The preprocessing convert the images to 224x224, can we do higher resolution?

[CanyonWind]

When we compare the similarity between higher resolution images, e.g. 512x512 or 1k x 1k, downsampling to 224 will cause losing of details. Is there any way to compare between higher resolutions and how do you estimate the impact of DreamSim score, as the models were pretrained on 224 x 224 and haven't seen higher-res images? Thanks

[ssundaram21]

Hi -- great question. At the moment DreamSim can be run with higher resolutions, but is trained on 224x224 images (as are the CLIP/DINO backbones). We are working on variants that are calibrated for higher resolutions and will update this issue when we have them up and running!

[imneonizer]

It would be interesting to see any benchmarks which shows whether calculating embeddings with high res images may be 512x512 yields any better accuracy.