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
577 stars 42 forks source link

About Refusion #34

Open Trevor-Philips-cbd opened 1 year ago

Trevor-Philips-cbd commented 1 year ago

Hello, for your Refusion method, I can't seem to find the corresponding part in your paper in the code, such as the part described in the article using GT and LQ to get latent features. I refer to your demo files in each task, and found that just changing the .yml can switch between IR-SDE and Refusion, but I can't find the difference with the IR-SDE method except the backbone network for predicting noise.

But when I looked at the unet-latent file in your project, I found that you separated this part of the code. My question is, did you train the latent model and diffusion separately?

Algolzw commented 1 year ago

Hi! Yes, we train the latent model separately and use the pretrained U-Net weights for the latent-diffusion model as in the latent-dehazing code.