Before I continue, I wanna make a tiny follow-up paragraph to my last Touhou Project related post, outlining a few things:
We've been a good few years into this "new age" of "A.I.", "deep learning", "large language" and "generative" models, with the promise that this would open up a whole new world of opportunities in artistry and technical work. On the technical side, there is some truth to this. I'll be the first to admit that these big LLMs helped out quite a bit in the implementation step of my most recent Soul Calibur reverse engineering work, but little if at all with the analysis steps that came first. They did help with scripting some mundane work in more famous environments like Blender Python, but with anything obscure I did it better all on my own. Relying on them will by no means make you an expert coder, no matter how many quacks and middle managers (what's the difference, really) will try to convince you otherwise.
All these LLMs are founded on the principle of the aforementioned deep learning. The actual science and engineering behind it is far too complex to get into here, but what matters is that one of the most key ingredients to making it work is already existing data. Which data ends up being used for the "training" process depends on what the desired output is. As an example, developer oriented LLMs will most certainly base their knowledge on educational material for development, as well as actual code from around the Internet. This has... a whole basket of ethical and legal implications, considering that taking others' code verbatim with no attribution isn't exactly viewed in the best light (Huh... now where have I seen this before?), especially with the added muddiness thanks to how LLMs form their outputs, making it difficult to pinpoint which code is taken and which, if any, license terms were violated, which isn't a guarantee in the case of very permissive licenses.
You might argue that this is a rat race to the bottom that amounts to, done ignorantly or not, back-and-forth theft of code between low skill developers. My opinion is that this really only accelerated the process of the development quality drop. Developers have and always will take from others, for both simple and complex tasks, because that's just what developers have been doing for years, although most of them will respect a given piece of software's license terms if they are actually defined.
However, can we really say the same for the creative flip-side? Generative image models work on pretty much the exact same principles, except this time the input data is images, which can be photos, paintings, sketches, you name it, they all boil down to being creative works. This is where we enter a major divide, which from my observations of a few recent events (and just general cultural sentiment) points to it being mostly generational. To cut to the chase, younger Millennials and further younger demographics appear to be overwhelmingly negative towards generative models, no matter the use case, because of the fact that they rely on the works of other authors with zero licensing. Gen X and older seem to be at the very least mildly positive and see it as the next stepping stone to potential greatness. Others see it either as a tool to be used when appropriate... or as the second coming of Computer Christ. Please note that I don't have any concrete numbers to support this in the overall scope of society, these are just based on observations of the various environments I'm in. On the other hand, if I narrow it down to a select few groups, then we can all observe a few patterns. One such group that I hold rather dearly is the demoscene, which for decades has been defined by it's environment that encourages a combination of artistic and technical prowess, and holding many related competitions in various categories every year. In spite of the fact that these competitions are held by and for computer nerds, the artistic and competitive aspects are highly valued in the scene.
So what happens when we introduce generative models into the mix, and how far can we go before all hell breaks loose? Enter Skywards, a mediocre, uninteresting and poorly directed Amiga collage demo that didn't even come close to winning any awards at Revision 2025's Amiga compo, although SoDa7 did produce a pretty decent backing track for it. Under normal circumstances, everyone would've forgotten about it's existence the day after the compo finished... instead it very much remains in the spotlight, just by the fact that nearly all of the images in the collage have been created with generated models, and then simply color crushed and downsampled for AGA platforms.
You can tell from the upload on Pouet above that this struck a nerve with a lot of people, me included. It's one thing for a demo to be mediocre, it's something else entirely when it's barely even man made. At some point we might as well make computers battle it out between themselves, no human input in the process whatsoever, because why even compete if the competitors are themselves artificial. "Oh but it's just the graphics, the coder and musician still did their part as usual", you might say. Alright, I'm not gonna take that away, I even praised the track just a few sentences back..... what makes you think it's only full demos that could be like this? Would it surprise you that shoving generative model image outputs as entries has been attempted before in compos that exclusively focus on graphics? Yeah, thought so. Oh would you look at that, Mr. Critikill has worked on both of these prods? I smell a pattern.
I know a good number of names seen in the comments of both pages, so I know who's been around for longer and who hasn't. The theoretical groupings I mentioned earlier are absolutely a reality here. You'll see a whole bunch of all or nothing opposition towards these technologies, counter-balanced by those who treat it as an evolutionary step. There's more examples by the way, don't you worry one bit. As for why this is the case, I wouldn't really be able to tell you. Maybe the greybeards treat it the same as when computers were finally becoming viable for digital illustrations and image editing in the late 80s.
Alright, I'll go over semi-related stuff for the Touhou Project crowd, cause that's the only thing anybody ever wants to read about here these days.
If you're in this scene, you certainly already know of or even played the demo of the 20th mainline entry. Just gonna put in on the record, I didn't. I only ever launched th17 exactly once for a test and never again, and haven't even bothered to download any further releases apart from 19 to see just how big of a mess it was. Unfortunately, this latest entry is not exactly easy to ignore considering just how much hub-bub there's been about it this past week. It all stems from what was at first speculation, but later a confirmed fact that several graphical elements in the game were created using generative image models. My honest initial reaction was "... why is anyone even surprised about this?".
I'm serious, how is this newsworthy? If it didn't happen before it was gonna happen sooner or later. Well, apparently people find this shocking because ZUN supposedly held a dislike towards generative models only a year or so ago, only to twist this around with his latest release. I tried looking into this deeper, but I could not find any statement coming out of his mouth to support these claims. The closest to anything he said about them prior was in the Th19 omake, which provides the usual flavor of non-answer wank to all this, except for one part where he says that it's the best choice for those who value pure "efficiency".... Well yea, no shit.
I genuinely need to ask again, why is anybody surprised about this? ZUN has been nabbing things left and right from the very beginning. Just because he's doing it in a way that a large amount of people currently find reprehensible doesn't change the fact that he's really just doing what he's always done best. Apart from that there's really nothing to add to the conversation, but somehow this has been enough to reignite it and give me a pretty noticable boost in incoming traffic to the site... for some reason.
The following is what I actually wanted to focus on: just like I outlined before with the demoscene, nearly all of the negative sentiment has been coming from younger demographics, while older ones seem to be, at worst, neutral. While I would love to pinpoint the reason why this is the case, I'm unfortunately not all-knowing. Maybe it really is as simple as kids being passionate and artsy, and old men being jaded and "efficiency" oriented.
Alright, I'm done here. This one has been a bit anemic, I'm sorry for that, but that's what I get when I write while being dead tired. If you think I should add some more stuff, you know where to reach out. Now piss off and good night.