How do you feel about AI upscaling?
bad. i haven't done much in the way of research on the current meta of "AI" upscaling technology so maybe there's some magical new trick that's suddenly made it work good, but i doubt it. what i can speak to is practical use cases i've found in the wild.
on the amateur/consumer end, upscaling gets used to transform low resolution media into high resolution media. you see this a lot with documentary footage from the turn of the 20th century, where horsedrawn carriages are romping down the dirt roads of new york city in 4k60fps and colorized from black and white. like most things "AI"-related, this has a window of genuinely impressive novelty that lasts for about twenty seconds, maybe up to a minute if the result is decent. wow, it's like i'm looking through a window into the past! pretty quickly though you start noticing the smear frames, the interpolated faces, the hallucinated artifacts. watching these upscales for any real length of time, for me at least, rapidly approaches nausea-inducing. most information present in these videos is artificial, which is basically fine except that they're sold as "more realistic" or "more immersive," preferable to the inferior original footage. but the only true information is that which was present in the "inferior" original. with the amount of work it takes to manually correct these errors to maximize verisimilitude to the original, you might as well just do a 4k scan of an original print and go through the traditional remaster process.
you see this as well on torrent sites, where almost any sufficiently popular new anime or tv show will have an upscaled ultra-high framerate version that looks like ass and is really easy to get confused for the genuine article if you don't know what to look for. those files USUALLY get buried by real-quality torrents and don't seem to be particularly popular (most people usually just feel tricked). which is in itself a pretty common theme among mainstream deployments of "AI": its ubiquity is equaled only by its unpopularity. nobody asked for it, it's not solving many problems better than the solutions that already existed, the output is deceitful crap, and yet it is being foisted upon us from every direction with astonishing confidence that the technology is something it just emphatically is not.
then there are the more professional implementations. the biggest one for me, long before the "AI" moment we're in now, was Peter Jackson's 2018 WWI documentary "They Shall Not Grow Old." this film takes digitized war reels and upscales/colorizes them through various machine learning technologies to make the footage "feel more present" and help modern viewers connect with the past in a way they hadn't been able to before. i saw that one in theaters and mostly felt that it was awkward and disrespectful, featuring many of the telltale signs of "AI" video output that we've come to know and love today. the outgrowth of this technology has been its use in 4k "remasters" of films like Terminator 2 and Aliens, which forego the costly process of an all-new scan using state-of-the-art hardware and instead seem to just upscale the original blu-ray transfers. the results look like ass, oversharpened edges turning every face into tanned leather, background characters gaining phantom details not present in the original, film grain being traded for a thick gloss of vaseline.
that seems to be a big priority, actually, the removal of film grain. of anything, really, that differentiates celluloid from digital. there's a nasty and pervasive attitude among top brass losers that there is an Objectively Good way for movies to look and an Objectively Bad way for movies to look, and what makes the difference has basically nothing to do with the contents of the media itself and everything to do with raw numbers and marketable buzzwords. they want crisp clean 4k 10bit rec.2100 HDR footage, and they want it as quickly and cheaply as possible. it's the exact same kind of surface-level tech fetishism that gave us motion smoothing and frame interpolation as auto-on settings in consumer televisions. they want something that looks good on a shelf or in a catalog, something flashy and new with big numbers that sells units or subscriptions. they want automated processes to "clean up" messy footage without the hassle of having to hire a studio of (unionized) people to spend hundreds of hours doing the same job by hand. which, again, could be fine, and indeed plenty of professional film restoration outfits DO use machine learning tools to simplify the process, have even innovated them in some cases-- but their outputs are highly limited and extensively reviewed by human beings, something that just does not seem to be the case with James Cameron's sloppy "AI" "remasters."
and it just doesn't take very long at all to dig up James Cameron and Peter Jackson openly discussing their desire and even NEED to build these automated processes, not because of what it brings to the art of cinema but because of the money it saves by cutting out (unionized) labor. i know there's a contingent of brave warriors on this site insisting that automation is automation and we all just have to get with the program, but there IS a difference between a machine that effectively automates a tedious task with no apparent loss in quality, and a machine that exists to cut corners without any regard for quality whatsoever. there's machine-learning-driven automations in cinema that are generally good, like the blue eye replacement tech in Dune 2 or even the mob population & behavior tech built for Jackson's own Lord of the Rings movies. there are certainly arguments to be made in some of the particulars of how those technologies originated, but i don't think there are any vfx workers banging down the doors pissed off that they don't have to manually add a blue filter on the pupil of every single background extra or individually animate the arm movements of a thousand-strong army in a 3 hour effects heavy blockbuster anymore.
and that's just it. i put "AI" in scarequotes because it's a marketing gimmick meant to lump dozens of unique, often totally unrelated technologies into a single basket to trick people into thinking it's all the same thing. it takes a tool with a specific function and transforms it into magic. all algorithms are grimoires now. the prevalence of "this"AI" as a buzzword is a result of how thoroughly controlled all the pillars of our economy are by professional money-havers that don't use the products they own and have nothing but resentment for the customers who do. the top-down flood of "AI" integration has done about as much to harm the public view of the technology as the shit output of the technology itself. like, for instance, video editing software Davinci Resolve recently introduced a bunch of "AI" tools... most of which are just further polished versions of machine-learning effects already present in prior versions or other more expensive software suites. there IS generative "AI" stuff in there, but that's maybe three of the dozens of features they added. but because "AI" is the buzzword that marketers want to see, all of these effects have been dubbed "AI"-- which really only accomplishes the obfuscation of how these effects actually work. it takes a relatively simple or at least straight-forward algorithmic process and arbitrarily encloses it in a black box until it's impossible to tell which ones are generative and which ones aren't without citing outside sources. which sucks! i like resolve! its "AI" dialog separator is actually quite good for if you're recording in a relatively noisy environment, and that's just the latest iteration of a machine learning application that's been the bread and butter of this kind of program for over a decade! it's not "AI" god damn it!
the confusion is intentional. our rhetorical inability to distinguish between generative video/text and basic programmatic task automation gives cover for all the crap that doesn't do what it's supposed to. it lets them point to successful use cases of "AI" as an indication of the tech's broader revolutionary potential, even though they're just plain not the same tech. it's like advertising a hammer to sell a screwdriver, except the screwdriver is made of recycled plastic and has a nonstandard head that doesn't fit any known screw type. put another way: it's a lie. they're lying. because they're liars.
"AI" upscaling fits right in here. it offers the apparent ability to qualitatively improve lossy footage at remarkably small cost (money, time, effort), but a survey of the results proves that actually there IS a cost: the cost of laziness. when you cut corners, half-ass the process, make fewer people do more work and rush them to meet an insane deadline, "AI" or no "AI" the end result is gonna be crap. the simple reality is, if you want a good job done well, you will need people who know what they're doing to do it. a good job done well takes time, effort, and money. it doesn't matter how much machine learning you throw at it, how many aspects of the task are automated; at the end of the day, the work of film restoration is going frame by frame to remove imperfections without altering the character or original intentions of the work. it is impossible to automate any sizeable portion of that task without the use of a machine that also has a theory of mind, a concept of historicity, an extensive comprehension of restoration best practices, a deep knowledge of film history and the reference materials required to confirm the processes of deceased filmmakers/deprecated technologies, and of course an understanding of why restoration is important and why it is done the way it is done and why doing it other ways is worse and often actively harmful.
that's the goal that "AI" CEOs insist they're working towards, but i simply do not believe that we're anywhere close to cracking the AGI problem. i don't think it's possible with existing technology, and quantum computing has so far proven to be vaporware. the thing i've found myself thinking a lot these days is: just because we can imagine a thing, that doesn't mean it can actually be built. fiction (from which so many big tech ideas are drawn) is not bound by the laws of physics. by the same token, i think we're being really naive about the social gullibility of even very intelligent people. just because the stochastic parrot tells us what we want to hear doesn't make it any less of a parrot. there is an astonishing refusal to acknowledge how easily people, in general, are fooled by chatbots, something we've known about since the invention of chatbots, instead acting like being successfully deceived proves that it was never a deception in the first place. these are the tactics of a con artist. it's magical thinking, and unfortunately for these guys magic doesn't exist. the tasks that CAN be effectively automated by machine learning aren't worth even the remotest fraction of the tech's valuation. everything else is laziness being marketed as efficiency. the difference is self-evident in the results.
so yeah, in short i'm not a fan of "AI" upscaling, at least when it's being applied carelessly and cynically.