Open zelenooki87 opened 5 months ago
RIFE is the only model that is in the top rank in terms of perceptual quality with available pre-trained models, according to this ranking.
I would assume IFRNet would be the easiest to implement because there is a Vulkan ported version.
This is false.
Oh, that was my layman's assumption. Some videos I interpolated with ema-vfi and ifrnet in a conda environment seemed to me to be somewhat better than rife. In one of the previous issues I opened, you wrote to me that ema-vfi is too slow to consider adding it. Has anything changed in the meantime? You added swinir and scunet and big respect for that. :) Greetings
Now the GridSample
operator is the bottleneck that affects all modern video frame interpolation models.
@WolframRhodium What about to implement new interpolation method? https://github.com/mulns/PerVFI
or this one? (seems better) https://github.com/JHLew/MoMo
You can do these yourself.
Is it possible to add other interpolation methods? I would assume IFRNet would be the easiest to implement because there is a Vulkan ported version. Is it possible to add some other interpolation from this benchmark page? https://github.com/AIVFI/Video-Frame-Interpolation-Rankings-and-Video-Deblurring-Rankings Thank you very much in advance