Why do the CLIP-SIM metrics I ran show a relatively large discrepancy?

[xinyunxing]

[kornelhowil]

Hi! Thank you for your interest in our work and sorry for the late reply.

Are these results averaged across all prompts and images? Did you follow our procedure as described in the appendix, specifically in section B2?

CLIPGaussian models were trained according to the settings described in Appendix A, except for the "Fire" condition, which used a higher feature learning rate (0.02). Additionally, for the lego object with the "Fire" condition, we set λ_bg = 100.