Visualize which areas in the image contribute to the loss

[gafniguy]

Hi,

Thanks for the paper and making it so accessible and easy to use.

For pixel-aligned losses, it's easy to show which areas of the generated image are wrong, by doing e.g L1(generated, GT) pixel-wise and showing it as an image, resulting in some sort of error map.

What do you think would be a good way to get something like this for LPIPS? I would like to see which areas in the image contributed to higher LPIPS distance.

Thanks

[richzhang]

Hi, thanks for the question. Look at setting spatial=True here: https://github.com/richzhang/PerceptualSimilarity/blob/master/lpips/lpips.py#L23

[gafniguy]

Perfect, thank you so much.