The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Add a penultimate_hidden_states to the clip vision output.
comfyanonymouse85be36bd2c1
diff --git a/comfy/clip_vision.py b/comfy/clip_vision.pyindex daaa2f2bf..9b95ae003 100644--- a/comfy/clip_vision.py+++ b/comfy/clip_vision.py@@ -49,12 +49,16 @@ class ClipVisionModel():precision_scope = lambda a, b: contextlib.nullcontext(a)with precision_scope(comfy.model_management.get_autocast_device(self.load_device), torch.float32):- outputs = self.model(pixel_values=pixel_values)+ outputs = self.model(pixel_values=pixel_values, output_hidden_states=True)for k in outputs:t = outputs[k]if t is not None:- outputs[k] = t.cpu()+ if k == 'hidden_states':+ outputs["penultimate_hidden_states"] = t[-2].cpu()+ else:+ outputs[k] = t.cpu()+return outputsdef convert_to_transformers(sd, prefix):