Topaz Video AI Alpha v4.1.0.0.a (Proteus v4)

My first impression much more detailed than the previous alpha yes I am happy so far
I still use relative to auto and only 7 recover detailed information and always 2 X enlarging
the video VHS is a visit from hungarian Youngsters in 1994 to Holland flying with a fokker friendship as a start and cockpits where open and food in the plane in touristclass was like a michelin star restaurant
a precious video totally corroded blurred because of the ferro VHS tape but with every topaz release coming a bit more to live

It is not slow in converting but I use a 4080 and not a 4050 4050 is 8 times cheaper and 3 time slower

No watch closely iris mq has still the Craquelé if you look closely unless you love a damaged badly conserved mona lisa

So the Protheus v4 is better

Thank you and happy New Year to you and all at Topaz!
David

Examples are much better than descriptions no matter how vivid they are​:+1::100:

protheus MQ is different than Protheus Fine tune Second enhancement

Why You use Nyx 1 in manual mode try to use a more generic artemis hq with dione clear as second phase

or protheus only v4 auto recover detail 7

Well I didn’t want to spam a bunch of screenshots, but since you asked, here you go. Aforementioned black hole issue with Proteus v3 fixed in Proteus v4. The black hole is on the door to the left of the characters. This is a still, but it looks quite chaotic in motion.


And here’s a demonstration of Iris v2’s deblurring of blurred background characters vs Proteus v4, which better respects the input cinematography. (Proteus v3 is similar in this regard.) Compare the girl’s eyes and uh, forehead horns.

Iris v2:


Proteus v4:

And finally here’s a synthetic test I set up by taking an existing blu-ray, downscaling it to 360p, and using various Topaz models to re-upscale it back to 1080p. This demonstrates that the Proteus models offset the picture to the up-left compared to the “ground truth” blu-ray source and the other models. You may have to download or view these images in their own tabs to make the comparison.

Ground-truth BD screenshot:


Downscaled to 360p, then upscaled via Artemis HQ:

Downscaled to 360p, then upscaled via Iris v2 (“MQ”):

Downscaled to 360p, then upscaled via Proteus v3:

Downscaled to 360p, then upscaled via Proteus v4:

6 Likes

I understand. Thank you so much for providing them. Really brings it home as to what your seeing. And the details you provide are generous and helpful. Much appreciated​:+1::100::trophy:

Why the downscale step if you don’t mind?

Like I said it’s a synthetic test. It lets you compare the upscale to a “ground truth” to judge the qualities of the upscaler. It’s not a “real world” usage scenario.

2 Likes


Comparing Proteus v3 and Proteus v4

3 Likes

I am just getting Error without any further message why.
On the left: IrisMQ on the right Proteus v4

logsForSupport.zip (175.1 KB)

No one to blame except for having a weak gpu that is the rtx 4050

Proteus v4 seems to need more memory or so…can´t run XMedia Recode in the background it leads to error.
The quality of Iris & Proteus v4 is nearly the same…i can see more details on Iris…in times the Alienfaces would be seen Proteus is blurring it out…
But overall the details of iris are better and the colors look more natural…Proteus has more green/yellow. Maybe in times the Alienfaces accour too much Proteus could be an alternative.

Here i try to get the Previews of Iris on the left side and Proteus on the right.
I posted this error also in the past that sometimes it´s not possible to choose the last preview…now again always Iris(the first preview) is choosen when i click on the last preview(Proteus).
If i want Proteus again on the right side i have to delete the Preview of Proteus choose the right side and render a new Preview instead on just clicking of the rendered thing.
logsForSupport.zip (229.3 KB)

I don’t believe the devs have said that a tensor version of the model is available yet. Until that appears you’re stuck in the slow lane.

1 Like

in my previous desktop computer i had the radeon RX 570 and since 2020 the software is slow as a snail.
so what do i need to get myself? a NASA server to get more than 3 fps?

A rx 570 is also not particularly fast as well if you look at fp16/fp32 performance lol. 3 fps is honestly about right when it comes to ML tasks like this. A 3080 or something will already be much better.

1 Like

still i ask decent FPS i am not asking to run a file at 3000 FPS but something like 20 fps a software that has been over 3 years in development… that is PAID mind you should be quite optimized by now

Try using identical manual settings for the comparison. “Auto” is kinda yolo.

1 Like

Agree about not dimming. It’s unnecessarily bothersome.

1 Like

Yes. Just put the “Unprocessed” label over the display and dispense with the dimming. It’s overkill.

4 Likes