Hi DreamDojo team,
Thanks for releasing the paper and code. I have a question about the temporal consistency loss in Eq. (4).
The relevant tensors appear to use the layout [B, C, T, H, W], but the released implementation computes:
vt_pred[:, 1:] - vt_pred[:, :-1]
This applies the difference along the channel dimension. Based on Eq. (4), I expected the temporal difference to be:
vt_pred[:, :, 1:] - vt_pred[:, :, :-1]
Could you clarify whether the released implementation is intentional or an indexing typo? Also, were the results in Table 5 and Figure 9 obtained using the channel-axis or temporal-axis implementation?
The paper mentions that this loss “enhances object completeness and reduces artifacts.” Were there any additional ablations or evaluations specifically supporting this observation, such as different loss weights, multiple seeds, or object-completeness metrics?
Thank you!
Hi DreamDojo team,
Thanks for releasing the paper and code. I have a question about the temporal consistency loss in Eq. (4).
The relevant tensors appear to use the layout
[B, C, T, H, W], but the released implementation computes:This applies the difference along the channel dimension. Based on Eq. (4), I expected the temporal difference to be:
Could you clarify whether the released implementation is intentional or an indexing typo? Also, were the results in Table 5 and Figure 9 obtained using the channel-axis or temporal-axis implementation?
The paper mentions that this loss “enhances object completeness and reduces artifacts.” Were there any additional ablations or evaluations specifically supporting this observation, such as different loss weights, multiple seeds, or object-completeness metrics?
Thank you!