Details of quality drift evaluation

[y0urOy]

Excellent work.

I would like to ask whether the Self-Forcing model reported in Table 1 of the paper was evaluated using the officially released checkpoint.

In my experiments, I used the official Self-Forcing weights:
https://huggingface.co/gdhe17/Self-Forcing/blob/main/checkpoints/self_forcing_dmd.pt

I evaluated the model on the same 200 MovieGen prompts, generating 30-second videos.
The results I obtained are:

Quality Drift: 4.76 （Different with 1.66 reported in the paper）
First 5s‘ Imaging Quality: 70.06
Last 5s’ Imaging Quality: 65.31

Could you please share more details about how Quality Drift is computed and evaluated in your Table1 experiments?