Regarding the fact that the clip library can only be used to read data using the PIL library, we hope that the official can modify the source code

[qsd-github]

My code is as follows:

model, preprocess = clip.load("./weights/ViT-B-32.pt", device=device) image = preprocess(Image.open("./xxx.jpg")).unsqueeze(0).to(device) # only PIL library can be used here, using cv2 library will result in an error.

I took a look at the source code: def load(name: str,..): .... return return model, _transform(model.visual.input_resolution) # The problem is that _transform()

def _transform(n_px): return Compose([ ToPILImage(), # Adding ToPILImage() here and importing the corresponding library can solve the problem of not being able to read cv files Resize(n_px, interpolation=BICUBIC), CenterCrop(n_px), _convert_image_to_rgb, ToTensor(), Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)), ]) This is my usage source code and solution. If you use it differently from my source code and support the cv2 library, please share your opinion. I suggest the official modify this location!