question regarding preprocessed_images

[fzhang612]

Hi,

Thanks for sharing this work.

Got a questing regarding this line

stylegan-encoder/encoder/perceptual_model.py

Line 21 in 46605b2

preprocessed_images = (preprocessed_images - drange[0]) * ((drange[1] - drange[0]) / 255) + drange[0]

I guess the purpose is to normalize reference image to [-1,1] because the discriminator requires it. The bug described in

stylegan-encoder/encoder/perceptual_model.py

Line 22 in 46605b2

# ("+ drange[0]" at the end is a fix of a bug in tflib.convert_images_from_uint8())

doesn't have any impact because it is the generator_output_tensor, not generated_image_tensor is fed into the discriminator, as specified in

stylegan-encoder/encoder/perceptual_model.py

Line 50 in 46605b2

return generator_output_tensor

. So I guess the correct way to normalize is simply preprocessed_images * ((drange[1] - drange[0]) / 255) + drange[0].

Please correct me if I miss something else.