While we're waiting for them to let people use DALL·E 2, have you gotten any interesting results from these text to art AIs?

https://huggingface.co/spaces/dalle-mini/dalle-mini
https://neuralblender.com/

There's a new one called Midjourney but they want you to contact them on Discord to participate in the beta.

DALL·E mini is now https://www.craiyon.com/

Finally, instructions for running text-to-image AI on your own computer:
https://reticulated.net/dailyai/running-your-own-ai-image-generator-with-latent-diffusion/

Stable diffusion seems to be getting popular.

Does someone have an easy to understand article that explains how stable diffusion differs from another CNN?

P8591
img2img with "The Original Honeymoon Mobile"

P8644
According to this guy being compute efficient is a big reason why it's taken off:
https://www.paepper.com/blog/posts/how-and-why-stable-diffusion-works-for-text-to-image-generation/

P8644
And of course there's the fact that they released the model unlike the DALL·E fucks who are still pining about "ethics."

P8646
Those people are smart. I would really like to see the encoder/decoder part working on its own, I feel like it could be used for image compression. You can get a 64x compression ratio with jpeg, but the result will be very ugly, so this one is pretty cool. My only concern is that the encoding/decoding could be computationally intensive. Can their code run on CPU? If so, I might try to play around with it.

P8646
Those people are smart. I would really like to see the encoder/decoder part working on its own, I feel like it could be used for image compression. You can get a 64x compression ratio with jpeg, but the result will be very ugly, so this one is pretty cool. My only concern is that the encoding/decoding could be computationally intensive. Can their code run on CPU? If so, I might try to play around with it.

P8676
4/g/'s general has a guide for setting it up on CPU only, so apparently it can work that way, although most people use GPUs for it.
https://rentry.org/cputard

"lambda plus js"

I present to you "a mechanical cow with gears in the style of Hayao Miyazaki, anime"

I tried to see if the encoder/decoder could be used for image compression. Well, it turns out it can. It's very slow and eats a lot of ram though.
From left to right, the original image, the output of the encoder, the output of the decoder and the same image jpeg-compressed to the same size.

P8851
Cool. Are you using CPU or GPU? When you say it's slow and takes a lot of RAM, is that encoding or decoding or both?
JPEG XL for comparison; it looks better than the JPEG, but the neural net version is still a lot better.

P8690
I wish these guides made it more clear from the start which GPUs they would work on. I wasted a lot of time trying to get software to work that was never compatible with my card in the first place. Tried this CPU one and I think it might work eventually on my potato if I left it running, but not in any reasonable amount of time.

"lambda plus js takes a new direction"

getting a headpat from a cute anime girl pov

This looks pretty neat.
https://nitter.net/t_takasaka/status/1570716749565231104
autotranslated:
>I am experimenting with automatically generating illustrations, animations, and 3D models in real time from a single sketch.
>You can change the shape, color, and texture of the parts with the sliders on the UI, and you can also select them randomly.
>We aim to create a system that assists people engaged in creative activities, from creators to those who have never drawn before.

P11580 kek

P11580 I think ai generated anime pics and art in general is for the better, from psycological point of view. Finally people who really can draw have a pass to say "KYS, nigger retard" to anyone who even remotely touches that ai junk or "assisted" workflow. Now every nigger tranny furry can finally create shit from shit to spread shit across the shit social "platforms". Can't wait for ai generated anime and so do you.

P11580
What does it take to do it in real time like that?

P11607
Judging from mouse movements, it does not take a lot of time, but you probably must have modern gpu, 4GB+, I think.

P11607 A couple of days? To be honest, AI sucks

NVIDIA created an AI that makes 3D models.

And now that it's obsolete, DALL·E 2 finally lets people use it without a waitlist (but with gay signup and filters).

P12308
Google also has one now:
https://dreamfusion3d.github.io/
Hopefully it's easy enough to do that we'll get an open source one soon.

Meanwhile Facebook has a video generating AI:
https://ai.facebook.com/blog/generative-ai-text-to-video/

P3497
Midjourney requires Dis[spoiler:shit]cord and also creepy.

P775
You forgot these:
> https://thispersondoesnotexist.com [spoiler: (JavaScript Not Required)]

> https://this-person-does-not-exist.com/en

P12969
> Google also has one now:
>https://dreamfusion3d.github.io/
> Hopefully it's easy enough to do that we'll get an open source one soon.

Yeah public can't use that one yet.

P15085
I got googly eyes....

P15096
Nice, it actually lets you change the settings unlike the other SaaS sites for this I've seen.

P15105
Top shelf is all I know

For those without good hardware who don't want to pay for shit, https://dezgo.com/ is getting a lot better than it used to be. There are a lot more options now including choice of models and negative prompts.

P22492
good stuff. Does one need a GPU to generate GPT-3 seeds/words?

How long until these AI drawing programs can draw text and hands?

I've been having fun with fixing a seed adding more stuff to the prompt to make it increasingly lewd.

P22680
P22681
[bold: sexy milkers]
What did you use for this...asking for a freind.

P22697
This was with whatever's running on https://dezgo.com/

Model: Anything 3.0 (anime)
Aspect: all the way to "Portrait"
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Guidance: 8 (left on default)
Sampler: DPM (left on default)
Seed: 3722215447

Prompts varying from
1girl, blond hair, blue eyes, twintails
to
1girl, blond hair, blue eyes, twintails, romantic, embarrassed, cleavage, bed, underwear, ecstasy, topless, nude, erect nipples, spread legs, panting, cowgirl position, penis in vagina, pov sex, ahegao, sweat

I think the negative terms I've been using are suppressing penises, or at least penis-in-vagina sex, for some reason.

Prompt: pov, sex, 1girl, nude, cowgirl position, penis
Model: Anything 3.0; Aspect: Portrait; Seed: 3041135178

First pic (with penis) is without any negative prompts; second pic is with these:
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry

P22740
On the other hand, maybe it's just hard to get it to draw PiV sex. I tried
pov, sex, 1girl, nude, cowgirl position, vagina, penis, penis in vagina, penetration
and this is what I got.

I generated a picture of a person under the age of 18 naked. This is considered controversial and casus beli for putting CCTV inside your house.

img2img is where stable diffusion truly shines

P22777
also ended up producing a bunch of pictures of girls in wheelchairs with unicorns of varying quality

or ponies, some of them are just ponies

P22778
Is Artchad obsolete?

P22807
Presumably not yet, being able to fix broken hands and other errors is still useful.

Is there a good way to tell it to tweak just one part of an image?

P22832
Yes you can. There are the "init_image" and "mask" command-line options.

P22930
How good of a GPU do you need to run it locally?

P8690's guide runs on CPU. That's what I used but it wasn't very good. I'll try to see if it improved or if other models are available.

P22945
I tried it, but my CPU is kind of potato too. Although I think not having enough RAM installed was a limiting factor too, so that's something I should try fixing. I temporarily added more swap, but that just caused it to slow to a crawl.

P22988
Yeah, using swap will be awfully slow. On my PC, ram usage peaks at 8.4GB and it takes around 15 to 20s per iteration on two cores. I don't think there is much you can do if your pc is too potato.

Someone on 4/g/ put together ChatGPT, image generation, and some other stuff to make a waifu AI.

https://boards.4channel.org/g/thread/90402185#p90402195

>How to make a superior AI waifu:
>
>0. Gaslight chatGPT into being an anime girl. Giving it example sentences works really well otherwise it talks like a dumb stupid lame bitch
>
>Also, make a SD prompt to generate generic images of your waifu. Do not instruct it to be in any specific location or holding anything. Just how she looks + clothes.
>
>repeat:
>1. Transcribe what you say and input it into chatGPT
>2. Ask chatGPT "Give me a list of keywords describing what is happening for an image generator"
>3. Append the most relevant keywords to your waifu prompt and generate an image based off that
>3, concurrently. Synthesize her audio response.
>4. Return both of these to the UI and show and play them.
>
>I love you /g/. Please be my friend :)

P23096
>replace Waifu pictures with historical pictures

P23096
>cucked
>fedchan
don't become some virutal girlfriend weirdo

P23133 I agreed. that's why my girlfriend is the man in the mirror.
He am very handsome...

P23096
It is truly over.

P23096
P23138
No. You will only have waifu in the Metaverse.

Has this ever happened to you?

P23137
fixed

P23157 Excuse me, I do not wish to appear controversial in this matter, are you the witch?

Trivia: the caption was written in Spencerian script

P11540
trying to make a better headpat

cute girl, arm up and forwards, perspective
negatives: hand, fingers, broken fingers, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
model: anything 3.0
seed: 4029851175

>broken fingers

poor girl P23185

P23159
All these images are so lovely. Waifu-tier stuff honestly.

P23196
No, retard, those are just some lifeless boring animu pics without any value or work spent on it.

P23196
They are cute but "waifu-tier", seriously? I feel like a waifu needs some story and character. You can't consider a standalone portrait as a waifu. I think you can make waifus out of ai-art, but you'll need to do something beyond a portrait. P22680 is good stuff, and so is picrel (too bad the guy nuked his deviantart).

P23200
sidenote: lowleg panties look great on girls, especially when they are slim and smol and cute.
I guess highleg panties also look great on girls, but that's mainly for girls who have a sweet ass or thighs and can or want to highlight them. Also highleg panties can stick out of lowleg outerwear like jeans and short shorts and they look very lewd.

Anyway, I like http://lambdaplusjs35padjaiz4jw2fugdoeutse262phqr72uf634s2wdbqd.onion/File/69/69551927c063cd1fc07f60f8baafffcfc9e6f44630357a075cfcbe2460399610.png and think she is waifu tier. I want to give her headpats and cook dinner for her. I don't really care about what you guys say.

P23197
>without any value or work spent on it

P23194
Maladies of the hands are sadly common among AI drawn people. Negative prompts attempt to cure (or hide) those maladies with varying degrees of success. Here's what she would look like without the "broken fingers" negative prompt.

P23206 the broken fingers really are the broken fingers

P23200
I like that I can generate her semi-consistently

>I can generate her
can someone generate historical men pictures?

P23205
>without any value or work spent on it
<you forgot to read the "retard" part
all gpu art is junk and boorus ban it, cope and seethe.
*headpat*

P23249
>gpu art
Jokes on you, I use CPU only

P23252

>CPU
Is that another kind of pepperoni, if you please, sir?

>CPU
>CP
MOOOOOOOOOOOOOOOOOOODS!!!!!!!!!!!!!!

P23285 "CP Unit"

P23205
Well, P23197 is definitely right about me personally not doing much work on the pictures in P23159. I just downloaded some pictures of daguerreotypes from the Library of Congress and did img2img on them. I used the Anything 3.0 model with the prompt "1girl" and a bunch of standard negative prompts, and I only discarded the few generations that either didn't have a girl in it or didn't resemble a historical photo.

lambda plus joshi shougakusei

P23302
very cute, I want to adopt her and teach her common lisp and haskell (as an excuse for me to actually learn these languages)

Stable Diffusion had a big update:
https://stability.ai/blog/stable-diffusion-sdxl-1-announcement