VAR generation roundtrip

[zaptrem]

@zhengkw18 When you generate an image with the reference model for VAR you decode it to pixels then reencode it to (now slightly different!) tokens. Why?

[zhengkw18]

This ensures that the self-generated latents and the latents from the dataset both lie in pixel-encoded space. We empirically find that this strategy performs better.