Thank you for your work.
You mentioned in the appendix that you directly shifted from a bidirectional 4-step model to self-forcing DMD training. I would like to ask：
- if this approach cause a temporal discontinuity in the visual effect (especially in the early iterations)？ Since the model does not possess autoregressive (AR) capabilities and only has DMD loss, can we still achieve good results by shifting the model from bidirectional attention to unidirectional attention?
- Is batch size important? If we train with a batch size of 8, will the performance drop significantly?
Thank you for your work.
You mentioned in the appendix that you directly shifted from a bidirectional 4-step model to self-forcing DMD training. I would like to ask：