I tried to retrained stylegan3-r on FFHQ256x256 with reduced channel numbers.
As this doc (https://github.com/NVlabs/stylegan3/blob/main/docs/configs.md) says,
i expected that reducing channel numbers will not reduce quality.
I set channel_base= 16384 and channel_max=256, so automatically doubled in model, and also set r1 to be 2.
Now FID reach 12.xxx but the results has serious weird artifacts and don't seem like FID 12.xxxx.
This didn't happened when i trained StyleGAN2 config.
i'm working on video generation in latent level. But provided weights are too heavy for my model. I expected StyleGAN3 can reduce the serious problem of aliasing. I already read related issues #185 , #77 etc. is there really no way to avoid this phenomenon?
I tried to retrained stylegan3-r on FFHQ256x256 with reduced channel numbers. As this doc (https://github.com/NVlabs/stylegan3/blob/main/docs/configs.md) says, i expected that reducing channel numbers will not reduce quality. I set channel_base= 16384 and channel_max=256, so automatically doubled in model, and also set r1 to be 2.
Now FID reach 12.xxx but the results has serious weird artifacts and don't seem like FID 12.xxxx. This didn't happened when i trained StyleGAN2 config.
i'm working on video generation in latent level. But provided weights are too heavy for my model. I expected StyleGAN3 can reduce the serious problem of aliasing. I already read related issues #185 , #77 etc. is there really no way to avoid this phenomenon?