Extend CLIP Score to image-to-image and text-to-text similarity

[jonasricker]

🚀 Feature

Extend CLIP Score to operate within the same modality (image-to-image, text-to-text)

Motivation

Especially image-to-image similarity would be useful to quantify the semantic similarity of two images/datasets.

Pitch

Extend the existing CLIP Score class to support images/texts for both arguments. This might serve as an inspiration.

Alternatives

Create separate metrics for image-to-image and text-to-text similarity, in order to better group them in the existing categorization. In that case, the three metrics could share some underlying code.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[rittik9]

@Borda I am interested in working on this.