So I just tried it with Gaia with grain and the result is almost even worse in my opinion.
Topaz just simply does not seem to be up to the task regardless of the settings or model.
1 Like
You should not go above 239% from the original, then the building model kicks in. Also called 4x.
I think the other programs instead use Lanczos from 2x (200%) and above to joint to the wished resolution.
You can try using literally any resolution but it makes practically no difference in the upscaling technique of the face itself:
If you say the competitors may use Lanzos for facial upscaling it just begs the question why Topaz does not use Lanzos for the same. The result is what matters here and I donât see Topaz offering any viable solution for this use case atm.
Garbage in, garbage out.
2 Likes
AVCLabs looks incredibly tempting. $199, for a lifetime licence!
I have tested it and it is pretty good but it have some problems with the textures going flat & looks a little animated on the edges.
Doesnât it come with slider options?
Yes but imagine you have very low res highly compressed video with faces Topaz cannot handle you pass the video first through AVCLabs or HitPaw Video Enhancer first and you can Enhance further with Topaz or other program.
Of course I donât state this is an ideal workflow and I wish Topaz would be at least on par or very close to these solutions mentioned so we could do everything in Topaz including faces.
meimeiriver in fact there are not many adjustment options just to mainly select model and the level of AI Enhancement, then 1-1 slider for Colorize, motion compensation and 4 for video settings(brightness, saturation, contrast, sharpen) and thatâs about it. The toolset is not the vast like in Topaz of course but when it comes to the faces these programs are the way to go as Topaz cannot offer anything close in the face restoration department atm.
3 Likes
My apologies to the devs. Yes, I strayed, and went and bought AVCLabs. What can I say? Itâs crap. 
Some of the options are kinda cute, but I didnât find their handling of (small) faces particularly better than TVAI. And even their Ultra mode leaves me with a disturbing lack of removing âcompression artifactsâ and such. Thereâs overall simply way too few options/sliders. Spent entire day tesing vids, but, at the end of the day, I happily went back to TVAI. 
1 Like
I actually agree with you, overall Topaz Video AI is much more superior when it comes to all the options and overall improvement on textures in the videos - I just found one specific usecase with specific type of low res video material AVCLabs Video Enhance can work better with when it coems to faces.
Thatâs why I brought it up and thatâs why I am puzzled how Topaz is so much behind in that particular usecase. I want Topaz to improve, for me it would be also ideal to do everything through Topaz Video AI. 
By the way if you bought it and wanna return it donât forget AVCLabâs 30 day money back guarantee should still apply. 
1 Like
Sadly, it doesnât. 
The fine-print says they will not refund for you having changed your mind. Oh well, lesson learned.
N.B. I can always use it as a 1-pass proccesor for them edge cases you mentioned; but that comes with great drawbacks, especially in the area of denoising, as once A.i. has messed with noise, said remaining noise will no longer be uniformly distributed (and can thus no longer be effectively removed).
I stand corrected:
I just had to jinx it, LOL. Yikes!
EDIT: And this is what the competion (AVCLabs) makes of this shot:
And back to TVAI, now with Proteus v4:
Proteus remains a wickedly good model.
I would totally use the deblur slider here probably to 100%, donât you want to apply?
Yes this is a sample of worst case we sometimes have and models fails. AVC Labs seems for me does less, is simpler and that is the reason itâs look better 
âŠbut only in such cases.
It does not lower the noise as TVAI, then bad âmoon faceâ is behind the noise. You can see the fluffy unstructured faces exists in the original when the fog is away âŠand then rhea/Iris runs over it, does recovery of this âpixel thingâ, faces that are not made for recovery. How you can told the model should recover and what not, big question, sometimes itâs better doing nothing.
