search

mulns / PerVFI

Official code base of "Perception-Oriented Video Frame Interpolation via Asymmetric Blending" (CVPR 2024), also denoted as ''PerVFI''.
https://openaccess.thecvf.com/content/CVPR2024/papers/Wu_Perception-Oriented_Video_Frame_Interpolation_via_Asymmetric_Blending_CVPR_2024_paper.pdf
Apache License 2.0
50 stars 4 forks source link

bad results, not much improvment over AMT #1

Open eisneim opened 5 months ago

eisneim commented 5 months ago

hi thanks for sharing this work! but after some experiments, the interpolation result is not good, any advice on how to get the best result?

method: raft+pervfi scale 4

https://github.com/mulns/PerVFI/assets/5195894/d69a5cb4-3f21-4909-8454-d58cd5cf7a9a

https://github.com/mulns/PerVFI/assets/5195894/19bb841b-ff97-48b2-91bb-99acd45372c4

0000075 0000027

still waiting for vidim this paper seems good at interpolate large motion

eisneim commented 5 months ago

comparing with MoMo which uses Diffusion model when creating in between frames, MoMo does better job

https://github.com/mulns/PerVFI/assets/5195894/ff9a4782-0157-498d-8f14-cd5d999bb526

tkkcc commented 5 months ago

Hi, would you like to share your input video.

eisneim commented 5 months ago

sample-video.zip and gdrive link: https://drive.google.com/file/d/1TrNLRIwkUcsbci2WPy9rzfyLzxgCn1oX/view?usp=drive_link

mulns commented 5 months ago

Currently, we only release model for scale X2. The model trained for X4 and X8 interpolation is under preparation