Long Video Training

Hi,

Thanks for releasing this great work!

when will the training code be released?

1. what sink sizes and window size do we use for a latent frame wise autoregressive model?
2. for the long Long Video Training is source anchored DMD an essential part? after naively implementing based on paper, i see some videos collapse to black videos.
3. for the source anchored DMD what is the real_source_guidance_weight? w_src?
4. do we weigh the losses of different segments difrerently? For example do i weigh the clips from further in the rollout more as they might have more error accumulation?

Note that i tried the LHAD from your paper ontop of stage 2 from causal forcing++