Some doubts about the multi-scale input and output of the super-resolution part about gigagan-pytorch HOT 5 CLOSED

@boya34 it isn't the original image, but the low resolution image being given to the unet. although paper was scant on details on upsampler, any resolution smaller than the low resolution image given to the unet, and the unet would simply learn to downsample

from gigagan-pytorch.

thank you for your reply, but I still feel a bit confused, let me explain in more detail.
As the paper says, the input image size of unet is 64 pixels, and the output image size is 512 pixels, so in version 0.2.6 the allowable rgb resolutions is [128, 256](bigger than 64 and smaller than 512), while the multiscale_input_resolutions is [64, 32, 16, 8], this mismatching will lead to an error (the related assertion is around line 1841 in gigagan_pytorch.py).
Besides, the rgbs choose the size bigger than original size which is 64 (in unet_upsampler.py), so the rgb_index is [128, 256, 512] , but it should be the same as multiscale_input_resolutions(the related assertion is around line 1627 in gigagan_pytorch.py).
The paper says the generator producing a pyramid {xi}, and x0 is the full image(64 pixels I guess), so I understand the meaning of multiscale_input_resolutions, could you please tell me more about the other two arrays? thanks a lot.

from gigagan-pytorch.

@boya34 you can resolve this error (and perhaps i should auto-resolve it for unet as it is confusing), by setting Discriminator(multiscale_input_resolutions = unet.allowable_rgb_resolutions)

i have also decided to just default to sending the discriminator the low resolution input image (going into the unet) as the lowest resolution rgb always

from gigagan-pytorch.

@lucidrains I get what you mean, Thank you again!

from gigagan-pytorch.

no problem, happy training!

from gigagan-pytorch.