Start here
Begin with probability and tensors. Then follow AE → VAE → VQ-VAE → VQ-VAE-2 → VQGAN → VAR. Return to the attention and evaluation guides whenever you need them.
Next, explore three branches: Transformer reconstruction, diffusion/flow decoding, and shared semantic representations. Finally, connect them to continuous autoregression. This is a teaching order, not a claim that every paper directly descends from its neighbor.
In each article, first read the problem and the change. Select architecture modules to inspect shapes and pseudocode. Then manipulate the mechanism. Adjustable teaching curves are not reported model results. Open the pinned source when you need implementation details.