How to calculate the distance between composed feature and target image feature?

[SunTongtongtong]

Hello there,

Thanks for publish the excellent work! I have following questions:

1. Based on the ReadMe usage2, we can achieve the fusion feature(based on reference image + modified text), how do you calculate the distance between the fusion feature and target image feature? Do you calculate it based on cosine similarity or euclidean distance？
2. Do we use the clip embedding for the target image?

Best

[SanghyukChun]

1. It is cosine similarity, i.e., Euclidean distance after L2 normalization. You can check the details in the code
2. Yes. We directly compute the similarity between edited feature (query image & query text) and the CLIP visual feature (target image)