Log in
Update May 2024
Time has really been flying for me. I hope everyone is doing well.
Checking my last update post, I was surprised at how far back the things I talked about are—I guess I found myself remembering what I’ve shared on Discord, and thought I’d posted blogs about it that I didn’t. Whoops!
I’ll take that as a sign that I should remind people that my Discord channel still exists. It’s mostly quiet, but I post more of my unorganized thoughts on there (not to mention I check daily for activity, if anyone has comments or questions). Here’s the link if you’re not already there.
Music
I figure I’ll lead with this: I’ve made an album for you! Here you go!
Well, okay, I’ve just been playing with Suno, which allows you to prompt an AI to generate music for you. I’m hearing impaired, so I can’t vouch for any of the audio quality, but I’ve been mostly impressed that it’s been possible to do at all.
I’ve been working on figuring out more of the “ins and outs” of prompting music, but my hope is that it might become a useful tool for creating some background tracks for games. Could make for a nice change of pace, for those who have heard enough Kevin MacLeod in their visual novels.
Cartoon Style
Something I shared on Discord, but probably not here: I was looking into other art styles AI might be able to produce reliably, and tried to see if I could get a more simplistic look with line art that might make it easier for me to do edits (to fix things like hands, that AI tends to massacre).
Most of these were basic character turnarounds with different prompts and settings, all using the same Open Pose ControlNet:
Something I really haven’t been able to control, however, is the shading style. It seems to just turn out differently based on the seed. Case in point, these two are the same prompt, with different seeds. One has very sharp-edged cel-style shading, and the other has more of an airbrushed look:
Either style would be relatively easy to make edits and changes to, though…but it would be nice if I could reliably get the kind of results I want on-demand, though.
To have the most control over the image content, my plan would be to use this in combination with 3D models. It’ll mean a lot of back and forth between tweaking the models, and the AI prompts, until I get combos that are a good fit for each other.
I did find that I need more than just a basic Image-to-Image pass, as things break down the more the AI is emphasized over the 3D model:
I think the samples above at different denoising levels illustrates what happens if the AI has too much control. But, in the in-between stages, it shows me what it wants to generate as far as eye size, position of facial features, and the like—I figure I can try to adjust the 3D model to be a better fit.
Trying similar attempts with ControlNet for the pose and basic elements of the image seems to help keep the AI version from getting too crazy.
AI Sex Change
This one came as a request to help create female versions of male characters, but it helped me figure out some new techniques.
I’ve never completely understood the “Reference” ControlNet model (still don’t), but I did discover that I could isolate the portion of the image I wanted to use as a reference, and the AI seems to understand what I’m going for.
In this case, I wanted to try to keep the facial features and some of the hair, so I removed the rest.
This seemed to get reasonable female versions of the characters (though they’ll likely require the same art filter the rest of the game’s assets use in order to blend in better.)
AI Details
Just to share some of my pain: I’ve tried a little bit to see if I can figure out how to use ADetailer for fixing image issues, such as hands. So far, I haven’t had too much luck.
Sadly, finding a good way to address garbled parts of an image is probably one of the biggest hurdles when it comes to actually being able to do anything productive with AI art.
Paper-Doll Character
I’m still working off-and-on trying to see if I can make a paper-doll-style character (intended for Clean Slate), and it’s been a bit of a challenge to generate all of the necessary parts while making them look like they flow properly together.
Typically, I’ll automate this process by having Stable Diffusion work its way through dozens of different prompts while I’m working my day job. It’ll take some hours, after which, I’ll review what it’s made, figure out what changes to try with the prompts, and ready it for the next day’s run.
It’s a bit of a disjointed process, and I’ll often forget what I was doing with it between renders.
Patreon TOS Changes
If you’re not aware, Patreon changes their TOS on a regular basis, each time adding additional restrictions to “Adult” creators. This latest one is going to make a lot of content unacceptable on the platform, and has made me a bit uncertain about whether or not anything I make in future will be allowed.
At some point, I suppose I’ll need to make the decision of whether I want to make Patreon/bank/credit-card-safe erotic content, or if I’ll have to switch entirely to SubscribeStar. Something to think about.
Still to Come
I really want to do something with video/animation at some point. I still haven’t taken the plunge and started experimenting, but I’m curious what can be done with AI. I’ve already seen some stuff done using Stable Diffusion and keyframes, and, of course, EbSynth is an option to use in combination.
I’ve also considered pixel art. I think the low resolution could be more forgiving.
One way or another, I’d just like to have more for people to look at than just still images.
Backer Content
I didn’t want to paywall the album (though I think it would be okay with Suno’s TOS), so I already blew the biggest thing I had, but in my backer post, I’ll talk a bit more about each of the songs.
I’ll also have some more samples of some of the other items I talked about above.
