Generating longer videos with lower resolution using AR model

[zlenyk]

Hello! QQ about my understanding of current architecture of the autoregressive model for video generation.

Current limit of frames generated comes from "max_seq_len" - how many tokens we keep as context for generating new tokens. We shouldn't generate more tokens to start loosing tokens from beginning.
But doesn't it mean that if we set video resolution to 2x smaller (so 4x fewer tokens), we should be able to generate 4x longer videos? If I'm not mistaken it's just the memory constraint.

Is there an error in that logic?
Thank you!