ECCV2022-RIFE
ECCV2022-RIFE copied to clipboard
hzwer
1.5 model has problems with patterns, that model 1.2 didn't have.
1.6 has a lesser defect in the top row of the building just below and to our left of the chimney in the center
what is the model difference? just other images in the training set or something?
OK - Model 1.6 is better with patterns than 1.5 but still has big problems with it. Model 1.2 was much better at patterns but had other issues.
lol i was super grossed out by this defect on the building until i realized that it was a smudge on my screen
this is nice looking stuff
New models still have problems with patterns. The differences between the results are very small from model to model. I made a collection of clips that cause problems. Here you can download the sourceclips (25fps) and try yourself.
https://www.magentacloud.de/share/hbx2la7qrk
Model 2.0 is quite good, but still has problems with patterns. Try my sourceclips https://www.magentacloud.de/share/hbx2la7qrk
Have you checked model 2.2? havent tested it myself with the source clips, but my pattern problems in my clips i tested are almost completely gone
Yes, I check each new model with my sourceclips. In some cases the result is better but not completely gone. Still enough problems to bother the filmfun.
Model 2.3 brought a very very (very) small improvement of this stuff. But there is nearly no visible change to model 2.2. These clips cause still big problems for Rife :(
@ChristianUndCo I add an command-line argument, you can set --scale=2.0 or --scale=4.0 to get better result for your cases.
Similar issue, not sure if you would like to take a look at these but these are just some videos I had issues with bad artifacting on. I did encode them to be a lot smaller so they aren't too big.
"Focus" - Has artifacting across the image on fade to black/white, and artifacts on flashes (1:24-1:48) "Problem" - Has artifacting on fast spinning moving line pattern (2:22-2:29) "Side To Side" - Has lots of artifacting on the wall even though the movement isn't too large and on fade to blacks (2:24-3:28)
Source clips (re-encoded) (deleted - you'll have to message me for them if you see this)
(I use x8 also)
If the model can even just improve those type of things that would be amazing. These are the only easily noticeable artifacts I've run into. Thanks for the work you put into making the best FI AI. :)
Screenshots of issues
(Fade to black)
(Fast movement pattern)
(Slower movement pattern)
Note: Upscaling to 4K does help with the last type of artefact
I just tested model 3.1. There is nearly no improve with patterns. Some stuff gets even worse. But upscaling from "576 SD content" to 720p or 1080p helps sometimes a little bit.