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.

Don't initialize CLIPVision weights to default values.

comfyanonymousfa2cca056c96
diff --git a/comfy/clip_vision.py b/comfy/clip_vision.py
index 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_utils
from .utils import load_torch_file, transformers_convert
import os
import torch
@@ -6,7 +6,8 @@ import torch
class 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,