search

yuval-alaluf / stylegan3-editing

Official Implementation of "Third Time's the Charm? Image and Video Editing with StyleGAN3" (AIM ECCVW 2022) https://arxiv.org/abs/2201.13433
https://yuval-alaluf.github.io/stylegan3-editing/
MIT License
660 stars 72 forks source link

about smoothing #44

Closed TaekyungKi closed 1 year ago

TaekyungKi commented 1 year ago

Hi thanks for this interesting research.

I recognized, after the smoothing, that first 2 and last 2 frames (total 4 frames) are removed when back to video, right? Formally, video_frames[2:-2] is actually survived. That's because of the smoothing formula. it starts computing the smoothing with frame number 2 (index i-2 for non-negative as paper). Is there any way to include those 4 frames while maintaining the smoothness ?

Best,

TaekyungKi commented 1 year ago

i copied the first and the last frame 2 latents, respectively and concatenated them with the original. This approach is not fancy but, it does not hurt the consistency and can match the original frame length. it works well because the adjacent latents are already quite close and smoothing procedure is just giving the latent vectors a bit of consistency. i did this with video of 25 fps. i tried several tricks but this is the most stable. i hope it helps someone.