rectified-flow-pytorch
rectified-flow-pytorch copied to clipboard
lucidrains
Why LPIPS loss uses classification layers of VGG and not original LPIPS repo which are calibrated for perception
Hello, Thankyou for detailed and yet easy to follow implementation of Rectified Flow paper.
I was going through the original implementation by authors and the authors use the original LPIPS library as noted below,
https://github.com/gnobitab/RectifiedFlow/blob/5a1fd4dd3ea7db764ce370a84ce35f9c8b15fde6/ImageGeneration/sde_lib.py#L28
But in your implementation LPIPS is just the vanilla VGG loss at the classification layer https://github.com/lucidrains/rectified-flow-pytorch/blob/1f59e7e9122439931968126d4ac662c1f0ce24ac/rectified_flow_pytorch/rectified_flow.py#L69
Any reason for not using the actual repository of LPIPS and instead using the original VGG? I believe in LPIPS VGG is modified wherein some calibration is done to align with human perception.
Have you done so because of the below issue, https://github.com/richzhang/PerceptualSimilarity/issues/72#issue-941220573
Or nothing in specific?
Thankyou
