Topaz Video AI: Not Fixing Faces or Adding Details

I am new to TOPAZ and I have a lot to learn. I thought TOPAZ was like a magic button used to upscale and fix videos that were generated using AI like - RunWay Gen3 or KLING. For example, I have a small group of people looking up to the sky (overhead shot) generated in RW GEN3 at 1280 x 720. All I want to do is upscale 2X and fix the faces and some small details in the scene. I have tried all of the models with not much success. What can I do to have TOPAZ enhance the faces? When I try to increase details, sharpen, etc it just looks worse. Is there a good tutorial that can help me?

Thank you!!

If it helps you understand, TVAI existed years before AI generated videos were made available to the public. All of their models are trained on real videos that are made and/or owned by Topaz.

The Iris and Rhea models are made to enhance faces. Turning up the sharpen slider seems to increase the amount of face enhancement. (Rhea is intended for large resolution videos.)

One thing TVAI has never been able to do is create something that looks good out of something that looks bad. Or as other have said: garbage in equals garbage out. Some other reasons you might not be seeing any results is from the opposite end of that statement. Where the AI determines that the quality is good and won’t increase it any more. So for example, you can have a bad looking image, but since it doesn’t have any compression artifacts or anything else on the long list of things to correct, the AI has no other choice but to conclude that it is intentional.

1 Like

Yeap, I have heard of the G in G out! LOL So what do you suggest I do? Should I stick with Iris? I am only going from 1280 x 720 to 2X. And it sounds like the Image going IN needs to be good? PLease see attached image. Are those faces to far gone to fix?

I am using Iris medium quality mostly. Maybe just give it a try!

You might get something by using the add grain feature. The biggest issue with this image is that it does look intentional.

What does Intentional mean?

OK let me try

That the distortions on the faces are not the result of image compression or being out of focus, or bad lighting.

AH!! I see. I understand. But I thought TOPAZ was supposed to fix that?

Depends on the source and context. Ideally, it’s not going to make faces clear that are supposed to be out of focus—but it does tend to do that. My experience with the special face enhancement models is that they tend to make monster faces when the face is a certain size. Proteus is much better about leaving well enough alone and being able to do a good general enhancement. It cannot get the same level of facial detail that Iris or Rhea can get, but the whole image is more consistent with itself and the intended focus.

1 Like

OK thank you. I guess my last question is: what are the correct settings for Detail Recovery, Sharpening, etc? Is there a way for me to test each one and preview in the correct order?

I’ve only ever upscale from DVD to FHD or 1280 by 720 to FHD. Both need very different settings. 1280 by 720, Proteus 4 Auto has always given the best results for me.
For DVDs Proteus 4 Manual:
Fix compression 20 to 50.
Improve detail 12 or less.
Sharpen 22 or less, or Anti-alias/Deblur 10 to 24. Not both active at the same time. They compound each other.
Reduce noise 0.
Dehalo 0.

Recover detail should always be at 0. (It defaults to 20.)

Trying to upscale to 4K will need different settings. Proteus 3 is very different in how the settings respond to the video.

1 Like

do you still use Proteus v3 in any use case of yours? or you are solo on Prob-4 (when proteus is in use of course)?

Yes I am only using Proteus 4. Proteus 3 only worked if there was little to no noise in the source. Because I don’t need to denoise first with Proteus 4, I am able to preserve more details on every video than with Proteus 3.

1 Like

Thank you for all the help!

Thank you for all the help! I will try these settings.

@ForSerious and others: Could I get your opinion on going from 480p to 1280 or 1920…

The 480p source is blocky noise YouTube video that I want to improve. When I try, the faces get unreal looking.

I’m very conservative in how much enhancement I let TVAI do.

My first thought is to try Proteus 4 Manual, and only change the ‘Fix Compression’ slider. I’d start at like 40. This will do both denoising and hopefully deblocking. There’s no reason to do an intermediate step with only Proteus. Just upscale straight to 1920.

Thanks!!