search

danielroich / PTI

Official Implementation for "Pivotal Tuning for Latent-based editing of Real Images" (ACM TOG 2022) https://arxiv.org/abs/2106.05744
MIT License
897 stars 112 forks source link

Effect of lpips type #46

Closed luchaoqi closed 1 year ago

luchaoqi commented 1 year ago

Hello,

I find that the option of lpips_type = 'alex' tend to affect my inversion results: with lpips_type = 'alex', it introduces undesirable checkerboard/noise-like artifact which gives me a sense of overfitting (input left and reconstruction right) image image with lpips=type='vgg', results are smoother which gives me a sense of underfitting: image image you may need to zoom-in a bit to tell the difference. any suggestions in this case? do I need to tune options like LPIPS_value_threshold = 0.06 to find a sweet point between this trade-off?

danielroich commented 1 year ago

Hi @luchaoqi, playing with the hyper-parameters of the L2 and the lpips as you suggested is the best option in my opinion. Moreover, you might change the convergence criteria in the configuration file as well so the current alexnet would not overfit