About CLIP training on nosied images #44
Description
Hey! I think GLIDE is a wonderful work. But I have a question about CLIP training on nosied images.
I want to know why CLIP can be trained on nosied images. I think if t (range from 0 to 1000) is large(maybe close to 500 or more), then the noised images hardly contain any semantic information. In this case, I want to know CLIP model how to encode similar features from noised images and text and I also think it may cause model to not converge (because it is hard to encode similar features between noised images and text)