Any tips on preserving smooth faces?

[skv89]

I use 3BQ8 model, which works wonders but the problem is it amplifies facial imperfections. Wrinkles, facial bumps, pores, all become extremely amplified. Is there anyway to keep the face smoother? I'm not asking for face beautifying but I would prefer face beautifying over amplifying imperfections.

[naxci1]

Can you share a picture of the workflow?

[skv89]

[noland81]

latent noise scale: 0.001-0.010. You can try to increase the smoothness.

[skv89]

I just tested it but doesn't really help. The only thing that kind of help is using Topaz Iris, which smooths out the faces a bit. Oh well, it is what it is.

[naxci1]

In my experience, setting temporal_overlap to 0 is better, as I haven't seen any benefit from it; it only slows down the speed. For best results, at least 3x magnification is required. These models are trained for 3-4x upscaling and give the best results with them. The same is true for Topaz models; even FlashVSR gives the best results at 4x. Currently, the workflow uses 2x upscaling. If I were you, I would reduce the video by 50% and then upscale it to 3x or 4x. It's possible to do this losslessly with a shutter encoder. I use lab for color correction.

[urixl]

In my experience temporal overlap is necessary.

[naxci1]

I haven't seen any benefit from it in the videos I used; it might have different effects in different videos. Therefore, it's necessary to do tests beforehand for each video, before performing a long process.

[skv89]

I use Davinci Resolve to downscale using Lanczos. It is the best downscaling method I've tested and the resulting quality is better than what I get using FFMPEG and Handbrake. Below are the settings I use to downscale in Davinci. You can also use ProRes instead of MP4 but file size is much larger. After extensive testing I get the best result downscaling with Davinci and then applying Proteus Natural prior to SEEDVR2.

I do see improvement in temporal overlap. I do use 3x in some cases but sometimes having to downscale to 1/3 for poor quality 1080p videos, the small distant faces lose so much detail that the resulting face generated by SEEDVR2 makes it look like another person. The main issue I'm having is that SEEDVR2 3BQ8 is so good at creating/amplifying details that it creates and amplifies imperfections in faces which I don't like.