Initial version

__init__.py

@@ -0,0 +1,4 @@
+from .clipperpweight import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS
+
+
+__all__ = ['NODE_CLASS_MAPPINGS', 'NODE_DISPLAY_NAME_MAPPINGS']

clipperpweight.py

@@ -0,0 +1,41 @@
+import torch
+
+
+class CLIPTextEncodePerpWeight:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"text": ("STRING", {"multiline": True}),
+                             "clip": ("CLIP", ),
+                             }}
+    RETURN_TYPES = ("CONDITIONING",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning"
+
+    def encode(self, clip, text):
+        empty_tokens = clip.tokenize("")
+        empty_cond = clip.encode_from_tokens(empty_tokens, return_pooled=False)
+        tokens = clip.tokenize(text)
+        unweighted_tokens = [[(t, 1.0) for t,_ in x] for x in tokens]
+        unweighted_cond, unweighted_pooled = clip.encode_from_tokens(unweighted_tokens, return_pooled=True)
+
+        cond = torch.clone(unweighted_cond)
+        for i in range(unweighted_cond.shape[0]):
+            for j in range(unweighted_cond.shape[1]):
+                weight = tokens[i][j][1]
+                if weight != 1.0:
+                    token_vector = unweighted_cond[i][j]
+                    zero_vector = empty_cond[i][j]
+                    perp = ((torch.mul(zero_vector, token_vector).sum())/(torch.norm(token_vector)**2)) * token_vector
+                    cond[i][j] = token_vector + (weight * perp)
+        
+        return ([[cond, {"pooled_output": unweighted_pooled}]], )
+
+
+NODE_CLASS_MAPPINGS = {
+    "CLIPTextEncodeMultiLayer": CLIPTextEncodePerpWeight,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "CLIPTextEncodeMultiLayer": "CLIP Text Encode (Perp-Weight)",
+}