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

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
So, these were made by SDXL with a pixel art lora.

View attachment 5249097View attachment 5249098View attachment 5249099View attachment 5249100View attachment 5249102View attachment 5249103View attachment 5249104View attachment 5249105View attachment 5249106View attachment 5249132

Since pixel art is really lossy re: fidelity/information, it has the advantage of not really pointing out mistakes. I didn't bother editing them, but it would've been easy to. All I did was run a script for limiting the color palette and dithering, if necessary. It wasn't even really good and could've been done better. (If they're blurry, it's because your browser is scaling them, try opening them in a new tab) I also accidentally everything to jpg, sorry.
I love pixel art. These came out amazing!
 
I'm not sure, here's another AI influencer and it looks like however they are making it is having trouble.
If I had to guess, I'd say it's AI generated but heavily touched up to remove all the jank. How they managed it I have no idea but as costly as it would be there's a fairly obvious way they could recoup their time and money investment with content that lifelike.
 
Hey there KF I have some more stuff for you guys, I haven't posted on this thread in like.. a few months but here some new img's i've done since I last posted
The last 3 images in this set are kind of "remasters" of obscure PC-98 games Hyoryu, Angel Army, and Red Cobra
 

Attachments

  • 00298-3256950402.png
    00298-3256950402.png
    430.8 KB · Views: 61
  • 00011-3829188528.png
    00011-3829188528.png
    467.4 KB · Views: 61
  • 00004-1966857446.png
    00004-1966857446.png
    565.2 KB · Views: 54
  • 00053-1899977764.png
    00053-1899977764.png
    568.4 KB · Views: 58
  • 00069-501129034.png
    00069-501129034.png
    646.3 KB · Views: 57
  • 00097-4048606012.png
    00097-4048606012.png
    519.9 KB · Views: 53
  • 00129-3077305476.png
    00129-3077305476.png
    766.9 KB · Views: 56
Is this a deepfake of an existing model or something? I've never seen an AI video depict that much motion without looking like reality is melting.
This is a whole new level of deepfake if that body isn't real. I'd love to see someone do a deep dive into how the sausage is made (ping me if you find out), but in the meantime I found this technique using img2img controlnets and ebsynth, which seems kinda close if you squint:
CivitAI Tutorial (needs login)
YouTube (age restricted)

You still get some reality melting, but I imagine if you use something like a lora and maintain the same seed for img2img, it might work better?

Example coomerbait results:
TikTok
Twitter (I refuse to call it X)

Shit's going to get even more creepy real fast.
 

Attachments

  • Video to Video Stable diffusion tutorial with Multi Controlnets and Ebsynth and Segment Anythi...mp4
    39.8 MB
  • Agree
Reactions: A Hot Potato
My position on LoRAs and Hypernetwork training for artists has always been that they are inadequate for capturing the major themes of an artist which makes them desired. Today I trained a LORA on an artist in SDXL and I don't feel that sentiment anymore. It cost me $3 to train a model to shit out copies of art that are near identical to the original artists work. The artists are right about becoming obsolete. Even the distinct ones don't stand a chances now. The LORA can even handle multiple subjects. Something I never thought SD could crack.
 
So which furry porn artist was it
I've had mixed results fine-tuning with furry artists. Originally they seemed like a good fit because they produce a shitload of 1:1 images that I could train into a hypernetwork. After running Gallery-DL on an artists gallery, I was able to harvest a sample size of around 1000 suitable images. More than enough for training a Hypernetwork using cosine. The main thing is the colors and shading especially with clothing involved. Furry pornography is about colors and character expressions though more so than animals, so it's a wash anyway or it least it was when I started out doing fine-tuning.

Of course I've only tried training off of 1.5/2 SD. Maybe SDXL LORA is the silver bullet along with the new tagging extension in automatic1111. Maybe I'll rent a 4090 for a day and retry it on the old dataset. SDXL is really receptive to negative prompts, so I may even be able to use a negative prompt to extract the furry out and end up with a normal person. I'll share the results if I can make it SFW and human.
 
Last edited:
Back