JingyunLiang / RVRT

Recurrent Video Restoration Transformer with Guided Deformable Attention (NeurlPS2022, official repository)
https://arxiv.org/abs/2206.02146
Other
354 stars 33 forks source link

Parameters description #19

Open r3a1d3a1 opened 1 year ago

r3a1d3a1 commented 1 year ago

Could you please give a brief explanation how --tile and --tile_overlap arguments affect quality and speed so that enthusiasts don't have to read the paper to use the code?

P.S. I noticed that the tile option should be at least 'x 64 64', otherwise the interpolate doesn't work, and that with only 4GBs of VRAM, I couldn't use the 'x 128 128' option for inference on REDS. Would be worth mentioning in the README.

Thanks!