
+class DynamicTileSplit:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "tile_width": ("INT", {"default": 512, "min": 1, "max": 10000}),
+                "tile_height": ("INT", {"default": 512, "min": 1, "max": 10000}),
+                "overlap": ("INT", {"default": 128, "min": 1, "max": 10000}),
+                "offset": ("INT", {"default": 0, "min": 0, "max": 10000}),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE", "TILE_CALC")
+    FUNCTION = "process"
+    CATEGORY = "ipadapter"
+
+    def process(self, image, tile_width, tile_height, overlap, offset):
+        image_height = image.shape[1]
+        image_width = image.shape[2]
+
+        tile_coordinates = generate_tiles(
+            image_width, image_height, tile_width, tile_height, overlap, offset
+        )
+
+        print("Tile coordinates: {}".format(tile_coordinates))
+
+        iteration = 1
+
+        image_tiles = []
+        for tile_coordinate in tile_coordinates:
+            print("Processing tile {} of {}".format(iteration, len(tile_coordinates)))
+            print("Tile coordinate: {}".format(tile_coordinate))
+            iteration += 1
+
+            image_tile = image[
+                :,
+                tile_coordinate[1] : tile_coordinate[1] + tile_height,
+                tile_coordinate[0] : tile_coordinate[0] + tile_width,
+                :,
+            ]
+
+            image_tiles.append(image_tile)
+
+        tiles_tensor = torch.stack(image_tiles).squeeze(1)
+        tile_calc = (overlap, image_height, image_width, offset)
+
+        return (tiles_tensor, tile_calc)
+
+
+class DynamicTileMerge:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "blend": ("INT", {"default": 64, "min": 0, "max": 4096}),
+                "tile_calc": ("TILE_CALC",),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "process"
+    CATEGORY = "utils"
+
+    def process(self, images, blend, tile_calc):
+        overlap, final_height, final_width, offset = tile_calc
+        tile_height = images.shape[1]
+        tile_width = images.shape[2]
+        print("Tile height: {}".format(tile_height))
+        print("Tile width: {}".format(tile_width))
+        print("Final height: {}".format(final_height))
+        print("Final width: {}".format(final_width))
+        print("Overlap: {}".format(overlap))
+
+        tile_coordinates = generate_tiles(
+            final_width, final_height, tile_width, tile_height, overlap, offset
+        )
+
+        print("Tile coordinates: {}".format(tile_coordinates))
+        original_shape = (1, final_height, final_width, 3)
+        count = torch.zeros(original_shape, dtype=images.dtype)
+        output = torch.zeros(original_shape, dtype=images.dtype)
+
+        index = 0
+        iteration = 1
+        for tile_coordinate in tile_coordinates:
+            image_tile = images[index]
+            x = tile_coordinate[0]
+            y = tile_coordinate[1]
+
+            print("Processing tile {} of {}".format(iteration, len(tile_coordinates)))
+            print("Tile coordinate: {}".format(tile_coordinate))
+            iteration += 1
+
+            channels = images.shape[3]
+            weight_matrix = torch.ones((tile_height, tile_width, channels))
+
+            # blend border
+            for i in range(blend):
+                weight = float(i) / blend
+                weight_matrix[i, :, :] *= weight  # Top rows
+                weight_matrix[-(i + 1), :, :] *= weight  # Bottom rows
+                weight_matrix[:, i, :] *= weight  # Left columns
+                weight_matrix[:, -(i + 1), :] *= weight  # Right columns
+
+            # We only want to blend with already processed pixels, so we keep
+            # track if it has been processed.
+            old_tile = output[:, y : y + tile_height, x : x + tile_width, :]
+            old_tile_count = count[:, y : y + tile_height, x : x + tile_width, :]
+
+            weight_matrix = (
+                weight_matrix * (old_tile_count != 0).float()
+                + (old_tile_count == 0).float()
+            )
+
+            image_tile = image_tile * weight_matrix + old_tile * (1 - weight_matrix)
+
+            output[:, y : y + tile_height, x : x + tile_width, :] = image_tile
+            count[:, y : y + tile_height, x : x + tile_width, :] = 1
+
+            index += 1
+        return [output]
+
+
+NODE_CLASS_MAPPINGS = {
+    "DynamicTileSplit": DynamicTileSplit,
+    "DynamicTileMerge": DynamicTileMerge,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "DynamicTileSplit": "DynamicTileSplit",
+    "DynamicTileMerge": "DynamicTileMerge",
+}

nodes.py

@@ -1,179 +1,18 @@
-import sys
+from ComfyUI_SimpleTiles.standard import TileSplit, TileMerge, TileCalc
-import os
+from ComfyUI_SimpleTiles.dynamic import DynamicTileSplit, DynamicTileMerge
-
-import torch
-
-sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy"))
-
-
-IMAGE_SIZE = 1472
-TILE_SIZE = 4096
-OVERLAP = 64
-
-# Splits an image in four tiles and returns them as a list
-class TileSplit:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required":{
-            "image": ("IMAGE", ),
-            "tile_height": ("INT", {"default": 64, "min": 64, "max": 4096}),
-            "tile_width": ("INT", {"default": 64, "min": 64, "max": 4096}),
-            "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
-            }
-        }
-
-    RETURN_TYPES = ("IMAGE",)
-    FUNCTION = "split"
-    CATEGORY = "utils"
-
-    def split(self, image, tile_height, tile_width, overlap):
-
-        height, width = image.shape[1], image.shape[2]
-        overlap_x = overlap
-        overlap_y = int(overlap * (tile_height / tile_width))
-
-        tiles = []
-        for y in range(0, height-tile_height+1, tile_height-overlap_y):
-            for x in range(0, width-tile_width+1, tile_width-overlap_x):
-                tile = image[:, y:y+tile_height, x:x+tile_width, :]
-                tiles.append(tile)
-
-        # Convert tiles list to a tensor if needed
-        tiles_tensor = torch.stack(tiles).squeeze(1)
-
-
-
-        return  [tiles_tensor]
-
-class TileMerge:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required":{
-            "images": ("IMAGE", ),
-            "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
-            "blend": ("INT", {"default": 64, "min": 0, "max": 4096}),
-            "final_height": ("INT", {"default": 2048, "min": 0, "max": 9*4096}),
-            "final_width": ("INT", {"default": 2048, "min": 0, "max": 9*4096}),
-            }
-        }
-
-    RETURN_TYPES = ("IMAGE",)
-    FUNCTION = "blend_tiles"
-    CATEGORY = "utils"
-    def blend_tiles(self, images, overlap, blend, final_height, final_width):
-        tiles = images
-        tile_height, tile_width = images.shape[1], images.shape[2]
-        original_shape = (1, final_height, final_width, 3)
-        overlap_x = overlap
-        overlap_y = int(overlap * (tile_height / tile_width))
-
-        batch, height, width, channels = original_shape
-        output = torch.zeros(original_shape, dtype=tiles.dtype)
-        count = torch.zeros(original_shape, dtype=tiles.dtype)
-        idx = 0
-
-        # for 3x3: custom_order = [0, 2, 6, 8, 1, 3, 5, 7, 4] # First 4 corners, then the sides, then the center
-
-        # Calculate grid dimensions
-        rows = (height - tile_height) // (tile_height - overlap_y) + 1
-        cols = (width - tile_width) // (tile_width - overlap_x) + 1
-
-        # Calculate the center of the grid
-        center_row, center_col = rows // 2, cols // 2
-
-        print("Rows: {}, Cols: {}".format(rows, cols))
-        print("Center row: {}, Center col: {}".format(center_row, center_col))
-
-        # Calculate the order in which to blend the tiles
-        # Order based on distance from center
-        distances = []
-        for i in range(rows):
-            for j in range(cols):
-                distance = abs(i - center_row) + abs(j - center_col)
-                distances.append(distance)
-
-        # Sort the tiles based on distance from center
-        reverse_custom_order = sorted(range(len(distances)), key=lambda k: distances[k])
-        custom_order = reverse_custom_order[::-1]
-
-        print("Custom order: {}".format(custom_order))
-
-        
-
-
-        ys = [y for y in range(0, height-tile_height+1, tile_height-overlap_y)]
-        xs = [x for x in range(0, width-tile_width+1, tile_width-overlap_x)]
-        for idx in custom_order:
-
-            y = ys[idx // len(ys)]
-            x = xs[idx % len(xs)]
-
-            tile = tiles[idx]
-
-            weight_matrix = torch.ones((tile_height, tile_width, channels))
-            # if not center tile
-            for i in range(blend):
-                weight = float(i) / blend
-                weight_matrix[i, :, :] *= weight  # Top rows
-                weight_matrix[-(i + 1), :, :] *= weight  # Bottom rows
-                weight_matrix[:, i, :] *= weight  # Left columns
-                weight_matrix[:, -(i + 1), :] *= weight  # Right columns
-
-            old_tile = output[:, y:y+tile_height, x:x+tile_width, :]
-            old_tile_count = count[:, y:y+tile_height, x:x+tile_width, :]
-
-            weight_matrix = weight_matrix * (old_tile_count != 0).float() + (old_tile_count == 0).float()
-
-            # Blend the old tile with the new tile
-            tile = tile * weight_matrix + old_tile * (1 - weight_matrix)
-
-            output[:, y:y+tile_height, x:x+tile_width, :] = tile
-            count[:, y:y+tile_height, x:x+tile_width, :] = 1
-
-        # Normalize the output and return
-        #output /= count
-
-
-
-        return [output]
-
-
-class TileCalc:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required":{
-            "tile_height": ("INT", {"default": 64, "min": 64, "max": 4096}),
-            "tile_width": ("INT", {"default": 64, "min": 64, "max": 4096}),
-            "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
-            "tile_width_n": ("INT", {"default": 3, "min": 1, "max": 9}),
-            "tile_height_n": ("INT", {"default": 3, "min": 1, "max": 9}),
-            }
-        }
-
-    RETURN_TYPES = ("INT", "INT")
-    RETURN_NAMES = ("final_height", "final_width")
-    FUNCTION = "calc"
-    CATEGORY = "utils"
-    def calc(self, tile_height, tile_width, overlap, tile_width_n, tile_height_n):
-        overlap_x = overlap
-        overlap_y = int(overlap * (tile_height / tile_width))
-
-        final_height = tile_height * tile_height_n - overlap_y * (tile_height_n - 1)
-        final_width = tile_width * tile_width_n - overlap_x * (tile_width_n - 1)
-        print("Final height: {}, Final width: {}".format(final_height, final_width))
-
-        return [final_height, final_width]
-
-
 
 NODE_CLASS_MAPPINGS = {
     "TileSplit": TileSplit,
     "TileMerge": TileMerge,
     "TileCalc": TileCalc,
+    "DynamicTileSplit": DynamicTileSplit,
+    "DynamicTileMerge": DynamicTileMerge,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
-    "TileSplit": "TileSplit",
+    "TileSplit": "TileSplit (Legacy)",
-    "TileMerge": "TileMerge",
+    "TileMerge": "TileMerge (Legacy)",
-    "TileCalc": "TileCalc",
+    "TileCalc": "TileCalc (Legacy)",
+    "DynamicTileSplit": "TileSplit (Dynamic)",
+    "DynamicTileMerge": "TileMerge (Dynamic)",
 }

standard.py

@@ -0,0 +1,179 @@
+import sys
+import os
+
+import torch
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy"))
+
+
+IMAGE_SIZE = 1472
+TILE_SIZE = 4096
+OVERLAP = 64
+
+
+# Splits an image in four tiles and returns them as a list
+class TileSplit:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "tile_height": ("INT", {"default": 64, "min": 64, "max": 4096}),
+                "tile_width": ("INT", {"default": 64, "min": 64, "max": 4096}),
+                "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "split"
+    CATEGORY = "utils"
+
+    def split(self, image, tile_height, tile_width, overlap):
+        height, width = image.shape[1], image.shape[2]
+        overlap_x = overlap
+        overlap_y = int(overlap * (tile_height / tile_width))
+
+        tiles = []
+        for y in range(0, height - tile_height + 1, tile_height - overlap_y):
+            for x in range(0, width - tile_width + 1, tile_width - overlap_x):
+                tile = image[:, y : y + tile_height, x : x + tile_width, :]
+                tiles.append(tile)
+
+        # Convert tiles list to a tensor if needed
+        tiles_tensor = torch.stack(tiles).squeeze(1)
+
+        return [tiles_tensor]
+
+
+class TileMerge:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
+                "blend": ("INT", {"default": 64, "min": 0, "max": 4096}),
+                "final_height": ("INT", {"default": 2048, "min": 0, "max": 9 * 4096}),
+                "final_width": ("INT", {"default": 2048, "min": 0, "max": 9 * 4096}),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "blend_tiles"
+    CATEGORY = "utils"
+
+    def blend_tiles(self, images, overlap, blend, final_height, final_width):
+        tiles = images
+        tile_height, tile_width = images.shape[1], images.shape[2]
+        original_shape = (1, final_height, final_width, 3)
+        overlap_x = overlap
+        overlap_y = int(overlap * (tile_height / tile_width))
+
+        batch, height, width, channels = original_shape
+        output = torch.zeros(original_shape, dtype=tiles.dtype)
+        count = torch.zeros(original_shape, dtype=tiles.dtype)
+        idx = 0
+
+        # for 3x3: custom_order = [0, 2, 6, 8, 1, 3, 5, 7, 4] # First 4 corners, then the sides, then the center
+
+        # Calculate grid dimensions
+        rows = (height - tile_height) // (tile_height - overlap_y) + 1
+        cols = (width - tile_width) // (tile_width - overlap_x) + 1
+
+        # Calculate the center of the grid
+        center_row, center_col = rows // 2, cols // 2
+
+        print("Rows: {}, Cols: {}".format(rows, cols))
+        print("Center row: {}, Center col: {}".format(center_row, center_col))
+
+        # Calculate the order in which to blend the tiles
+        # Order based on distance from center
+        distances = []
+        for i in range(rows):
+            for j in range(cols):
+                distance = abs(i - center_row) + abs(j - center_col)
+                distances.append(distance)
+
+        # Sort the tiles based on distance from center
+        reverse_custom_order = sorted(range(len(distances)), key=lambda k: distances[k])
+        custom_order = reverse_custom_order[::-1]
+
+        print("Custom order: {}".format(custom_order))
+
+        ys = [y for y in range(0, height - tile_height + 1, tile_height - overlap_y)]
+        xs = [x for x in range(0, width - tile_width + 1, tile_width - overlap_x)]
+        for idx in custom_order:
+            y = ys[idx // len(ys)]
+            x = xs[idx % len(xs)]
+
+            tile = tiles[idx]
+
+            weight_matrix = torch.ones((tile_height, tile_width, channels))
+            # if not center tile
+            for i in range(blend):
+                weight = float(i) / blend
+                weight_matrix[i, :, :] *= weight  # Top rows
+                weight_matrix[-(i + 1), :, :] *= weight  # Bottom rows
+                weight_matrix[:, i, :] *= weight  # Left columns
+                weight_matrix[:, -(i + 1), :] *= weight  # Right columns
+
+            old_tile = output[:, y : y + tile_height, x : x + tile_width, :]
+            old_tile_count = count[:, y : y + tile_height, x : x + tile_width, :]
+
+            weight_matrix = (
+                weight_matrix * (old_tile_count != 0).float()
+                + (old_tile_count == 0).float()
+            )
+
+            # Blend the old tile with the new tile
+            tile = tile * weight_matrix + old_tile * (1 - weight_matrix)
+
+            output[:, y : y + tile_height, x : x + tile_width, :] = tile
+            count[:, y : y + tile_height, x : x + tile_width, :] = 1
+
+        # Normalize the output and return
+        # output /= count
+
+        return [output]
+
+
+class TileCalc:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "tile_height": ("INT", {"default": 64, "min": 64, "max": 4096}),
+                "tile_width": ("INT", {"default": 64, "min": 64, "max": 4096}),
+                "overlap": ("INT", {"default": 64, "min": 0, "max": 4096}),
+                "tile_width_n": ("INT", {"default": 3, "min": 1, "max": 9}),
+                "tile_height_n": ("INT", {"default": 3, "min": 1, "max": 9}),
+            }
+        }
+
+    RETURN_TYPES = ("INT", "INT")
+    RETURN_NAMES = ("final_height", "final_width")
+    FUNCTION = "calc"
+    CATEGORY = "utils"
+
+    def calc(self, tile_height, tile_width, overlap, tile_width_n, tile_height_n):
+        overlap_x = overlap
+        overlap_y = int(overlap * (tile_height / tile_width))
+
+        final_height = tile_height * tile_height_n - overlap_y * (tile_height_n - 1)
+        final_width = tile_width * tile_width_n - overlap_x * (tile_width_n - 1)
+        print("Final height: {}, Final width: {}".format(final_height, final_width))
+
+        return [final_height, final_width]
+
+
+NODE_CLASS_MAPPINGS = {
+    "TileSplit": TileSplit,
+    "TileMerge": TileMerge,
+    "TileCalc": TileCalc,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "TileSplit": "TileSplit",
+    "TileMerge": "TileMerge",
+    "TileCalc": "TileCalc",
+}