similarity filter

README.MD

@@ -26,7 +26,7 @@ The intention for this repository is to build a suite of nodes that can be used
 - **QuickShapeMask**: Rapidly generate shape masks (circle, square) with customizable dimensions.
 - **DTypeConverter**: Convert masks between different data types (float16, uint8, float32, float64).
 - **FastWebcamCapture**: High-performance webcam capture node with resizing capabilities.
-
+- **SimilarityFilter**: Filter images based on the similarity of previous frames and block execution if the similarity is too high.
 ### VAE Tools 🖼️
 - **TAESDVaeEncode**: TAESD VAE encoding node.
 - **TAESDVaeDecode**: TAESD VAE decoding node.

__init__.py

@@ -1,6 +1,6 @@
 from .controls.value_controls import FloatControl, IntControl, StringControl
 from .controls.sequence_controls import FloatSequence, IntSequence, StringSequence
-from .controls.utility_controls import FPSMonitor
+from .controls.utility_controls import FPSMonitor, SimilarityFilter
 from .controls.motion_controls import MotionController, ROINode, IntegerMotionController
 from .quick_shape_mask import QuickShapeMask
 from .tiny_vae import TAESDVaeEncode, TAESDVaeDecode
@@ -14,8 +14,10 @@ NODE_CLASS_MAPPINGS = {
     "IntSequence": IntSequence,
     "StringSequence": StringSequence,
     "FPSMonitor": FPSMonitor,
+    "SimilarityFilter": SimilarityFilter,
     "MotionController": MotionController,
     "IntegerMotionController": IntegerMotionController,
+
     "ROINode": ROINode,
     #"IntervalControl": IntervalCo  ntrol,
     #"DeltaControl": DeltaControl,
@@ -44,8 +46,10 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "TAESDVaeDecode": "TAESD VAE Decode (RyanOnTheInside)",
     "DTypeConverter": "DType Converter (RyanOnTheInside)",
     "FastWebcamCapture": "Fast Webcam Capture (RyanOnTheInside)",
+    "SimilarityFilter": "Similarity Filter (RyanOnTheInside)",
 }
 
+
 WEB_DIRECTORY = "./web/js"
 
 __all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS'] 

controls/utility_controls.py

@@ -3,6 +3,7 @@ import time
 import numpy as np
 import cv2
 import torch
+import random
 
 class FPSMonitor(ControlNodeBase):
     """Generates an FPS overlay as an image and mask"""
@@ -121,3 +122,77 @@ class FPSMonitor(ControlNodeBase):
         self.set_state(state)
         
         return (state["cached_image"], state["cached_mask"])
+
+class SimilarityFilter(ControlNodeBase):
+    """A node that filters out similar consecutive images to prevent unnecessary workflow execution using StreamDiffusion's approach."""
+    
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "image": ("IMAGE",),
+                "always_execute": ("BOOLEAN", {"default": False}),
+                "threshold": ("FLOAT", {"default": 0.98, "min": 0.0, "max": 1.0, "step": 0.01}),
+                "max_skip_frames": ("INT", {"default": 10, "min": 1, "max": 100, "step": 1}),
+            }
+        }
+    
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "update"
+    CATEGORY = "real-time/control/utility"
+
+    def __init__(self):
+        super().__init__()
+        self.cos = torch.nn.CosineSimilarity(dim=0, eps=1e-6)
+
+    def update(self, image, threshold=0.98, max_skip_frames=10, always_execute=False):
+        print(f"[DEBUG] Input image object: {hex(id(image))}, shape: {image.shape}, device: {image.device}")
+        
+        # Get state with defaults
+        state = self.get_state({
+            "prev_image": None,
+            "skip_count": 0
+        })
+
+        # First frame case
+        if state["prev_image"] is None:
+            state["prev_image"] = image.detach().clone()
+            state["skip_count"] = 0
+            self.set_state(state)
+            return (image,)
+
+        # Calculate cosine similarity
+        cos_sim = self.cos(
+            state["prev_image"].reshape(-1), 
+            image.reshape(-1)
+        ).item()
+
+        # Calculate skip probability using StreamDiffusion's approach
+        if threshold >= 1:
+            skip_prob = 0
+        else:
+            skip_prob = max(0, 1 - (1 - cos_sim) / (1 - threshold))
+
+        # Generate random sample
+        sample = random.uniform(0, 1)
+
+        # If we should skip (probability check)
+        if skip_prob >= sample:
+            # Check if we've skipped too many frames
+            if state["skip_count"] >= max_skip_frames:
+                state["prev_image"] = image.detach().clone()
+                state["skip_count"] = 0
+                self.set_state(state)
+                return (image,)
+                
+            # Skip frame - return ExecutionBlocker to prevent downstream execution
+            state["skip_count"] += 1
+            self.set_state(state)
+            from comfy_execution.graph import ExecutionBlocker
+            return (ExecutionBlocker(None),)
+        
+        # Frame is different enough - process it
+        state["prev_image"] = image.detach().clone()
+        state["skip_count"] = 0
+        self.set_state(state)
+        return (image,)

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui_realtimenodes"
 description = "These nodes are for real-time applications of ComfyUI."
-version = "1.0.0"
+version = "1.1.0"
 license = {file = "LICENSE"}
 dependencies = ["opencv-python"]