finetune CLIP and the difference between current repo vs previous version

[YangjiaqiDig]

Hey, thanks for your excellent work!
After reading the paper and checking the code, I have two questions:

1. Looks like the new version in the current repo doesn't require any CLIP fine-tune steps anymore like what is mentioned in the paper.
2. The current repo's performance is better than the previous version, even without CLIP fine-tuning on your target datasets, so what is the main impact of the performance improvement? Does it prove that fine-tuning CLIP does not matter for the results?
   Thanks!