Axel Mertes has a great idea for: New AI Scaling Method

Continuing the discussion from [Video Enhance AI Feature Requests]

This is his original post below:

Feature Wish: NEW AI Scaling Method

I just recently stumbled over this amazing two minute paper video, that blew my mind:

Enhance! Super Resolution Is Here! :mag:

If you think GigaPixel AI and VEAI are outstanding in terms of scaling, that this is squarely nothing compared to what this approach seems to promise:

(Screen grabs from the video linked above from Two Minute Papers youtube channel)

I want to do THIS with my old PAL interlaced videos…
No idea if that ONLY works with faces right now and how much computing resources it might take to make this work for full images, but the direction is clear I guess.

The BIG questions are: Will this work, will it work “in time” and will it work temporary stable for video/film usage?

Regardless of all this, if those results are true then this is mind boggling.

i saw the video too on two-minute papers, but I tried it in the google colab and the results were bad, I don’t know I do something wrong or what, but here is the colab to test it:
Google Colab

PS: it will take about 40 Min to enhance one photo

but here is another colab with much better results :

hope you test the two different colab and tell me your feedback.

There are many software can restore face better than GAI / VEAI.
For example, Remini APP, MyHeritage web and Tencent ARC’s GFPGAN.

Here is an example I upscaled with Tencent ARC’s GFPGAN.

I also tested with sample image shown in the Google SR3 video.

Here is the upscaled image by Tencent ARC’s GFPGAN
(It may not look as detail as Google SR3, but it only took a few seconds to process)

There are also AI software built in Face restoration for video such as, AVCLabs Video Enhancer AI.

BSRGAN test image source

VEAI (ALQ 12 4x):
BSRGAN test image source MConverter.eu_4.00x_640x640_alq-12.mp4_000001.274

AVCLabs (Ultra Upscale x4)
BSRGAN test image source_Ultra Upscale(x4).mp4_000000.198


That’s very interesting!! Yeah, they should definitely incorporate some of these types AI models capable of restoring like this. It would be cool if it was like an entire scene that it could restore rather than just faces.


I am not sure… I will have to try it out myself sometime when I get around to it. Thanks for sharing your input.

1 Like