Chunked Frames prevents divergence. Overlapped Conditioning prevents chunk-to-chunk discontinuity.
Variable Length training and inference enables the model to generate videos of arbitrary lengths, as shown at the 1st and 59th seconds.