search

nihaomiao / CVPR23_LFDM

The pytorch implementation of our CVPR 2023 paper "Conditional Image-to-Video Generation with Latent Flow Diffusion Models"
BSD 2-Clause "Simplified" License
432 stars 43 forks source link

null_cond_prob #46

Closed h-bouzid closed 2 weeks ago

h-bouzid commented 1 month ago

Hi dear @nihaomiao, Could you please explain to me what is the role of null_cond_prob variable? I couldnt fully grasp it.

and why do you change it from training 0.1 to test 1?

Thank you for your time

nihaomiao commented 1 month ago

Hi, @h-bouzid, thanks for your interest in our paper! During training, we use classifier-free guidance, which needs to switch between unconditional and conditional generation modes. This is controlled by null_cond_prob. During testing, as shown in Table 6, using guidance can slightly improve the FVD score but double the inference time (since we need to run both unconditional and conditional generation at each reverse step). So we choose 1 as the default setting.