DinoV3 compatible with transformers version 4 and 5

README.md

@@ -14,7 +14,7 @@
 
 | Date | Description |
 | --- | --- |
-| **2026-04-20** | Added node "Fill Holes Nicely with Meshlib" |
+| **2026-04-20** | Added node "Fill Holes Nicely with Meshlib"<br>Fixed DinoV3 Features Extractor |
 | **2026-04-06** | Added new "DINO-Lock" functionality<br>New fill_holes in Sparse Generator<br>Thanks to Easymode on Discord |
 | **2026-04-05** | Added node "Extract Images from Video"<br>Can be used with "Sparse Generator with ReconViaGen" |
 | **2026-04-04** | Added node "Sparse Generator with ReconViaGen" |

trellis2/modules/image_feature_extractor.py

@@ -83,11 +83,21 @@ class DinoV3FeatureExtractor:
         hidden_states = self.model.embeddings(image, bool_masked_pos=None)
         position_embeddings = self.model.rope_embeddings(image)
 
-        for i, layer_module in enumerate(self.model.layer):
-            hidden_states = layer_module(
-                hidden_states,
-                position_embeddings=position_embeddings,
-            )
+        # transformers < 5
+        if hasattr(self.model,'layer'):
+            for i, layer_module in enumerate(self.model.layer):
+                hidden_states = layer_module(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                )
+        elif hasattr(self.model,'model') and hasattr(self.model.model,'layer'): # transformers >= 5
+            for i, layer_module in enumerate(self.model.model.layer):
+                hidden_states = layer_module(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                )
+        else:
+            raise Exception("Cannot extract features")
 
         return F.layer_norm(hidden_states, hidden_states.shape[-1:])
         

trellis2/trainers/flow_matching/mixins/image_conditioned.py

@@ -85,11 +85,21 @@ class DinoV3FeatureExtractor:
         hidden_states = self.model.embeddings(image, bool_masked_pos=None)
         position_embeddings = self.model.rope_embeddings(image)
 
-        for i, layer_module in enumerate(self.model.layer):
-            hidden_states = layer_module(
-                hidden_states,
-                position_embeddings=position_embeddings,
-            )
+        # transformers < 5
+        if hasattr(self.model,'layer'):
+            for i, layer_module in enumerate(self.model.layer):
+                hidden_states = layer_module(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                )
+        elif hasattr(self.model,'model') and hasattr(self.model.model,'layer'): # transformers >= 5
+            for i, layer_module in enumerate(self.model.model.layer):
+                hidden_states = layer_module(
+                    hidden_states,
+                    position_embeddings=position_embeddings,
+                )
+        else:
+            raise Exception("Cannot extract features")                
 
         return F.layer_norm(hidden_states, hidden_states.shape[-1:])