Stable Diffusion Thread
#21
Some very interesting stuff in this thread.

I want to get into this kind of thing.

If I want to build a PC for large batch (1000s) img2img generation (ideally "high res" without the need to upscale post generation) what should I choose between a 4070ti or 3080 ? (I don't really want to buy a used 3090 because I have very bad luck)

Thanks!
#22

Corridor Digital's use of diffusion in this video is incredibly innovative. They are creating a new form of animation that can give consistent results and can be completed quickly. I look forward to seeing how this will be used in the future.
[Image: HfVqWXY.jpg]
I simply follow my own feelings.
#23
[Image: R0Z8l9I.png]
[Image: CVdTbm2.png]
[Image: LJBsq4T.png][Image: fF4ix4H.png]
[Image: 9xCvpR9.png][Image: EXsg61y.png]
[Image: RZpCmfZ.png]
[Image: U63evIy.png]
#24
Stable diffusion.

[Image: 7563059cc0918b3bebc341632522f753.jpg]
#25
Early experiments with ControlNet Human Pose. Unfortunately it wouldn't let me use lowvram or medvram with it so the images were very low-res.

"Congratulations on your election, Chancellor."

[Image: jJ7lMLR.png]

As for the other tools, I've found depth to be extremely useful for "thing, but in style of other thing" generation, far more than img2img. Normal wasn't nearly as useful as it seems because automatic normal map generation sucks. Others can be useful for certain niches.
#26
"Now hold the hell up. You're telling me this guy was a nigger?"

[Image: axQvAzq.png]

"More drama on Civitai. I think I'll sound in."

"Srgrafo's style of C Cupped anime girl is a perfectly natural object of attraction for any healthy post-pubescent male. The absurdity of tasteless and philistinic Americans denouncing ecchi as 'child porn' beggars description, an

[Image: UOpweME.png]

"Hmph. I know that you enjoy 'memeing hard' on imageboards and such, but you must understand this is scary talk for people irl (in real life). The inevitably lolita-adjacent appearance that SD was bound to produce from a dataset of this kind should have been apparent to the creator. This is just a horrible idea for a Lora style model. Browse the subreddit devoted to the works of the creator in question and you'll see what half of his fanbase is into

[Image: UuSWblB.png]
#27
The world's best Soybooru diffusion model is completed

[Image: b7QgHnw.png]
[Image: WXX0Rh5.png]
[Image: Nl64X3D.png]
[Image: 8VhsmFC.png]
[Image: CMbftuh.png]
[Image: nRmeNzy.png]
[Image: JrS0aCa.png]
[Image: XTGr4SC.png][Image: UfFI8GP.png]
#28
These are so funny. I love them.
#29
(02-26-2023, 06:32 PM)Albicacore Wrote:
Corridor Digital's use of diffusion in this video is incredibly innovative. They are creating a new form of animation that can give consistent results and can be completed quickly. I look forward to seeing how this will be used in the future.

This isn't a new form of animation by any stretch, however it is a very good innovation in rotoscoping and a glimpse at how I suspect AI will be used in animation. Using the AI they expedited the rather tedious rotoscoping process greatly, and from what I can tell which if your model is good only leaves some cleanup/other details left for the animator. I expect this expedition of tedium to extend to other parts of production like backgrounds.
#30

[Image: CLiHcpq.png]
[Image: cB7NdAs.png]
[Image: 9KvsaT5.png]
[Image: ArVTsOC.png]
[Image: TTcGLFP.png]
[Image: lrB5mve.png]
[Image: SKav5u3.png]
[Image: LXjdMYI.png]
[Image: dhkNvN5.png]
[Image: oVOIrSH.png]
Heart 
#31
Phenomenal work JF_, if trannies weren't furries this would be the direction of "retro revival" for 3D and games. The AI 'jaks as well are hysterical.

Mr. Nobody was also my favorite from Doom Patrol.
[Image: 0b065d24f0c9a57c24af80cac8885ba2.jpg]
#32
Stable Diffusion XL 1.0 is out, but who needs it? Checkpoints now are absurdly good and Hires Fix creates incredible images well beyond the intended size.

I've just discovered that ComfyUI does VRAM management so much better than AUTOMATIC1111's UI that I'm finally able to generate decent images with a fairly unimpressive graphics card. The Gigachad LoRA is especially fun to play around with. I've generated them in color because it gives me more leeway to play with levels when I turn them grayscale, but I've left these untouched to show the raw potential:

LIKE A TRIZZBO IN THE DAAAAAAAAARK
[Image: NbOUlef.png]

My ComfyUI setup is linked at the end. Very few duds among them - it usually gets the prompt perfectly on the first try. I have a ton of negative prompts, but I don't think the negative prompts even do anything. I tried putting them all in to the positive prompt to generate a joke image but it still came out looking normal.
#33
Should I get a Nvidia card? There wasn’t any good software for AMD a few months ago.
#34
(07-27-2023, 04:40 PM)BillyONare Wrote: Should I get a Nvidia card? There wasn’t any good software for AMD a few months ago.

Nvidia works off the shell with most stuff - see ComfyUI that Frank is using. You *can* get it (and other things) to work with AMD, but it will often involve coding and reworking python scripts. Nvidia is presumed default by most individuals making ML packages.
#35
Thanks. I’m not good with IT stuff so I’ll save up for a nvidia. Unless something good comes out for AMD soon.
#36
I don't know if this fit with the current thread,but this an future scenario written by an ai where Adolf
Hitler is recontextualized as if he was born
in America in 2009. https://www.reddit.com/r/ChatGPT/comment...o_writing/

"Born in 2009 in America, Adolf Hitler was
given the name "Adam Hightower" by his
parents. As a child, Hightower was known
for his strong leadership skills and
tendency towards authoritarianism. He
was an excellent student and excelled in
sports, but struggled with social

interactions and often bullied his peers.".....
#37
Study: Gigachad in Rockwell's Freedom of Speech
[Image: 2FpNfL0.jpg]
Experiments with T2I (more performant chink ripoff of ControlNet) and combining LoRAs.


Study: Gigachad in Corporate Memphis
[Image: oEoabv1.jpg]
Exploring the comedic potential of everyone's favorite art style with a daring LoRA synthesis.

Miscellaneous images
[Image: NJ9nsmt.jpg]
#38
Lost in a Roman wilderness of pain
And all the children are insane
All the children are insane
Waiting for the summer rain

[Image: 69fe49f1-80b6-4963-9967-7b81dea0fe64.jpg]



[-]
Quick Reply
Message
Type your reply to this message here.




Users browsing this thread: 1 Guest(s)