Clarification Request on Differences Between RollingForcing and MAGI-1

[mikeallen39]

Dear Authors,

Thank you very much for sharing your insightful paper—I found it highly interesting and thought-provoking. While reading, I noticed some conceptual similarities between your approach and MAGI-1.

However, I’m currently trying to better understand why your proposed "rolling forcing" appears to yield improved performance compared to MAGI-1. Would you be willing to elaborate on the key differences between your method and MAGI-1, especially regarding the design and advantages of rolling forcing?

I’d greatly appreciate any clarification you could provide—it would help deepen my understanding of your contributions.

Thank you again for your excellent work!

[Kunhao-Liu]

Thank you for your interest in our work. While our Rolling Forcing shares some similarities with MAGI-1 in terms of denoising frames with varying noise levels, there are key differences that Rolling Forcing introduces to improve long video quality:

1. In MAGI-1, frames with varying noise levels use causal attention, meaning each frame can only attend to previous frames. In contrast, Rolling Forcing employs bidirectional attention for frames in the denosing window, allowing frames to attend to both past and future frames for mutual refinement and thus reduce error accumulation.
2. We integrate the attention sink mechanism into Rolling Forcing, which persists the initially generated frames as a global context anchor that maintains long-term global consistency.
3. During training, we perform distillation using self-generated frames as context, rather than frames from ground-truth datasets. This eliminates the train-test discrepancy known as exposure bias.
4. As a result of this distillation, Rolling Forcing is approximately 100x faster than MAGI-1 in terms of latency on a single GPU.