You seem to be missing the big picture entirely. The big picture is, AI or not, we're still guessing what belongs in those pixels. there's no magic way of finding out what those pixels should be, as the data is perpetually lost without a higher resolution original source. While, yes, AI is much, much better at guessing (if it's a good AI, anyway), but it's still guessing. For some reason you're trying to defend AI beyond this point, and I don't see why. I'm not saying AI is working at the same level as a bilinear algorithm, but I am saying that AI is, at the end of the day, just guessing, because it is just guessing. Of course, i mention bilinear, bicubic, etc for a good reason: the OP and a few others want to understand what's going on. Obviously, someone who doesn't already understand bilinear transformation (as evidenced by the fact they say have an interest by don't point to any knowledge of where they are in their information journey) isn't about to digest the various forms of AI. There was an implicit request for a better understanding of what's going on, so trying to pick some sort of battle with me over whether or not I'm giving AI enough credit (which is what DiffuseReflection was going on about, which you then responded to) is not even remotely helpful, as it's coming at me not only for something that's not wrong, but at the far advanced end of a scope that multiple people are approaching for the first time.

Just for you, I decided to download the video and run it through ffmpeg to output indivual frames out in png. Already data will be lost from all the conversions, but given I saw exactly what you're referring to, I wanted to take a shot myself. I highly doubt it's AI on the face of it, 'cause we're looking at trying to run something realtime. So focusing on frames 372 to 840...

First thing i notice is that frames 555 to 574 show quite a bit of blurring between the lip and the chin. Meanwhile, the parts that aren't moving (or moving much) are incredibly crisp. 589 to 630 shows it really well on the forehead. Now to understand what's going on here, there's 2 things you need to know: video compression is about descaling sections of images, rather than upscaling images. Secondly, there's more than one video compressor being used in this video.

The way a video compressor generally works comes down to focusing on several "key frames" then the frames after it, rather than being full frames, only contain the bits that change. I don't happen to have a tool handy that i know how to use to show this clearly, but i think you can see this very plainly, yourself. If you were to look at the "LOKI" to his right (your left), you'll notice it doesn't move. Neither does the wall, the dirt on the wall, the CB, the disney plus logo, etc. Since they don't move, and since the non-moving pixels make up the majority of frames, what typically happens is that you have a "key frame" which contains a full frame, then several frames after that frame (until the next key frame) will actually only contain the data for that which has changed from the previous frame, because over 50% of the data from frame to frame is not changing, so if you can keep it that way, on frame data alone you can cut the size between key frames in half just by reducing key frames. The next bit is, well, cutting down on the information between those key frames. Not only do you have a recording framerate issue (bluring will happen as a result of movment, but since certain parts aren't moving, they're clear), but you can further cut down on the size of of the frames outside of the key frames by using really, really lossy compression (blurring) to make it so the colors in the blurry bits are more similar, thus easier to compress (saying something like 88 99 72 90 27 90 19 40 17 is bulky, but if you can turn that to say 87 87 87 87 87 87 87 87 87 87 by blurring, you can then say 87 x 10 and it takes much, much less space [simplification, as we have more complex algorithms than simply repeating a single byte]). Since our eyes are already used to it, it's not as noticeable.

Now, here's the rub. The ZOOM and Skype video codecs tend to be really, really lossy, and that's the first filter you see. Then, on top of it, we have whatever filter is being used for OBS or whatever the interviewer is using to record the screen. Then, what's more, is the video editing program that she's using is going to do this, yet again. Now, since I downloaded it, I can see what the metadata of the file is. Which says... "handler_name : ISO Media file produced by Google Inc." which, in turn, tells me that Google threw another compression algorithm on top of the 3 already being used. This is also assuming the camera did no compression of it's own. What you see is weird is that the compressors didn't blur anything other than the movement (the things outside of the key frames) so the whole scenery and even parts of him that barely move look really, really crisp, but he, himself, looks like a smudge, and given he's the focal point, not the scenery, this looks really, really uncanny to you.

I'm not sure ImageMagick will have all the filters, but getting an equivalent effect could indeed work. Honestly ImageMagick is confusing in it's layout and usage, and you are pretty much going blind with inputs and then looking at the results. If you had a simple job to do or you do it in bulk (for example, to fix transparency issues i split the alpha as a separate section, have a super fast program scan and do the work, then put it back in) then once you get the right commands and behavior as you expect then yes it's a great tool.

I'm still new to GIMP as well, but at least i can now get basic stuff done i used to use PSPX for.

Yeah, specifically image magick for batch jobs or quick jobs. My primary usage of image magick is part of upload script that i use for uploading images to my website on command line. I might have like 10 images at a time, and i want to convert them to PNGs to auto-scrub all the tags and for a few other reasons. gifs are also converted to mp4, iirc. Then uploaded and the uploaded urls are then spit back at me. imagemagick handles the conversion process, except for gif images (ffmpeg). GIMP i use for any editing of the image, while MS paint i'll use for quicker resizing of images for uploading to gog (just quicker to get in and out of mspaint).

Thanks kohlrak. Very informative. Here's a little gift for you.

At least quote me when you're trying to sarcasticly blast me. Glad to see we've went from talking about image upscaling to creating fake people, which is kinda related, but not actually the same thing.

You see sarcasm where there is none, sir! I genuinely appreciate the "Just for you" time you took to analyse and explain the artefacts on that video. I genuinely learned something. Obviously, being interested in these things, I thought you'd get a kick out of that little online AI toy. Sorry if I gave the wrong impression.

Well in my experience ImageMagick commandline goes with the 'image options' approach. So assume you have an image and a mask/alpha as separate files, you'd do 'convert image1 image2 -alpha copy outfile'. If you are working with the same image multiple times you'd do use -clone to specify a previous frame, but if you change the frame you may get more than you expect. Example for printing something i did -clone 0 +append -clone 0 +append, got 4 not 3 of something, so when i did it a 4th time i got 8 instead of 4.

A lot of options use -, but a few use + instead. so -append adds to the bottom and +append adds to the right of the image.

The -fx feature is nice, but doing something should be very simple because complex stuff takes a VERY long time, so much i wrote a bmp program to do the job of the -fx i gave it because it's 1000 times faster for that specific filter.

In the event you change the image fundamentally (change canvas size, multiple images making use of transparency, etc) -composite will rebuild the image otherwise it assumes you're doing a sequence and you'll get outfile-1 outfile-2, etc.

Forcing output to be a 24bit bmp file you need -type truecolor -define "bmp:format=bmp3"

Sorry, that's my mistake. I've gotten too cynical and defensive over the years.
I actually had a need that only gimp could fulfill here on GOG: Falcon 4's patches needed to be a very obscure format. Now i got my Squirtle Squadron again.

Completely understandable. Think nothing of it. Text is such a crude form of communication.

Top 5 Image Upscaler Tools for Creatives in 2022
#1 - Stockphotos.com AI pixelated images(https://imageupscaler.com): High-Tech and User-Friendly.
#2 - Gigapixel AI by Topaz Labs: Ideal for Batch Upscaling.
#3 - AI Image Enlarger: Efficient and Affordable.
#4 - Bigjpg: Upscaling for Illustrations.
#5 - Deep Image: Great for Mobile Users.