hzwer / Practical-RIFE

We are developing more practical frame interpolation approach.
MIT License
542 stars 60 forks source link

Which time ratios are supported? #76

Open bonlime opened 2 months ago

bonlime commented 2 months ago

Hey, I see that in inference image by default you use powers of 2 for interpolation, but also support passing arbitrary "ratio". I'm trying to understand how RIFE is being trained and what are the actual possible inference timestamps.

Do you use random [0, 1] during training or only use powers of 2 as possible options? If powers of 2, what is the largest i can use? The reason i ask is because generally models don't work well for values outside of their training domain, so i'm being very cautious here

hzwer commented 1 month ago

During training, I will randomly select three frames from index 1-9. For example, if I select 1 5 8, then time=0.571.