Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • Want to keep track of this thread?
    Accounts can bookmark posts, watch threads for updates, and jump back to where you stopped reading.
    Create account
4chan is telling me things are starting to get pretty ridiculous in regards to AI generation

NSFW
ai1.jpg
ai2.jpg
1703472327712014.png
 
Wow, when you tell SDXL with the latent upscaler turned on to produce "detailed" pictures, it gets detailed. SD1.3 with ((intricate)) at the front of the prompt levels of detail.

It's definitely a step up, both in terms of speed and quality. I'm surprised, everyone told me it would be garbage because the training dataset was so limited due to copyright fags, but it actually seems really good to me. I need to learn an entirely new style of prompt though, neither the old!SD word salads nor the more conversational DALL-E3 prompts seems to work well, I had to use a sort of in-between style.
Definitely glad I bought a 4090, SDXL churns out large batches really quickly even if I turn on the refiner or upscaler.
 
Microsoft is a company run by retards.

Their newest shtick appears to be to memory hole / shadow ban undesirable accounts. I tried for two hours to create images with BING AI and I got absolutely nothing. There was nothing offensive or racist in the prompts, just normal every day shit I wanted to do for fun.

I always got the “we are busy try again later” notice. And wouldn´t you know, I logged in into a second account and I could create 60 images in a row with no problems. Interesting. Guess I called them dickless homosexual goatfuckers too many times, could be, I don´t know. Goes both for BING AI Image Generator and Designer. I did not use a VPN so it looks like it is not IP blocked but solely related to an account.

@MMS-China since you are our resident China faggot I am glad to inform you that Microsoft kowtows in front of your beloved Chairman Xi. It is impossible to create a proper flag of the Hong Kong Special Administrative Region of the People's Republic of China and the Macao Special Administrative Region of the People's Republic of China, and instead you get ... this...

7b645e47a974.jpeg 93fd-bdc855143544.jpeg 826a-2301dd0c8e2d.jpeg

@snov fully agree, SDXL can be pretty neat. I used Mage.Space for about a year before I jumped ship. The difference to DALL E3 is wild. Try using prompts and look at the difference, for better or for worse. SDXL can be fun, and it was amazing especially in comparison to the old SD 1.5 and 2.1 but the fine-tuned models sort of ruined the fun. Too easy to create great images, and I am a easily getting bored when I do not have to challenge myself and find workarounds.
 
Lemmingwiser said:
Did you ask for fat whale with vibrator instead of finger, man's fucking legs dissapearing behind a wall and underage futa blowjob? wtf is wrong with you?
Click to expand...

What did I just say, I found these in the 4chan AI thread. It's interesting watching the degeneracy of the Westerners.
 
MMS-China said:
NSFW
Click to expand...
Near said:
:null: I am featuring this thread because discussion elsewhere on the Internet is dominated by pornography and very unsavory types. Please enjoy the death of art and the beginning of the AI era responsibly.
Click to expand...
If you're just going to use my site to dump your weird obsession with hating white women and worshipping bug ladies I will send you back to /pol/.
 
If anyone's been thinking of making their own LoRAs with booru captioning, I recommend the BooruDatasetTagManager as it lets you apply tags globally to your dataset and alter tags for each image-caption pair without opening each individual .txt file in Notepad.
 
MMS-China said:
Uh oh, this doesn't look good

1704581356301259.png
Click to expand...
Having looked into this a little more, he’s just prompting the supreme realism 1.5 checkpoint to produce a woman holding a paper or a drivers license, and then photoshopping the text in.
If I were going to do this I’d probably just take a photo of myself holding up the paper, and use inpainting to change the face. Photoshopping text onto verifications is like the first thing they look for. It’s something about the pixels or something.
 
MMS-China said:
4chan is telling me things are starting to get pretty ridiculous in regards to AI generation

NSFW
Click to expand...
Was it really necessary to post that last one? I mean any of them really but esp. the last.


Anyway, my reason for clicking on this thread is actually to share this how-to someone did on keeping a character consistent across multiple generations.

ComfyUI and although what he does works for the face, he doesn't show it for multiple poses and I suspect it wouldn't work. I don't know how you would get a consistent character from multiple angles. I imagine you would have to train a model on a specific 3D model or real photographs of a single person from different images. Still, I found it a fun video seeing what you can do with ComfyUI and what some of the different widgets in all this actually mean.
 
Stupid question: What exactly is the difference between a model and the XL version of that model? Same for loras.
 
Lord of the Large Pants said:
Stupid question: What exactly is the difference between a model and the XL version of that model? Same for loras.
Click to expand...
It's one higher.

Fine.

XL means it's based on and only runs with the SDXL model(s) and software(and VRAM) that can handle it.
So, the model needs to match the LORA and whatever software you use has to support it.

SDXL is generally trained on 1024x1024 images compared to the base Stable Diffusion at 512x512. So that's the optimal size to start with.

You can run it with an 8GB video card, but it's a tight fit, and in my case I have to close most of my web browser windows as they use too much GPU memory, fucking Chrome.

Few more notes, the base SDXL comes in 2 parts. The main model and a refiner model. It's optimized for those two and most software supports the 2 stages. Almost every other model you'll find doesn't use the refiner as they've been trained without it.

You will sometimes see models that say "Baked VAE" or "Needs VAE" in this case you may need to tell your software to use the VAE in the model(Baked) or a separate one. Using the wrong option causes all kinds of visual artifacts. There is a standard SDXL VAE that's used for most models where it's not "Baked". VAE in simple terms converts the output of the model, a Latent image, into an actual human viewable image.
 
davids877 said:
It's one higher.

Fine.

XL means it's based on and only runs with the SDXL model(s) and software(and VRAM) that can handle it.
So, the model needs to match the LORA and whatever software you use has to support it.

SDXL is generally trained on 1024x1024 images compared to the base Stable Diffusion at 512x512. So that's the optimal size to start with.

You can run it with an 8GB video card, but it's a tight fit, and in my case I have to close most of my web browser windows as they use too much GPU memory, fucking Chrome.
Click to expand...
Yeah, I'm staying on 1.5 for the time being cause there's a lot more LoRAs offered for it, and I don't feel like funneling all my memory into image generation when I'm doing other things.
 
davids877 said:
You can run it with an 8GB video card, but it's a tight fit, and in my case I have to close most of my web browser windows as they use too much GPU memory, fucking Chrome.
Click to expand...
To expand on this, I recommend avoiding SDXL models if you have an 8 GB VRAM card or not more than 16 GB RAM; you can get away with it, but it's going to be a lot slower than using a non-XL model and upscaling 3-3.5x, plus your computer won't be good for much else during the process.
 
I’ve been enjoying this a lot.

Halloween kiwis
IMG_1529.jpeg
IMG_1528.jpeg
IMG_1530.jpeg
Kiwi typing up a lolcow thread
IMG_1531.jpeg
I wanted to make one with a bowtie
IMG_1532.jpeg
Bri’ish
IMG_1533.jpeg
Chilling with dog
IMG_1534.jpeg
Kiwi playing a game
IMG_1535.jpeg
Kiwis getting ready for war
IMG_1536.jpeg
IMG_1537.jpeg
They love the piano
IMG_1538.jpeg
IMG_1539.jpeg
Warrior
IMG_1540.jpeg
 
Last edited:
You must log in or register to reply here.
Back
Top Bottom