search

Algolzw / image-restoration-sde

Image Restoration with Mean-Reverting Stochastic Differential Equations, ICML 2023. Winning solution of the NTIRE 2023 Image Shadow Removal Challenge.
https://algolzw.github.io/ir-sde/index.html
MIT License
558 stars 41 forks source link

About the training of the latent model #50

Open ruiw293 opened 1 year ago

ruiw293 commented 1 year ago

When training the latent model, the test results fluctuate between good and bad periodically。 The training sets I used are DIV2K and Flickr2K.

ruiw293 commented 1 year ago

1J`05(5JNJW`@96P`@R }D9

ruiw293 commented 1 year ago

XVDU1@@YSWWPKLLN$}R$OW2 And no matter how the training is done, the best result is just to make the reconstruction of latentGT+hiddenLQ more similar to LQ HQ： 00000_HQ LQ： 00000_LQ result： 00000

Algolzw commented 12 months ago

Hi! Have you solved this problem? The result of latent_GT + hidden_LQ should be close to the GT image.