swapping-autoencoder-pytorch icon indicating copy to clipboard operation
swapping-autoencoder-pytorch copied to clipboard

some details about the paper and project

Open hughwcq opened this issue 3 years ago • 1 comments

Hi,it`s a amazing work!But there are some details i want to know. In the paper:"A manipulation vector for snow is discovered by taking mean difference between 10 user-collected photos of snowy and summer mountain. The vector is simply added to the texture code of the input image (red) with some gain." and may I know more detail about the value of gain and how to get the Continuous interpolation from "winter“ to "summer"?

hughwcq avatar Feb 25 '22 10:02 hughwcq

The strength, or the magnitude of the manipulation vector is decided by the distribution of the 10 user-collected photos. For example, if the snowy mountain collection contains many visuals of summer mountain, and if the summer mountain collection contains some snow images, the manipulation vector will contain "weak" signal, and the value of gain needs to be larger. I recommend setting the value of gain at 2.0, and adjust it by looking at the output images. The continuous interpolation is simply achieved by gradually increasing the gain from 0.0 to some large value.

taesungp avatar Mar 01 '22 20:03 taesungp