Question regarding dynamic sequence length adaptation

[HL-hanlin]

Thanks for the great work!

While going through the codebase, I noticed that the _adapt_sequence_length function performs linear interpolation from the current sequence length in BAGEL to the target sequence length in Wan. Could you share the intuition behind this interpolation?

For example, have the authors conducted ablation studies on alternatives such as using the current length directly as the target length for Wan, or zero-padding to a fixed maximum sequence length (e.g., 512)?

Thanks!