Dearest Rachel –
Even the most inspirational saying can be rendered ridiculous with the wrong (or, if you’re a comedian, the right) preface.
Life is like a sewer; what you get out of it depends on what you put into itTom Lehrer
Such was what’s been going through my mind as I’ve been experimenting with this whole topic of artificial intelligence-created artwork, and the program that is supposed to accomplish it. For all the promise it holds (and I assure you, it holds much promise), it’s not the easiest thing to put together, or, once it is assembled, to give instructions to. Once I get the hang of this, I may yet turn you into a masterpiece (or rather, an entire collection of masterpieces), but until then, there are going to be a lot of fits and starts along the way. It’s going to take a lot of patience, but at least the results are faster in coming (and infinitely superior) than if I tried to create any of these on my own.
I probably should explain why I’m going to all this trouble of setting up a program when there’s a website (in fact, by now, there are probably a number of websites thus set up) that will create all manner of illustrations and fauxtographs (hey, I don’t know if that’s a word yet, but given the circumstances, it ought to be) of you, and it does an impressive job of it without my needing to go further into the weeds. All of which is quite true, but for once, I’m dealing with a technology that I just want to understand better, and that should be explanation enough. Just like you used to work logic puzzles and sudoku games to maintain your own mind’s plasticity (especially since you wanted to stave off your mother’s fate of dementia – granted, you managed to avoid it, but not the way you or I would have wanted), so I need to be challenging my mind from time to time, and I’ve finally found something that’s intensely interesting. That it’s on the cutting edge of technology, with new things being discovered and released it seems multiple times a week only adds to the appeal and fascination.
But beyond that, there are practical limitations to the websites, some of which I’ve mentioned before. They require a number of photos; some as few as three, others rather more (the site I’ve been using recommends a minimum of twenty), but they generally top out at no more than thirty. Indeed, I assumed it was a limitation of the program itself until I dug a little deeper into it all. Now, thirty would be okay, but after a few dozen iterations, the poses generated clearly repeat themselves a lot because of how few images the system has to work with. Better to create a checkpoint file out of a hundred or more, so as to give it more to work with in terms of poses and expressions. But how? Well, that’s what I’ve been trying to discover.
Additionally, the websites don’t seem to have a proper handle on turning a photograph (or a collection of photographs) into a manga or comic book character. You look a little too realistic to fit into the world of this or that creator (or, in some cases, naming a manga artist will only result in a portrait akin to an ukiyo-e illustration – impressive, but nothing like what I’m aiming for). I haven’t yet gotten far enough into the system to remedy this (or perhaps I have, and simply haven’t had the opportunity to prove it), but I think my research will lead me to a better solution – and it will still be faster than commissioning artists to come up with something approximating what I’d like to see, because I can keep tweaking the results to get closer to a vision that matches mine.
With that being said, this is a walk-through of what I’ve gotten out of the program thus far. Keep in mind, I was only generating images for maybe two hours at the end of the day yesterday. Prior to that, I was doing some accounting work for the end of the month for camp – I knew that if I started the day generating images, I might not get back to doing any actual work for others (although, in hindsight, since each series of pictures took the computer several minutes to generate, I might have used that down time to do the work. Still, best to have complete focus on the task at hand, rather than have a divided attention possibly cause mistakes – some of what I had to unravel was surprisingly complicated. But that’s neither here nor there).
At this point, you might ask about the delay; it’s the last hour or two on the last workday of the week, and I’ve waited until now to start on this project I’ve been talking about for weeks? Yes, well… the process is more complicated than that. As it so happens, when creating a checkpoint file of pictures of you, I need to train the computer to know what, specifically to look for in terms of facial features, body proportions, and all that. Each photograph has to go through a series of checks, and naturally, the more photographs, the longer this process takes. Basically, this took up the entirety of Thursday – although while the offsite computer was reviewing them (I was using a collaborative network, with a remote server that had certain processing capabilities even my ‘office’ system lacked), I did go through a number of accounting-related tasks that I would build on throughout Friday morning and mid-afternoon. Indeed, I didn’t leave until after five that evening, and even then, it was because the offsite system kicked me out for using my processing limit as I was testing the checkpoint file it had generated.
The fact that it had, left me worried that either I wouldn’t be able to download the file for internal use, or that it would be broken and wouldn’t work on my own system. Fortunately, this proved not to be the case; upon opening the program Friday afternoon, the file installed without incident. Unfortunately, what it generated turned out to be nothing more than variations upon the photographs I’d trained it on.
I couldn’t figure out why this was all I was able to get as a result of my prompting, given that your file had been based off of one that had all those artists and styles embedded in it. Evidently, that’s irrelevant; in order to access whatever’s in the original checkpoint file requires merging the two files, which you can do (and not only that, you can select from a sliding scale of how much of file A and how much of file B, thereby creating a separate checkpoint file incorporating aspects of both; I used a 70/30 split, but might consider creating one with a 50/50 split as well. Apparently, you can even merge in a third file, but I didn’t get that far – I only had a couple of hours before I would need to get home, remember?). With that being said, here’s what it came back with when I asked for you as a magazine cover girl this time:
Speaking of swimsuits, I asked for an underwater photo, and this is one of the results:
Of course, there’s still a lot I need to learn in order to avoid certain issues I never anticipated. I asked the computer to imagine you as a Bond girl, and somehow wound up with you as Bond as well:
I had to learn to include “single subject” in my prompt to avoid this; before that, there were some particularly unusual results, especially when I asked for a painting of you as a Valkyrie:
The proportions on the group in the upper left are particularly egregious, although you might not notice them at first. The composition on the bottom is almost passable, if you assumed the version of you behind your main figure was simply adjusting the ropes on the ship, preparing it for departure – they even look different enough to not make one think they’re based on the same person.
While there’s so much more you can see that can be done, even with just this 70/30 merged checkpoint file as set up (I have others that focus more toward manga, comics, photography and so forth – and there’s also the question of the merger proportion), as I said, I only had so much time, and each request would take between two and five minutes to assemble – which means there were plenty more of these, but I’m running short on both space and time. And with things being put that way, I’ll leave you with one more image:
Hopefully, future combinations and tweaks will result in even more interesting pictures of you. I just wish you were able to see these for yourself.
Anyway, keep an eye on me, honey, and wish me luck; I’m going to need it.
3 thoughts on “What You Get Out of It”
Thanks! I’m so glad I stumbled across this post – it’s been a real eye opener and provided me with a lot of new information and its good to see the images generated. Many thanks for sharing your understanding!
Looking forward to seeing your images once you get the hang of this.
Would like to see more of the images that focus more toward manga, comics. AI art is an eye opener