Question about some implementation details

[kevinchiu19]

First of all, thank you for sharing such great work and code.

I have two questions here:

1. When giving inference to video_gen (image_pair), why is the reconstructed rendering result passed in instead of the original image directly?
2. I see that the 'noise_timestep' seems to be fixed in the inference code, and I don't see any DDPNet related code?

[Albertchen98]

I also have the feeling that the effectiveness of DDPNet is a question mark.

And there are also several points unclear in the paper:

• In Sec 3.3, they claim that a T' is introduced to eliminate the train-test-mismatch, but during LCM distillation, they seem still to sample t in the range of [0,T].
• why the reward function of DDPNet won't force t --> 0 ??? the following reward function should be better if they really applies DDPNet in their training process:
  $$r(x_0^r, t) = -\mathcal{L}_{MSE} - \lambda \cdot \text{sim}(x_0^{pred}, x_0^r)$$

wish the author could clarify these points