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

I’ll second the idea that even an extremely crude drawing with the right prompts gives incredible results when compared to the prompts alone.

I did an mspaint doodle with my mouse of a standing figure and described the person in words and I’ll be damned if it didn’t look 100 times better than running it with the words alone.

I have a very fuzzy understanding how this all works but my guess is that the model can spend more cycles making something nice when I’d doesn’t have to start by assigning rough locations and proportions.

I am seeing a weird issue where more iterations doesn’t necessarily give better results though. Sometimes artifacts are just warped and exaggerated instead of corrected.
 
I have a very fuzzy understanding how this all works but my guess is that the model can spend more cycles making something nice when I’d doesn’t have to start by assigning rough locations and proportions.
It's sculpting random noise, basically image recognition in reverse. So when the source of that "noise" is a rough sketch of what you're imagining*, naturally it'll land on something closer to what you imagined. This happens at the expense of "creativity" in terms of it serendipitously landing on something that you like even better that wasn't what you expected (although in a roundabout way you're right because that, in turn, tends to happen at the expense of needing to throw away 90% of a batch job to get one that flukes into coolness).
Denoising strength is the key here, you can get great results with a less-shitty sketch at lower denoising values, or raise the slider to get looser and less img2img-like interpretations of it as though it were shittier. Lower values also tend to let you get away with higher CFG values (which is basically strictness to your prompt, again at the expense of "creativity"--but that's usually what you want with img2img).
Also with higher denoising strength, try throwing in completely unrelated source images--the logic of their underlying structure still leads to interesting results even if it's not a sketch of what you wanted. And it's neat to see your tentacle hentai turn into a sweet spaceship or whatever.

*(Also reminds me of a cool tip I saw if you're aiming for more photorealistic img2imgs: take your shitty sketch into photoshop and use a filter tool to add a bunch of multicoloured noise to the whole image first.)
 
Last edited:
The things I get when I ask politely...

"The Ultimate Warrior on $100 bill"
grid-0051.png

"Your mom sneeds" (Refer the lower-right image for endless horrors)
grid-0052.png

"Rock's Modern Knife"
grid-0053.png
 
I am seeing a weird issue where more iterations doesn’t necessarily give better results though. Sometimes artifacts are just warped and exaggerated instead of corrected.
That's more "just how it works" than a bug... what's the issue? Bad anatomy and distorted colors can be partially mitigated by dropping the cfg scale. You can also try inpainting them away, but I've had bad luck with it personally except replacing faces and hands.
 
Got stable diffusion running well on my machine after the site went down. My nitro+ vega64 renders pretty quickly. But there seems to be some kind of memory leak or a inability with the program and my hardware to dump the memory from the last render. Which causes me to force restart the program to free up my vram. Annoying, but not world ending.
Here's some examples Ive generated. Really enjoying this. Using it alot to generate concept art for a freinds game. Currently investigating it to generate textures for 3d modeling. I've also taken some of my friends art and fed it into the AI to see what it what interpret from it. It's fascinating to see what it fixates on.
 

Attachments

  • 00161-1508862666-a bird dying in a pile of rubble,tone mapped, shiny, intricate, cinematic lig...jpg
    00161-1508862666-a bird dying in a pile of rubble,tone mapped, shiny, intricate, cinematic lig...jpg
    118.7 KB · Views: 31
  • 00160-2690619130-a bird dying in a pile of rubble,tone mapped, shiny, intricate, cinematic lig...jpg
    00160-2690619130-a bird dying in a pile of rubble,tone mapped, shiny, intricate, cinematic lig...jpg
    115.8 KB · Views: 30
  • acarinacanyonn.jpg
    acarinacanyonn.jpg
    296.8 KB · Views: 35
  • 00074-3865610695-rusty van in dark forest, night, raining, wet, lights on in van, spooky, creepy.jpg
    00074-3865610695-rusty van in dark forest, night, raining, wet, lights on in van, spooky, creepy.jpg
    44.9 KB · Views: 39
  • 00004-297737101-Planets, Moon, space ship, zdzisław beksiński.jpg
    00004-297737101-Planets, Moon, space ship, zdzisław beksiński.jpg
    39.1 KB · Views: 35
  • 00063-1462743029-a room with a chair and a painting on the wall and a window with a view of a ...jpg
    00063-1462743029-a room with a chair and a painting on the wall and a window with a view of a ...jpg
    287.2 KB · Views: 32
  • received_900077274730657.webp
    received_900077274730657.webp
    180.7 KB · Views: 27
  • 00156-1841300205-bearded dragon in a forest,tone mapped, shiny, intricate, cinematic lighting,...jpg
    00156-1841300205-bearded dragon in a forest,tone mapped, shiny, intricate, cinematic lighting,...jpg
    236.2 KB · Views: 33
  • thehellhead.jpg
    thehellhead.jpg
    106.1 KB · Views: 33
  • 00095-697867862-A giant arcane robot stands waist deep in a stormy ocean, facing away, surreal...jpg
    00095-697867862-A giant arcane robot stands waist deep in a stormy ocean, facing away, surreal...jpg
    28 KB · Views: 32
  • 00092-738288248-A giant arcane robot stands waist deep in stormy ocean, facing away, surrealis...jpg
    00092-738288248-A giant arcane robot stands waist deep in stormy ocean, facing away, surrealis...jpg
    35 KB · Views: 33
  • 00106-2528731319-A lonely bronze statue deteriorates in a field, , Antoine Wiertz, oil on canv...jpg
    00106-2528731319-A lonely bronze statue deteriorates in a field, , Antoine Wiertz, oil on canv...jpg
    37.5 KB · Views: 32
  • 00169-3570950351-a child sitting in a dark basement next to a computer, Ben Templesmith.jpg
    00169-3570950351-a child sitting in a dark basement next to a computer, Ben Templesmith.jpg
    82.4 KB · Views: 37
  • 00123-452298632-a train going through a post apocalyptic landscape, surrounded by corpses, gor...jpg
    00123-452298632-a train going through a post apocalyptic landscape, surrounded by corpses, gor...jpg
    157 KB · Views: 32
  • 00126-3425439884-a fireman in a post apocalyptic city looking solemly at charred corpses, gore...jpg
    00126-3425439884-a fireman in a post apocalyptic city looking solemly at charred corpses, gore...jpg
    107.7 KB · Views: 36
  • apocfuturecity.jpg
    apocfuturecity.jpg
    197.5 KB · Views: 34
Found these on Twatter the other day: "Marilyn Manson eating Marilyn Monroe"
Frightening. He should dress up as Marilyn Monroe. Make the scary into reality.

Narushrek.png


I finally got Dolan Dangers to show up but not how I want. I did the jew thing so he can be facing towards the camera and waving. I think I'm just too used to proompting for NovelAi model and all the other shit trained off of danbooru.
Dolan 1.png
Dolan  2.png

Lol .png
 
Last edited:
masterpiece, best quality, masterpiece, 1man, (hellsing:1.5), pistol, red moon, night, cemetery, action pose,
1668073747021.pngtmppgwjx4wq.png

Refer to the link for a /g/ anon's tier list of styles
 
Back