Latent-Refusion v.s. Refusion

[MU-MU-M]

Hi, I've only seen the code for refusion but no code for latent-refusion. Are there the codes for latent-refusion and the latent-refusion pretrained model in this package? Also, I have no idea about the model architecture of latent-refusion. Is latent-refusion just pretrained autoencoder + NAFNet? On the other hand, I mainly want to focus on deblurring application, but there is only ir-sde pretrained model and no refusion pretrained model for deblurring in this package.

Thanks.

[Algolzw]

Hi, the latent-refusion code is the latent-unet. If you want to perform latent-refusion on the deblurring task, you need to pretrain the latent-unet (U-Net autoencoder) on your dataset. Please refer to the latent-dehazing for more details.

[MU-MU-M]

Are both the training flow and inference flow of the latent-unet for deblurring application the same as latent-unet for dehazing? That is to say for deblurring, do we need to compute latent_lq, latent_gt, hidden_lq, and hidden_gt? But here is a question, when performing inference without ground truth (since actual deblur has no ground truth), how should I perform latent-unet testing?

Thanks.

[Algolzw]

In inference, we don't need the latent_gt and hidden_gt, you can just set them to None.