The dataset used when training UNet-latent (addition UNet for latent Refusion)

[novice0224]

I would like to ask the following two questions:

1. When training the latent UNet (the additional U-Net for latent Refusion), how many images are in your training dataset? (Specifically, how many images are there for HR dehazing and for Bokeh effect transformation in the training dataset?)

2. During the training of the latent UNet, you mentioned that the training patch size can be set to 256 × 256. I would like to ask if you randomly crop the large-sized image into 256 × 256 patches, or do you resize the larger image to 256 × 256?"

[Algolzw]

Hi! The number of training data for HR dehazing is about 45 (all are high-resolution) and we use the cropped patches to train the U-Net.