Use a simple CLIP model implementation instead of the one from transformers. This will allow some interesting things that would too hackish to implement using the transformers implementation.