fix(reference): moved reference image to IGlobal

Author: anirudhk-tech

nodes/src/nodes/visual_similarity_filter/IGlobal.py

@@ -37,6 +37,7 @@ class IGlobal(IGlobalBase):
 
     embedder = None
     config = None
+    reference_embedding = None
 
     def beginGlobal(self):
         self.device_lock = threading.Lock()
@@ -47,8 +48,10 @@ def beginGlobal(self):
             return
 
         from .embedder import FrameEmbedder
+
         self.embedder = FrameEmbedder(self.config)
 
     def endGlobal(self):
         self.embedder = None
+        self.reference_embedding = None
         self.device_lock = None

nodes/src/nodes/visual_similarity_filter/IInstance.py

@@ -38,7 +38,6 @@ def beginInstance(self):
         self._match_rows = []
         self._image_buf = bytearray()
         self._image_mime = 'image/png'
-        self._reference_embedding = None
 
         cfg = self.IGlobal.config
         self._fps = float(cfg.get('fps', 1.0))
@@ -52,7 +51,6 @@ def open(self, obj: Entry):
         self._total = 0
         self._forwarded = 0
         self._match_rows = []
-        self._reference_embedding = None
 
     def close(self):
         if self.instance.hasListener('table') and self._match_rows:
@@ -92,10 +90,11 @@ def _on_frame_complete(self, image_bytes: bytes, mime: str):
         self._total += 1
         timestamp = idx / self._fps if self._fps > 0 else float(idx)
 
-        # First frame: capture as reference and forward unconditionally.
-        if self._reference_embedding is None:
+        # First frame ever: capture as reference and forward unconditionally.
+        if self.IGlobal.reference_embedding is None:
             with self.IGlobal.device_lock:
-                self._reference_embedding = self.IGlobal.embedder.embed(image_bytes)
+                if self.IGlobal.reference_embedding is None:  # double-checked under lock
+                    self.IGlobal.reference_embedding = self.IGlobal.embedder.embed(image_bytes)
             self._matched += 1
             self._match_rows.append([idx, self._fmt_time(timestamp), '1.000'])
             self._forward_frame(image_bytes, mime)
@@ -104,9 +103,10 @@ def _on_frame_complete(self, image_bytes: bytes, mime: str):
         # Subsequent frames: score against the reference.
         try:
             import numpy as np
+
             with self.IGlobal.device_lock:
                 frame_emb = self.IGlobal.embedder.embed(image_bytes)
-            similarity = float(np.dot(self._reference_embedding, frame_emb))
+            similarity = float(np.dot(self.IGlobal.reference_embedding, frame_emb))
         except Exception:
             similarity = 0.0
 

nodes/src/nodes/visual_similarity_filter/requirements.txt

@@ -0,0 +1,2 @@
+transformers
+accelerate

nodes/src/nodes/visual_similarity_filter/services.json

@@ -1,25 +1,16 @@
 {
 	"title": "Visual Similarity Filter",
 	"protocol": "visual_similarity_filter://",
-	"classType": [
-		"video"
-	],
+	"classType": ["image"],
 	"capabilities": [],
 	"register": "filter",
 	"node": "python",
 	"path": "nodes.visual_similarity_filter",
 	"prefix": "visual_similarity_filter",
-	"description": [
-		"Filters video frames by visual similarity using CLIP embeddings. ",
-		"The first frame received becomes the reference; every subsequent frame ",
-		"is scored against it and forwarded only if its similarity meets the threshold. ",
-		"Includes pre-roll/post-roll context around matching segments."
-	],
-	"icon": "visual-similarity.svg",
+	"description": ["Filters video frames by visual similarity using CLIP embeddings. ", "The first frame received becomes the reference; every subsequent frame ", "is scored against it and forwarded only if its similarity meets the threshold. ", "Includes pre-roll/post-roll context around matching segments."],
+	"icon": "embedding-image.svg",
 	"lanes": {
-		"image": [
-			"image", "table"
-		]
+		"image": ["image", "table"]
 	},
 	"input": [
 		{
@@ -84,52 +75,34 @@
 		},
 		"vsf.clip-base": {
 			"object": "clip-base",
-			"properties": [
-				"vsf.similarity_threshold",
-				"vsf.fps"
-			]
+			"properties": ["vsf.similarity_threshold", "vsf.fps"]
 		},
 		"vsf.clip-large": {
 			"object": "clip-large",
-			"properties": [
-				"vsf.similarity_threshold",
-				"vsf.fps"
-			]
+			"properties": ["vsf.similarity_threshold", "vsf.fps"]
 		},
 		"vsf.custom": {
 			"object": "custom",
-			"properties": [
-				"vsf.embedding_model",
-				"vsf.similarity_threshold",
-				"vsf.fps"
-			]
+			"properties": ["vsf.embedding_model", "vsf.similarity_threshold", "vsf.fps"]
 		},
 		"vsf.profile": {
 			"title": "Embedding model",
 			"description": "CLIP model and configuration",
 			"type": "string",
 			"default": "clip-base",
-			"enum": [
-				"*>preconfig.profiles.*.title"
-			],
+			"enum": ["*>preconfig.profiles.*.title"],
 			"conditional": [
 				{
 					"value": "clip-base",
-					"properties": [
-						"vsf.clip-base"
-					]
+					"properties": ["vsf.clip-base"]
 				},
 				{
 					"value": "clip-large",
-					"properties": [
-						"vsf.clip-large"
-					]
+					"properties": ["vsf.clip-large"]
 				},
 				{
 					"value": "custom",
-					"properties": [
-						"vsf.custom"
-					]
+					"properties": ["vsf.custom"]
 				}
 			]
 		}
@@ -138,9 +111,7 @@
 		{
 			"section": "Pipe",
 			"title": "Visual Similarity Filter",
-			"properties": [
-				"vsf.profile"
-			]
+			"properties": ["vsf.profile"]
 		}
 	]
 }