Gyokuyoutama

If you are going to trace something, at least don't make it look obvious.

Aaaand it's finished!

 
I'm so so happy with how this turned out! I feel like I'm really getting better, looking at this. I think I have my sketching process to thank. It's very thorough, for better or for worse.

When you ask "what kind of touhou fanart do you want to see?" there are two types of people: The ones that say "mo' miji" and the ones that call you a "mo'ran"

Not been in the mood to draw recently, paying a bit more attention to my online classes, but I found this oldass drawing I did when I was like 14 or so? It was when I used to draw very often.
 
 
Some things never change, even 16 years later, I still love maid uniforms and android/cyborg vibes. I wish I could recapture her expression, I never manage to get that serious look when I try.
 
Also, I decided to make my own Minecraft skin for some reason. Decided to base it on a character from my old programing class project.
 
I think it turned out pretty good! Maybe I could make the gray parts a bit lighter, but I feel I managed to get the knight vibe into the skin. The "color noise" from the tool did help add texture to it.
 
And I also found this
 
What the fuck IS this? WHY did I make this? @Idiot Cube WHAT THE HELL IS THIS?

GOG is giving away Jazz Jackrabbit 2.

I was there since the dawn of time.
 
I will be there to put the chairs up and turn off the lights when its closing time.

I find it both pretty funny and interesting to see.
 
Newbies and people with no self-esteem go around doomsaying that the AI art will end their jobs.
Scummy people using it and trying to pass it as their art, usually editing the result images so people won't notice.
People getting mad about robot making better than they do after years of practice.
Others mocking how the AI can't make fine details or hands, even though they can barely shit a stick figure.
Some ignoring the work that took to program and train the AI and saying all the work is only done by repeating the same prompt a hundred times.
Ignorant folks accusing actual artists of using AI if their art style resembles anything they saw made by AI.
 
I actually enjoyed using DallE Mini and using the results of random prompts as a base on what to draw. I would try that Stable Diffusion one if my PC could handle it.
 
But basically:
 

I've got a friend who loves churning out AI art for D&D portraits. If he gives it a really long-winded description with lots of details, it can turn out several dozen images that more or less fit the concept, some of which don't have weird limbs or fucked up faces. But it works fine for a tiny image on Roll20 that just needs to passably resemble a dwarven archer or a light-themed human cleric.
 
Like Gyo said, the algorithm has to work with the data it's been trained on. And it has no idea how to make something new from two unrelated concepts. It not only lacks the creativity to imagine what Chef Kawasaki's supple, egg-shaped body would look like under that apron, it also lacks the intelligence to understand that it needs to do that.
 
The dirty secret of machine learning is that it's just letting the program fiddle with its own values, basically at random, until it produces desirable results. It doesn't know how it got the results we like, and adjusting any of the values manually probably won't result in any useful changes. There won't be a "Kawasaki cup size" value, for example.
 
Basically, I think the human artists don't have anything to worry about. Our brains, as slow and fallible as they are, can still do things that are simply unthinkable for a computer.

I think the AI art stuff is really neat. It's fun when you can submit any nonsense and get something back.
Seeing how the AI responds to the input is half the fun, there were similar shenanigans with AI Dungeon.
 
Trying to turn AI art into real money though becomes a really complicated issue really quickly and I'm not educated enough to speak on it.
I'm sure there are legitimate professional applications for AI art, I've heard backgrounds or concept sketches being cited as legitimate uses, but there are also people stealing art for training sets or trying to pass something off as their own.
 
Basically I think the technology is cool and people are the issue.
 
I don't see AI anything going away though. I'm now working at a tech company and a lot of the newsletters talk about new applications for machine learning.
They also talked about the future of the metaverse though, so maybe it's not a good metric haha.

Yeah I think there's no reason it won't be able to work out the current kinks, a lot of the things that people meme about (hands) aren't probably much more than a minor speedbump in the improvement process. As an enjoyer of fanart I think that at least part of the appeal is lost if there isn't someone behind the work itself, but at the same time if it's impossible to discern can I really say shit?
 
I don't really know how to feel about where it's headed tbh, like most things it'll probably end up being something lame and pervasive as hell and I really just want to bunker down and ignore it in its entirety

To put things another way, remember way back on bigSPUF when I made those 2hu comments using the walfas program?  I didn't draw a damn thing for those; I just used premade characters and models.  But I did compose shots, create dialogue, craft jokes, etc.  That's how I see AI being used down the line; it will be good at making the "elements" of your comic, but you'll still have to put them together.  Or if you want art of something very specific with many parts you'll have to take a lot of AI pictures and merge them together.  (I know some indie authors are already doing that to get cheap cover art.)

The reason that I take my position is the method used.  If something is not in the training set it can't be replicated by a machine learning AI, period.
 
There's a famous anecdote of a program that sorted pictures of dogs.  Given the initial pictures it was pretty accurate, but people noticed that every picture of a dog in the snow was classified as a husky, even if it was a poodle or a chihuahua or whatever.  The problem was that in the training set the only pictures of dogs in the snow were huskies.  Machine learning AI just looks for similarities to known categories.  It has no concept of, well anything really, but in particular nothing outside of the categories given to it.  In the example of the pictures with dogs it has no concept of "dog" vs. "snow" and so can't be told directly to just ignore the snow because it doesn't matter.  (This is also called the "frame problem" in AI.)
 
Now if we just want to classify pictures of dogs and nothing else we can fix the problem by including more pictures in the training set, making sure to include both huskies not in the snow and non-huskies in the snow.  The AI will never really have a concept of a breed of dog, but if we included a ginormous amount of data and make sure it is properly varied, then it can create an algorithm which is good enough for our purposes.  (But there will always be corner cases where it gets things spectacularly wrong in ways that no human ever would.)  We can't really manually update the algorithm to fix the issue because they are so complicated and bizarre, and it's not like there's a part of the code where the AI says "this is what I think a husky is."
 
Now let's apply these ideas to fanart.  The AI can only make variations of things which have appeared frequently in the fanart.  To give another example of how this can go wrong, there was a point in her stream where Pippa was trying to make art of Omaru Polka.  But every image came back as a loli, while Polka has adult proportions.  I can't be sure what is going on, but my very strong suspicion is that the "fennec fox" tag in the training data was dominated by Fennec Fox from Kemono Friends, and thus when the AI was told to make a Fennec Fox character it pushed it specifically to looking like the Kemono Friends Fennec Fox.  It doesn't have a concept of "give the character big ol' fennec ears and a tan colored fox tail."  The problem could be fixed by stuffing tons of varied fennec fox artwork into the algorithm, but there should be at least a little variation because of characters like Fenneko, Perfumer, Polka herself and some OCs (though I suspect that they might have removed furry characters like Fenneko to improve the output.)  But the Kemono Friends version so dominates the tag that you'd need a lot of extra fanart. Similarly "shark tail" is going to get you a character that looks like Gura Gawr, and if you try to make it different the shark tail is going to look all fucked up, because so much of the art with a shark tail is just art of Gura.
 
Don't get me wrong.  I'm not in the "AI art is crap and will never do anything."  It's already to the point where if I just want generic art of a character who isn't too unusual that it can easily fill my demands.  And it will get better from here.  But there are certain roadblocks in the way that the AI works that will prevent it from doing everything, no matter how much they refine it.  And that's without getting into really impossible feats, like having an AI that automatically generates a manga complete with panels, captions/speech bubbles and a coherent plot.  (On the other hand, AI probably will get to the point where you could use an AI to automatically create all the character and background art and then rearrange them into your own manga.  The most significant problem at the moment would be maintaining a consistent look across your OCs but that is a problem that I do think will be solved in a few years.)

To put things another way, remember way back on bigSPUF when I made those 2hu comments using the walfas program?  I didn't draw a damn thing for those; I just used premade characters and models.  But I did compose shots, create dialogue, craft jokes, etc.  That's how I see AI being used down the line; it will be good at making the "elements" of your comic, but you'll still have to put them together.  Or if you want art of something very specific with many parts you'll have to take a lot of AI pictures and merge them together.  (I know some indie authors are already doing that to get cheap cover art.)

The reason that I take my position is the method used.  If something is not in the training set it can't be replicated by a machine learning AI, period.
 
There's a famous anecdote of a program that sorted pictures of dogs.  Given the initial pictures it was pretty accurate, but people noticed that every picture of a dog in the snow was classified as a husky, even if it was a poodle or a chihuahua or whatever.  The problem was that in the training set the only pictures of dogs in the snow were huskies.  Machine learning AI just looks for similarities to known categories.  It has no concept of, well anything really, but in particular nothing outside of the categories given to it.  In the example of the pictures with dogs it has no concept of "dog" vs. "snow" and so can't be told directly to just ignore the snow because it doesn't matter.  (This is also called the "frame problem" in AI.)
 
Now if we just want to classify pictures of dogs and nothing else we can fix the problem by including more pictures in the training set, making sure to include both huskies not in the snow and non-huskies in the snow.  The AI will never really have a concept of a breed of dog, but if we included a ginormous amount of data and make sure it is properly varied, then it can create an algorithm which is good enough for our purposes.  (But there will always be corner cases where it gets things spectacularly wrong in ways that no human ever would.)  We can't really manually update the algorithm to fix the issue because they are so complicated and bizarre, and it's not like there's a part of the code where the AI says "this is what I think a husky is."
 
Now let's apply these ideas to fanart.  The AI can only make variations of things which have appeared frequently in the fanart.  To give another example of how this can go wrong, there was a point in her stream where Pippa was trying to make art of Omaru Polka.  But every image came back as a loli, while Polka has adult proportions.  I can't be sure what is going on, but my very strong suspicion is that the "fennec fox" tag in the training data was dominated by Fennec Fox from Kemono Friends, and thus when the AI was told to make a Fennec Fox character it pushed it specifically to looking like the Kemono Friends Fennec Fox.  It doesn't have a concept of "give the character big ol' fennec ears and a tan colored fox tail."  The problem could be fixed by stuffing tons of varied fennec fox artwork into the algorithm, but there should be at least a little variation because of characters like Fenneko, Perfumer, Polka herself and some OCs (though I suspect that they might have removed furry characters like Fenneko to improve the output.)  But the Kemono Friends version so dominates the tag that you'd need a lot of extra fanart. Similarly "shark tail" is going to get you a character that looks like Gura Gawr, and if you try to make it different the shark tail is going to look all fucked up, because so much of the art with a shark tail is just art of Gura.
 
Don't get me wrong.  I'm not in the "AI art is crap and will never do anything."  It's already to the point where if I just want generic art of a character who isn't too unusual that it can easily fill my demands.  And it will get better from here.  But there are certain roadblocks in the way that the AI works that will prevent it from doing everything, no matter how much they refine it.  And that's without getting into really impossible feats, like having an AI that automatically generates a manga complete with panels, captions/speech bubbles and a coherent plot.  (On the other hand, AI probably will get to the point where you could use an AI to automatically create all the character and background art and then rearrange them into your own manga.  The most significant problem at the moment would be maintaining a consistent look across your OCs but that is a problem that I do think will be solved in a few years.)

Nice little rabbit hole youtube channel if you are interested in that sort of thing:
 
https://www.youtube.com/channel/UChjOucyoXowoqMJ6L_hs3NA/videos
 
Uploads videos with names of messes of characters, several times an hour, and has been doing so for three months.  Uploads vary in length from a few seconds to over an hour.
 
Mainly looks to be podcasts.  Sometimes the URL for an archive site and the name of the podcast is in the title.  Titles always end in a date shortly before the upload date; this might be the time that the podcast was retrieved.  Some of it sounds more like radio, including political talk radio and music, though it is possible that this was retrieved from an online upload of a show or an internet stream of the content.  Very occasionally you will get static, which almost sounds like it came from an actual broadcast.  But I suspect that in those cases what was recorded was an internet stream of some guy's radio rather than the radio directly.  Everything else just has too high of sound quality and things almost always start right at the beginning of the program, which wouldn't happen if it was recording off the air.
 
Doesn't seem to be any rhyme or reason as to what is put on there.  I haven't even seen the same show uploaded twice, though obviously there's a lot of uploads and I'm not going to look through all of them.  Some of its in English, some is in other languages (French seems to be the second most common.)  Some of it is nerd podcasts, some Christian inspiration type stuff, some music (both from online and recorded off radio stations), some talk radio, some sports, etc.
 
So it's something made by an AI pulling content from a lot of online audio sources.  The rabbit hole comes from if you feel like answering the following questions:
Most obviously, why?  Usually the answer is some sort of content farm but these videos have been getting next to no views; the most popular video on the channel is still under 500 views. Is there any meaning to the titles, or are they randomly generated?  Certainly you can see a date at the end, and often a URL and sometimes a title.  But there's a number at the front (which is usually similar to previous numbers but does not increase sequentially from video to video) and sometimes a whole mess of letters and numbers (which seem to be hex, since you never see anything in the alphabet past f.) Is there any reasoning behind the sources?  That is, are they all available through a common pirate site or something?  Or is the AI just crawling the web and posting whatever sound files it can find?

Evolution of a trend.
 
Got sent this video, which got me thinking about its antecedents.
 
 
I'm fairly certain that the immediate reference is to a series of Kizuna AI videos, since these are vtubers.  The earliest I could find was this one (going back three years):
 
 
But this came from a trend of MMD videos, particularly involving Touhou characters, for example this one (four years further in the past):
 
 
The earliest variation of this I could find was this one (two years earlier, though it could have been before that on NND):
 
 
This in turn was a reference to the following video (from two years earlier):
 
 
But that's still 11 years after the release of the original Backstreet Boys song so it's very possible that this video was a riff on something else.  Unfortunately at this point we get into the pre-video streaming days so any viral videos that might be being referenced likely only exist on various personal hard drives or in a low quality format on some forgotten personal homepage.

I had this idea of trying to find old toys that I owned to display them in my room.
 
It took me some time, but I finally learned that the one toy that I REALLY liked was called "Thunder Gridman, Synchro Samurizing Servo".
 
I found one kinda cheap and is 99% complete. Was about to get it.....but it doesnt ships to South America. Fucken hell.

Hurricane: survived
Power never even went out. No flooding, no trees on the roof, just the usual shit ton of debris everywhere.

The Space Quest games were not only Sierra style games where you could easily die, but they were pretty merciless even for Sierra games.  For example, there's a part in Space Quest I where you need to make enough money to buy a new ship.  While you can get some starting cash from a few places, the only way to extend that to enough is to play a slot machine.  Now you can screw yourself in the usual way, i.e. by losing all of your money and not having any way to make more.  However some times you get a bad enough result that the slot machine straight up kills you.  It's completely random; there's no way you can prepare for it or avoid it.  (Though the game is nice enough to let you know it's a possibility by killing the guy playing the slots before you.)
 
The funny thing is that being this level of dick actually makes those games feel more fair, in a way.  It's almost impossible to play a Space Quest game for long and not realize that you are expected to save constantly, under different file names.  In contrast Kings Quest II is quite content to let you get through about two thirds of the game without letting you know that you fucked yourself over the first five minutes by walking over that bridge twice.  Space Quest is kind of like kaizo platformers; in no way is is it fair, but you knew what you were getting into.
 
Anyway, I had no idea that the guys from Andromeda were making another game.  I'll keep an eye on it.  That being said it seems like adventure kickstarters always go wrong, so I'm not exactly having high expectations.  But maybe just taking a decade is enough going wrong that the game will be fine.

Bit of a blog post here, but "Two Guys SpaceVenture" is finally releasing out of Kickstarter today!
 
The project was originally kickstarted back in June of 2012 as a successor to the Space Quest games.
Admittedly I never played the Space Quest games, but I have a fondness for point and click games.
Although the Space Quest games were Sierra, whose catalog were infamous for killing you over every little mistake!
 
I originally received the news through the SPUF forums because none other than John Patrick Lowrie, voice of the Sniper, is featured as one of the voice actors for the game.
At the time, I backed SpaceVenture to the tune of thirty dollars and waited.
This was also around the time that projects such as Wasteland 2, Pillars of Eternity, Doublefine Adventure, and the Shadowrun games were all getting funded.
All of those came out and most of them got sequels in the time since SpaceVenture was funded.
 
So here we are more than ten years later and it's finally ready to launch later today.
They've been releasing backer updates during all this time, but SpaceVenture still qualifies for the hall of "Games I waited a decade for."
 
SPUF might be gone now, but I originally only found SpaceVenture through SPUF, so it's kind of a strange feeling.

I already miss Marko.

I mean it is the most memed Hellboy scene, so not exactly a hard guess.
 
It's pretty funny how early into the series it shows up though.
 
"Here. Now you know what you're getting yourself into."

So I've been fiddling with Civilization 2 for a while.  I mean, really I've been doing a lot of retro gaming for a variety of reasons, but that one always hits me hard for some reason.  I particularly like making my own scenarios and messing around in them, even if they aren't as polished as the published scenarios.  The Fantastic Worlds expansion gives in game editors that are quite helpful in setting up new units, changing technologies, changing improvements, and doing some minor other fiddling (though there's a memory overflow error that sometimes happens when using them, watch out!)  But you'll quickly find that other scenarios have tweaks that you can't do in game.  The most obvious is the civlopedia, i.e. the list of all the units, techs, etc.  When you make a scenario it will go to the default one (which is probably irrelevant) but the published scenarios have a simplified version of it which accurately reflects the new units, techs, and so on.
 
I've done a lot of reverse engineering and figured out a bunch of this on my own.  A big thing is that Civ 2 stores all text used in the game in various text files, and if you have one of these in your scenario folder it will use that file instead of the default.  If you replace the civlopedia text it will even create a new menu for the information.  By copying and pasting the modified versions from another scenario you can get a pretty good idea pretty quickly on how to handle things.  Similarly there are a lot of graphics files which you can't change using the in game editors (ex. the barbarian leader unit, the pops faces in cities) but these can be changed by copying the relevant .bmp files into your folder and then editing them.
 
But you'll find that there are some things done in scenarios which are not as easy to crack.  Most notably, some scenarios (like the Midgard scenario) have either partially or wholly different tech trees for different factions.  There's no way to set techs as just for one tribe in the rules text files, and looking through the text files for the relevant scenarios shows no obvious differences.
 
Eventually I gave up and started looking for resources explaining this.  I wasn't expecting to find anything, but came across this book on archive.org:
 
https://archive.org/details/civ-ii-scenario-guide/page/244/mode/2up
 
It's a book written by the scenario designers where they go into great detail about how exactly the Civ2 rules work, what you can change and what is hard coded into the game, what will cause mysterious errors, how to do various tricks (such as having "exclusive" units; turns out this is done with some shenanigans involving dummy techs that the computer sees as inaccessible but which lead to usable techs), comments on each specific published scenario, and just general advice for how to design and polish scenarios.  They even go into how you can convert scenarios made in later versions of the game to be compatible with earlier versions and they discuss how viable each scenario is for multiplayer, even though they were designed when there was no multiplayer for Civilization 2.
 
It's kind of insane that something like this exists, but really this sort of thing was more common before 2005 or so.  I mean, manuals for older games are often hundreds of pages to begin with, so fan supplements being that long isn't so weird.  But then they started doing less and less documentation.  Now we are in the age of "it's probably on a wiki" or "check out a discord" even though you're not likely to find information even equivalent to a tenth of this book.
 
EDIT: Looks like some of the pages are scattered around in this scan, which is too bad.

I actually found that extremely funny. They made this big post full of buzzwords like saying "Switch emulation is a long existing problem", when the emulation is still very barebones, only possible on top grade hardware, only a few games can be played without glitches and only focused on a few first party games.
 
Like, are companies REALLY going to butcher their Switch ports performance even more so people that wouldn't pay them anyway can't emulate a game?
 
I say they TOTALLY should implement the DRM so legit costumers can suffer even further.