Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Fix some clip vision regression. (#15506)

comfyanonymousd9f9d2ba1291
diff --git a/comfy/clip_model.py b/comfy/clip_model.py
index d7d3f994c..26cc5d7ee 100644
--- a/comfy/clip_model.py
+++ b/comfy/clip_model.py
@@ -314,13 +314,18 @@ class CLIPVisionModelProjection(torch.nn.Module):
if "projection_dim" in config_dict:
self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)
else:
- self.visual_projection = lambda a: a
+ self.visual_projection = torch.nn.Identity()
if "llava3" == config_dict.get("projector_type", None):
self.multi_modal_projector = LlavaProjector(config_dict["hidden_size"], 4096, dtype, device, operations)
else:
self.multi_modal_projector = None
+ def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+ if "{}visual_projection.weight".format(prefix) not in state_dict:
+ self.visual_projection = torch.nn.Identity()
+ super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
def forward(self, *args, **kwargs):
x = self.vision_model(*args, **kwargs)
out = self.visual_projection(x[2])