Computer created text and images has obviously been big in the news over the past year and change, largely because of the release of ChatGPT and DALL-E (and the various iterations of them). Most of those types of implementations are essentially party tricks. It is pretty cool to type out "Give me a picture of _________" and a few seconds later, you have an image on your screen that roughly matches that description. Usually, though, the text and images that are generated have some inherent flaws when you look a little past the surface. Images of people, for example, frequently show distorted faces and/or hands.
I ran a test last January and asked ChatGPT to write me a biography of Paul Sampliner, hoping it might be able to pull in some data sources that I hadn't found had access to, but it gave me a thousand word bio and every statement about Sampliner was factually incorrect. Literally every single statement about him was wrong.
So it's a neat trick to see the computer generate something right in front of you, but it's practical usefulness can be a bit limited.
However, if you limit what you ask these AI systems to do and provide a distinct set of data to work from, they can be useful. I've seen reports of scientists having these systems run through variations of vaccine cocktails to narrow down the ones that have the strongest chance of being effective. They still need to be tested and studied, of course, but now they're studying 100 possible combinations instead of a million. So it you ask a computer to just extrapolate based on a sample set of known patterns, you can get to something more practical.
And that's where my AI experiment today comes from.
When Hasbro released their Moon Knight action figure at the tail end of 2022, I was excited because that's a favorite character of mine. But beyond that, while there have been many Moon Knight action figures over the years, I felt none of them translated very well. Frequently because toy companies simply took a generic figure and painted a crescent moon on his chest and that was about it. Further, the white plastic they always used was highly reflective and never seemed to jive with the night prowling hero; how could you sneak up on bad guys if your suit is the only thing visible? But this new version took its cues from the show, and gave him more of a wrapped-in-dull-bandages-like-a-mummy look which I thought was exceptionally clever.
However, I didn't especially care for the Khonshu figure they released. The design is also based on the show, which is fine, but I prefer the comics version which had more of a featureless face surrounded by a nemes, that striped head cloth often worn by Egyptian pharaohs. In the comics -- the ones I grew up with -- Khonshu never really appeared directly; he simply spoke telepathically to Marc, often using a life-size statue of himself as a kind of conduit. And that's what I wanted a figure of to accompany my Moon Knight action figure: a statue of Khonshu as he was shown in the comics.
If I had a 3D file of him, I could print one on my 3D printer, but everyone who was producing Moon Knight related designs has been focused on the show. And since that design came from artist Bill Sienkiewicz, who was actively trying to make something for
comics, it wasn't particularly accurate from a historical perspective so I couldn't use anything based on actual Egyptian works either.
(As an aside, Khonshu in ancient Egypt was frequently depicted with the head of a falcon. So the Marvel Cinematic Universe depiction of the figure with a bird skull for a head is closer to historically accurate than what's in the comics. But I saw Sienkiewicz's version first, and it was the only version I saw for decades so I still prefer that.)
So I'm left with trying to make my own. My modeling skills are pretty rudimentary, though, and sculpting a figure would be a huge task, particularly for what I want to be able to do. I looked briefly at modifying a Hero Forge figure, and I got something I was reasonably happy with, but their figures are scaled to be no more than an inch or two high. The proportions would look wildly out of whack and the details would be crude if I tried simply printing that figure at 600% or so.
There was a Marvel Select action figure back in 2006 that included a Khonshu statue of the type I was interested in as an accessory. Simply purchasing one posed a few problems, though. First, the accessory by itself typically seems to cost $60-$70 US while a complete set (of which I didn't even want the actual Moon Knight figure) was often more than $100 US. Second, it's hollow; they only sculpted the front and the back is just left completely blank and open, which means you can only use it from directly head-on. But here's where the AI comes in.
I uploaded three photos of the toy into an app called Kaedim. Not particularly high resolution or anything; they were literally three images I found online, each shot from a different angle. The software sat and processed those images for a couple hours. (I left to run errands and came back later, so I don't know exactly how long it took.) What it presented me with was a 3D rendering of the toy, as it interpreted it from those three photos...
But beyond just rendering the basic object in 3D, it's tried to figure out what would make sense for the parts that aren't shown in the photos. It's created and entire back that indeed continues the nemes, collar piece, and cloak. What's more, it continued on stuff you can't even see and wouldn't even be relevant for what I'm trying to do -- namely, the figure's entire torso is rendered under/behind the cape as is his right arm! It even figured out the fingers on the closed hand, despite the images not showing that very clearly at all.
It rendered what it thought the figure should look like in three dimensions based on the limited data set of a few photos.
Now, this is just a 'generic' 3D rendering, and would absolutely not print well directly. But I was able to take the file into Tinkercad (a free, online CAD program) and fill out some of the details a bit to make this more printer-friendly. I also created a staff with a few simple cylinders and ran the two files through my printer...
The final print is still a little rough in places -- I had some issues with the printer itself midway though -- and there's probably a few ways I could streamline the file a bit more, but I find this to be an extremely functional and useful way to utilize AI. Again, it's not trying to generate something from "nothing" (which no AI program does anyway -- that's part of the ethical debate about what's being fed in as samples to begin with) but it was able to process along a finite, not really creative task of filling out a model based on incomplete data.
Theoretically, I could do this with other objects, but I cannot think of any other example I'd want to try at the moment. But it's definitely an idea to keep in the back of one's pocket, I think. If you're interested in the files I ended up printing from, I do have them
uploaded to Thingiverse here.
0 comments:
Post a Comment