About training/inference speed improvements of TiTok

Thank you for the great job! I just have some doubt about the efficiency of TiTok.

In the paper of TiTok, throughput (samples/s/gpus) of TiTok is basically larger than MaskGIT. However, take TiTok-L-32(32 tokens) for example, quantized latent tokens(32) is concatenated with mask tokens(256) to make sure decoder have enough tokens to generate image, so the tokens fed into decoder should be 32+256=288, larger than MaskGIT(256 tokens). Since time complexity of transformer is O(N^2)，I just wonder why TiTok is faster than MaskGIT.

I'm new to the area of image generation, so I would be much grateful for your reply. If I have any misunderstanding about TiTok, please point it out.

The issue names the TiTok paper, TiTok-L-32, and MaskGIT but no repository file, test, or entry point. Start by reading the cited token-count and throughput discussion, then compare the relevant TiTok and MaskGIT explanations. Done would require a clear, maintainer-confirmed explanation of the apparent speed difference.

Written by the indexing model from the issue text.