The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Don't initialize CLIPVision weights to default values.
comfyanonymousfa2cca056c96
diff --git a/comfy/clip_vision.py b/comfy/clip_vision.pyindex efb2d5384..a95707e46 100644--- a/comfy/clip_vision.py+++ b/comfy/clip_vision.py@@ -1,4 +1,4 @@-from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, CLIPImageProcessor+from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, CLIPImageProcessor, modeling_utilsfrom .utils import load_torch_file, transformers_convertimport osimport torch@@ -6,7 +6,8 @@ import torchclass ClipVisionModel():def __init__(self, json_config):config = CLIPVisionConfig.from_json_file(json_config)- self.model = CLIPVisionModelWithProjection(config)+ with modeling_utils.no_init_weights():+ self.model = CLIPVisionModelWithProjection(config)self.processor = CLIPImageProcessor(crop_size=224,do_center_crop=True,do_convert_rgb=True,