So engine need to reconstruct 'current' frame using only samples data from 'other' frame (it can use 'current' frame only as 'objects positions reference'). In simplest form with 2 frame only input - we have only 'current' and 'other' frame. I'm sure there more options when using PyTorch directly but I have not experimented with that.Įdit: can someone test the 2x_LD-Anime_Skr_v1.0.pth model? It crashes on me.ĭTL: The 'training condition' for motion compensation process is simple enough: For given number of frames with marked 'current' frame it is required to create 'current' frame view using only 'other' frames. I'm all ears if anyone has any suggestions. But while researching how to convert between the two formats I only saw that mentioned once or twice. Small ones are quick (few seconds), the largest one I converted (64mb) took around a minute.Ģ) I have not compared them to the original version. Credits to NMKD (n00mkrad)Ģ) are results bit identical to the pytorch version?ġ) Depends on the size of the model. Here are all the checkpoints for DeGif that I have. I mention it because sometimes an earlier checkpoint produces better results.

Avs video editor 32 bit