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

Colon capital V said:
Chat it might actually be over for us if AI has gotten this good.
Click to expand...
Full Krea model:
Highly realistic photo of a slightly obese bearded North American man with brown hair and green eyes. He points at a piece of cheese screaming. Wears a red flannel shirt. Action takes place at a supermarket.
Right: Flux.schnell with the 512 LoRA
Highly realistic photo of a slightly obese bearded North American man with brown hair and green eyes. He points at a piece of cheese screaming. Wears a red flannel shirt. Action takes place at a supermarket.
Definitely fares better as a standalone model.
 
dumbass
Behold, the dumbest shithead to ever live.
mongoloid
I've hoarded a whole bunch of LoRA's and I haven't been really using them.
And I should be using them more cuz they're pretty damn fun.
 
Is that supposed to be Jack or Legoshi from Beastars, only more autistic?
 
Baraadmirer said:
Is that supposed to be Jack or Legoshi from Beastars, only more autistic?
Click to expand...
It's supposed to be an autistic border collie as this used to be my avatar a few good years ago so close enough. I've been associated with a border collie for years, and I've been known as a fat autistic Polish furry for years, so why not.
1754782830389.webp
If you go back to the first page of this thread you can find my post when I tried i2i with SD1 on this image, so I've been on the grind since the early days.

Fun fact: somehow that old i2i gen ended up on u18chan's AI thread of all places, reuploaded by someone else. Fucking weird how some content can proliferate.
 
Slav Power said:
Full Krea model:
View attachment 7730011
Right: Flux.schnell with the 512 LoRA
View attachment 7730018
Definitely fares better as a standalone model.
Click to expand...
It basically just needs to remove the blur and mimic bad human stuff like angled cameras, lighting and improper zoom.

It's kinda funny. The odd 'AI porn humor' thread on 4chan, you'll see almost flawless futa dick vore chocolate explosion tomfoolery next to the 'best' attempt at real AI porn a year prior to the former. Day and night. Much like AI art and other products, the humanity is part of the appeal, so doubt it'll ever truly take over, but it for sure will replace porn artists who are funded only by f-list idlers and other such lowlives not above stealing and mimicing to make a profile called 'LOLI NIGGER RAPIST'.
 
I am confused by how people make anything in Stable Diffusion. I went through all the effort to install it and it just makes Cronenberg body horrors when i tell it to do a person.
 
Ughubughughughughughghlug said:
I am confused by how people make anything in Stable Diffusion. I went through all the effort to install it and it just makes Cronenberg body horrors when i tell it to do a person.
Click to expand...
What prompt are you using? Generally you're going to want to add a negative prompt for things like bad hands and extra limbs.

This is something I generated a year ago:

20240828_205258-2772106982.webp
 
Baraadmirer said:
What prompt are you using? Generally you're going to want to add a negative prompt for things like bad hands and extra limbs.

This is something I generated a year ago:

View attachment 7778964
Click to expand...
Professor Chad?
I would have ChatGPT listen to me croon about the portrait I wanted to make, then have it generate the prompt (which included the negatives for extra limbs and such), and then I'd feed it in and it would go catastrophically. Kept doing so even when I'd roll it back to something as simple as, say, "man in forest."

I know it has to be a me thing, I'm running 1.5 locally and I have two models installed.
 
Ughubughughughughughghlug said:
Professor Chad?
I would have ChatGPT listen to me croon about the portrait I wanted to make, then have it generate the prompt (which included the negatives for extra limbs and such), and then I'd feed it in and it would go catastrophically. Kept doing so even when I'd roll it back to something as simple as, say, "man in forest."

I know it has to be a me thing, I'm running 1.5 locally and I have two models installed.
Click to expand...
ChatGPT gives more natural language for it's prompts which doesn't work as well for SD 1.5, it does better with prompts that are almost more list like. here's what a prompt looks like for 1.5 and the results:
Screenshot 2025-08-14 121025.webp94057356.webp
if it's not your prompts it could be you need to change your sampler, tweak the CFG scale, adjust your step count or find a better checkpoint.

there are far better models now than SD 1.5 though. I'd recommend flux for realistic stuff and illustrious for anime/cartoony stuff. both work better with more natural language prompts like what you get from ChatGPT and have a lot more training data so their outputs are far better than SD 1.5.
 
Ughubughughughughughghlug said:
Professor Chad?
I would have ChatGPT listen to me croon about the portrait I wanted to make, then have it generate the prompt (which included the negatives for extra limbs and such), and then I'd feed it in and it would go catastrophically. Kept doing so even when I'd roll it back to something as simple as, say, "man in forest."

I know it has to be a me thing, I'm running 1.5 locally and I have two models installed.
Click to expand...
Ah, I remember when I would run 1.5 locally. I've moved on to Illustrious.

I'd suggest looking online for text embeddings, which are basically shortcut words that stand in for multiple tokens that the model understands. CivitAI has embeddings for download.

For example, the prompt and negative prompt for the image I posted above was:
Code: 
absurdres, (upper body:1.1), (pointer:1.3), (mature male:1.1), (muscular male:1.3), classroom, tweed suit, holding, chalkboard, glasses, pointing

Negative prompt: 7dirtywords, bad-hands-5, JuggernautNegative-neg, EasyNegative, FastNegativeV2, female, heterochromia, negative_hand-neg, blurry, (worst quality, low quality, normal quality), (head out of frame:1.2)

It also depends on how your models were trained: do they recognise natural language, or have they been fed with booru tags?
 

Attachments

  • ComfyUI_00043_.webp
    ComfyUI_00043_.webp
    1.8 MB · Views: 33
  • ComfyUI_00115_.webp
    ComfyUI_00115_.webp
    1.5 MB · Views: 38
  • ComfyUI_00243_.webp
    ComfyUI_00243_.webp
    2 MB · Views: 41
  • ComfyUI_00250_.webp
    ComfyUI_00250_.webp
    2.2 MB · Views: 36
  • ComfyUI_00253_.webp
    ComfyUI_00253_.webp
    2.8 MB · Views: 30
  • ComfyUI_00288_.webp
    ComfyUI_00288_.webp
    1.3 MB · Views: 28
As a break from what I normally generate, here's a cute little otter spirit from Touhou 17 (no LoRA). Haven't been able to generate a wolf or eagle spirit in their canon colours yet.
20250526_144423-2669640299.webp
 
Ars Technica: New AI model turns photos into explorable 3D worlds, with caveats (archive)
On Tuesday, Tencent released HunyuanWorld-Voyager, a new open-weights AI model that generates 3D-consistent video sequences from a single image, allowing users to pilot a camera path to "explore" virtual scenes. The model simultaneously generates RGB video and depth information to enable direct 3D reconstruction without the need for traditional modeling techniques. However, it won't be replacing video games anytime soon.
Click to expand...
The model demands serious computing power to run, requiring at least 60GB of GPU memory for 540p resolution, though Tencent recommends 80GB for better results. Tencent published the model weights on Hugging Face and included code that works with both single and multi-GPU setups.
Click to expand...
 
You must log in or register to reply here.
Back
Top Bottom