Open mrgreen3325 opened 2 years ago
Hi, during inference the LR image y should be combined with a randomly drawn z for upsampling. y should contain some features that favor the inverted process -- this is one advantage of joint training the downscaling and upscaling procedures, which is the motivation of many previous works that use encoder-decoder architectures, and our model can also partly benefit from this.
Thanks for your great work. I read your paper too, and I got several questions. May I know that does the z is only used in the loss function computation? Do we need to provide an input LR y with z during influence? Is the downsampling process from x to y also embed some information in y, such that y contains some features that favor the inverted process for upsampling, or the process is just similar to the bicubic downsample? Thank you.