About generative task

[zwl666666]

Great work! However, I have some questions regarding the image generation part.

Specifically, how is the pixel decoder implemented? Could you provide more details? My current understanding is that a decoder is needed to be retrained to map the discriminative tokens back to the pixel space, but achieving such high-quality generation results based on this alone leaves me a bit puzzled. Or is it that the VQGAN part is entirely fixed and does not require training, with the discriminative tokens serving only as conditional inputs to generate the corresponding VQ tokens, which are then mapped back to the pixel space by the VQ decoder?

Looking forward to your response.