Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
If you have an email ending in @hotmail.com, @live.com or @outlook.com (or any other Microsoft-related domain), please consider changing it to another email provider; Microsoft decided to instantly block the server's IP, so emails can't be sent to these addresses.
If you use an @yahoo.com email or any related Yahoo services, they have blocked us also due to "user complaints"
-UE
If you use an @yahoo.com email or any related Yahoo services, they have blocked us also due to "user complaints"
-UE
Comments
Other stuff I got that I thought was interesting/funny:
It's slightly different to what I expected - I guess that's what you mean by these troubles with getting the AI to grasp the Frazetta style - but I could totally use a bunch of these images as D&D character art. And one as a piece of gay art. As for the rest, I believe I could see a few boomers treating it as legitimate devotional focus much like with that image of prequel Obi-Wan that's been making the rounds a decade or so ago.
One???
I'm trying not to do AI art right now but I did finally learn to use LoRAs and now I really, really want to try them out.
re:subject and artsyle, maybe use that technique where you structure the prompt in paragraphs. It works for some people, though of course never for me.
Funny thing is it seems to only create head and shoulder portraits, and there appear to be ways to make it generate very weirdly specific things.
It occurred to me that maybe I used the wrong resolution, but trying portrait did not help.
(I like how the second one has a vaguely postal stamp look to it).
I genuinely think there's no way to make this thing generate anything more than head and shoulders portraits.
As for weirdly specific things, "accent" makes it give everybody lipstick, no matter how hard you try to mask it.
Other pieces.
Aside from that, now that I've learned to use LoRAs, I want to use Stable without LoRAs and see how far I can push it. I could always learn to "train" my own Stable, but that seems like it'd require even more computing power and I don't care for it.
So what I've been doing is writing prompts in two or three parts (I used to do this, but not as seriously as I am now).
One thing I had to re-learn is that base Stable has a thing against Asian men (how is the conservative position the one getting mugged by reality here?).
The prompt for this literally contained "Asian man". I think it also contained "woodsy cabin", which is why it literally carved a cabin out of a tree...?
"Triadic colors" usually gives me odd geometric backgrounds, but this time it gave me random planets (and dinosaur birds?).
Same prompt, and more typical. I always set seeds to slightly negative, which is why the pieces come out so different I think.
The best I've managed so far with no extra models and LoRAs is this;
I added an extra color or something here I think, and also changed the resolution, and it lost the pseudo-realism (but still looks nice, hopefully imgur won't Douyin me again).
This is also nice, but for some reason "dandelions" always becomes "giant dandelions".
It did not win.
The second one is almost okay, but still not right.
I even tried "hat boxes", knowing very well that this would add hats but nooooo.
Sometimes, AI just messes with you and creates hobbits for some reason. I'm guessing it's the resolution I picked.
I have accidentally created a gallery of gay pride.
After a while, even the hand gestures got very gay.
Merry Christmas~
I recently learned that "ethically sourced" models actually exist, those based on public domain, cc0 or licensed stuff. It's something I was hoping for so as to get around the moral quandary in using artists' works to compete against them. The ones I know are Mitsua and Ethical Diffusion (which has since been deleted) and there's next to no info on them out there. Let's experiment:
First try:
Let's try again and compare with other models (note: the first two are models built on top of base SD (i.e. better),
middle two are base SD(edit: wrong picture), last two are Mitsua, the ones we're interested in:Maybe it's the sampler?:
Maybe it's the SD VAE?:
Ahh, that's more like it. Let's try again (note: middle one is base SD 1.5):
It's promising. Let's try again but with modern-day stuff (since it's based on stuff they're allowed to use, it's biased towards old stuff in the public domain). Let's also include SD 2 (note: I don't know what "ema only" means or entails):
Ehhh...
(It shows that the base for anime stuff is vRoid.)
I don't yet know how to train LoRAs, but if it's like how it works in other models it goes a very long way in getting you what you're looking for.
But after I kept changing the wording, it got better.
I also ran this same image through a new XL model (Dreamshaper 7) and I kept getting this very samey look that was similar to the base even if I tried specifying the style.
Same thing happened with this prompt that had been giving me issues for ages anyways (tl;dr, AI does not believe in beach balls).
This last one came out well but sadly it has the curse of Extra Limbs.
Then I tried without the beachball.
Then ran it through Dreamshaper, which... did not exactly improve the style.
It's not bad. It's just not what I wanted, and you can see how it's basically the same style as earlier, even though this prompt has some extremely obvious style cues.
Related to that:
This was actually sort of different from what I expected from the prompt, but it's very artsty.
Found a cool new Vector-style LoRA and made some CCP Propaganda.