Hi. You're reading this message because you've probably asked about something involving AI upscaling. As you can tell, this happens a lot. I'm tired of explaining so please read this copy-pasted prepared response.
So, you have a low-resolution video and you want to make it 1080p (or more, I don't care). Well, here's how to do that.
Step 1: Open your video in your favorite media player
Step 2: Press the full-screen button
Congratulations! Your video is now 1080p! (Or whatever resolution your monitor has)
But wait, you say! That's not what you mean! You want to u p s c a l e your video, not just make it bigger. And you've heard how A I™️ is so great at that and can upscale video to high resolutions! Well, I'm making a deliberate point with this. Upscaling is upscaling. It's taking a low-resolution image and somehow turning it into a higher-resolution one by trying to guess at what values the pixels in between should have. It doesn't matter if it's your video player or some AI model doing it. Upscaling methods fundamentally only have the low-resolution image to work with. No AI can magically recover details no matter how much it claims to. All upscaling will always be a trade-off between different artifacts like blurring, ringing, aliasing, warping, invented details, etc.
Unfortunately, the world of A I™️ is full of buzzwords and misleading promises and the belief that A I™️ is some magic black-box that can fix everything for you if you just press the right buttons. Well, that's bullshit. Repeat with me:
When people look for a "nice upscale" of some material, here's what they really want: They want something that looks "sharp" and "clean". And if you want that, I can't stop you. But you should know that being "sharp" and "clean" has nothing to do with being accurate and true to the source. The "clean" look of many AI upscales is achieved by wiping (intentionally added) grain and texture. That's not "clean", that's flat and lifeless. The "sharpness" is only achievable by creating haloing or warping lines, and many AI models will accidentally sharpen content that's not intended to be sharp.
Now, please don't get me wrong. I'm not saying that AI-based methods are inherently bad. There do exist very good AI-based upscaling filters that can perform very well when used correctly. One of them is waifu2x (with the cunet model), which is the current gold standard for rescales in the anime encoding scene. But it's this belief that anything AI-based is automatically a magic wand that improves everything it touches is what I want to take apart.
Now, depending on who's reading this there's a small chance that you just went "Oh, I have to try this out!", followed by "But this waifu2x model sucks! It's still so blurry and doesn't do anything to improve the video." If that was you, then you've missed the point. (If not, then I'm proud of you.) Models like w2x are great because they are very careful and don't do any excessive sharpening, warping, or cleaning. It's just a reality of upscaling that none of these things should happen to the video.
"But now I'm confused," you say. "You said that waifu2x is a good upscaling model, so why aren't you using it to upscale all your 480p videos to 1080p?" Now, this is a valid question. And sure, you can do this. If you want, you can run every 480p video through waifu2x before you watch it. There's nothing wrong with that. It's just that once we're talking about storing or distributing such upscales, that you have to ask yourself whether this is really worth it. The effect of waifu2x is an improvement, but it's not a gigantic one. Upscaling all of your videos using waifu2x would take a lot of processing time, and recompressing it in some video format will create additional compression noise (unless you crank up the bitrate, which needs more file size, which is another downside). And even if you did that and upscaled all your videos to 1080p, what if later on you want to watch it in 4k? Or what if the next new and fancy AI model comes along that performs slightly better than waifu2x? Would you upscale everything again?
So, in the end, most people decide that it's just not worth it. Using good real-time upscalers in your media player (e.g. mpv with --profile=high-quality
and antiringing, possibly even with a shader like nnedi3 or artcnn if your device is fast enough) already gets you very far and waifu2x is not enough of an improvement over that to warrant running everything through it. And that's why we're not doing it either.