why do you choose three different loss for different preception model？

[shoutOutYangJie]

I see "F.cosine_similarity", "MSE", "L1" 3 loss type for different preception model, why design it?

[ZhaoChuyang]

Hey! The MSE loss implementation is borrowed from dino_perceptual, and I think it is equivalent to L1 for optimization. For models originally optimized with cosine similarity (e.g., contrastive learning in CLIP and SigLIP), we keep cosine similarity as the perceptual metric. Our design principle is that the metric should align with the one used during the model's original training. For models not trained with distance-based learning, we use L1 by default.