openai-chatter

[patent grail]

h wut that looks sick too

did u inpaint to double the size omg

[zenith crater]

yup!

[patent grail]

😍

[turbid hatch]

Love yours! I need to try more inpainting. Here’s my favorite of what I’ve prompted, an old schooner passing the event horizon of a black hole.

[hollow dune]

yoooo this is beautiful

[zenith crater]

woahhhhh

reminds me of a wave crashing over it

[turbid hatch]

What’s profound for me, is that I have a tremor in my hands, and would never be able to draw something like this without ai. Pretty damn cool.

[pulsar topaz]

What's profound for me is I have no artistic abilities whatsoever, this lets me take hand in creating things

[turbid hatch]

Oh tremor or not I’m right there with ya haha, tremor just gives me a nicely packaged excuse! Alright back to my day job you lovely, brilliant and creative people! Pleasure chatting with you all

You all have art to generate too! Better not Dilly Dalle…😀

[soft hull]

The only thing more stupid than the DALL-E Mini thing is that people are now referring to GPT-3 as OpenAI 😭

[hollow dune]

i'm not past the 50 image limit but i'm way to tired to generate more

good night everyone!

[zenith crater]

night night!

[wise wraith]

anything happening with the servers?

[modest sonnet]

LOL. I just came here to post absolutely the same screenshot. Wondering why "flat earth" is agains the policy

[zenith dome]

I can't get over how good Dall-E 2 is at food photography

[patent grail]

its so scary

[soft hull]

Try the prompt

"Gourmet meal prepared in a bucket"

It's so perfect

[zenith dome]

I'd rather not waste a prompt on it if you've already done it yourself, can you post it?

if you still have it

@soft hull

[wise wraith]

ayo-

[trim whale]

the idea, as near as i can tell, is to drive a bunch of traffic to AIs with less restrictive policies in 1-2 years once they catch up (hopefully whatever committee is setting policy at openAI realizes they're on the clock to figure something more sustainable than the hays code)

[blazing atlas]

It seems to just really want to make gorgeous food photos. I tried the "secret language of DALL·E" experiment and got a bunch of incredible, vaguely Asian dishes, including this:

https://labs.openai.com/s/GQb8yiQjmfFNXZpXFodNWhsu

[fathom star]

So, I thought of a character in my head and then described her as a prompt, then chose my favorite drawing of her. I've now used that picture to help make tons of similar pictures with my character. After I get a good set of pictures I'm going to lay them out like a comic strip or something and tell a story with it. Here's some of my favorites so far. Each picture has the original image on the left side of it that the AI model uses as a reference. Then I would type in a prompt to describe what I want the right side of the picture to be doing, and it uses the artwork and prompt together to make another original image.

thanks to @obtuse maple for the tip in how to create this!

On that top right image, I actually asked it to make a boy wearing a green shirt. I then used that image to make new ones for another character I could put in the story. These other two two look like parents. Could use that too maybe. Make some more pictures generated from these. "dad driving a car" or "mom watering plants" It would burn through a lot of prompts to make all this, but the idea is totally doable.

I used the prompt of her laying (middle of the first image), but there's lots of empty space, since it's half an image. So I move the pic of her waking up over to the left side of the image, and asked it to fill in the right side of the image with the prompt "laying"

Sorry for the spam. I'm just blown away lol

[radiant birch]

I wonder if she ever blinks 😆 but seriously this is very cool

[fathom star]

Haha, I wonder the same thing! I should try to prompt it to have her close her eyes sometime

[wicked kiln]

I did a few similar experiments to carry over style

https://labs.openai.com/s/mvEHcleyCcvR9fz7jjjcXmaF

[fathom star]

oh awesome!!

[wicked kiln]

https://labs.openai.com/s/zi6CBKFftI2A1dU3P2gvyg0i

https://labs.openai.com/s/xgmMSTUBFQ3Y8JOCpq9NIf4A

[fathom star]

Wow, that's really cool. These would be easily passable as avatars in an indie game or something. Touch them up a little bit and there goes a lot of your work load as a game dev

[wicked kiln]

I also wrote this about how I did same char in another style https://www.reddit.com/r/dalle2/comments/v1tz4l/a_3d_render_on_the_right_that_mirrors_the_sketch/

[fathom star]

🤯

[wicked kiln]

Sorry I meant https://www.reddit.com/r/dalle2/comments/v1swsh/how_to_use_edit_mode_aka_inpainting_to_change/

Not sure this works but I have many more generations for the skull boy above in the community discord #976190440137187368 message

I think the generated drawings have a lot of artifacts. I wish there was a way to improve on that but maybe the model just needs tweaking to get this to work better

I’d love to be able to say generate this text prompt in this specific style and provide an image for the style

[hollow dune]

Deception (no ongoing major/ geopolitical conspiracies)

[haughty zealot]

It’s so unfortunate that enough people believe in flat earth that openai has to take it seriously

I’d love to make some Discworld inspired images

[hollow dune]

i wanted to see Asgard similar results of earth

but yeah, very unfortunate

[smoky crag]

does she have a name yet? any lore/info?

[haughty zealot]

That would have been really cool

[fathom star]

Nope, no real back story yet, but I was thinking of just calling her Red. I'll probably just tell a story about her waking up and going to school and some stuff after school.

[hollow dune]

maybe i could do a picasso style attempt but I don’t want to get striked

[smoky crag]

Red

[smoky crag]

I see

I should probably use Dall-E to create scenes from the story I'm writing

[fathom star]

You totally could! But if it were to be commercially used it could go against tos since they own the images

[pale parcel]

I think you could legally use dall e 2 as an "envisioning tool"

[smoky crag]

the only use would be to send them in the server where my dev team is so they can better understand/visualize the scenarios

my writing is super janked so I think that would be helpful lol

[fathom star]

Woah, too cool!

I have a challenge for anyone who wants to try. I'm trying to get DallE2 to draw something that resembles Catdog. When you use Catdog as a word it doesn't seem to have any reference to the cartoon. I've tried asking it to put heads on both ends, etc, but haven't gotten any sort of animal that resembles them yet. Can you guys think of any good prompts to describe this?

Without using any sort of image reference, I should add.

[smoky crag]

hmmm

[pale parcel]

uhhh so I've got a quick question

what is this

[smoky crag]

it's broken

[fathom star]

I had that happen once!!

[smoky crag]

happened to me a few times

[fathom star]

@soft hull

[smoky crag]

I don't know the actual reasoning behind it, but my guess is it just failed to generate an image

dk though

[pale parcel]

My thoughts exactly

[acoustic lark]

@fathom star

[fathom star]

Ohh hey cool!! What'd you type?

[smoky crag]

oh interesting

mildly disturbing, but interesting

[acoustic lark]

"animal with cat head on back and dog head on front" haha

[fathom star]

"Happy animal with cat head on back and dog head on front. Cartoon network" 🤣🤣

[acoustic lark]

I love it!

[pale parcel]

it just gave me a warning for trying to do an inpaint of a dall e 2 generated image of a crowd at woodstock 😂

[fathom star]

Weird lol

I've wanted to get a snail tattoo with my mushroom. Thought I'd try to get ideas

[flint mantle]

Have a bit mixed feelings about this: https://www.cosmopolitan.com/lifestyle/a40314356/dall-e-2-artificial-intelligence-cover/
On the one hand, I think its cool if Dall-E2 would be open for commercial use. On the other hand: OpenAI felt really concerned about safety when it comes to AI but violating their own content policy here without commenting it.

[pale parcel]

is there a list of banned dalle words we can find somewhere

[scenic linden]

congrats Matt for getting dalle 2 access i watched like every single of your AI videos

[radiant birch]

That would be nice, I recently got a warning asking for a "blunt" object

[pale parcel]

Thank you!!

[pale parcel]

Yeah I was trying to generate a “special attack” for a superhero

[topaz slate]

forbidden gang sign

[severe citrus]

OUTPAINTING????

That loooks awesom

[sleek gale]

Just found out it's impossible to generate a soda can in square shape

[severe citrus]

https://labs.openai.com/s/1QE3rZHRHonMiBzSsKrtYNek

closest i could get

[hollow dune]

pandora’s jar

[pale parcel]

Lol

[hollow dune]

can someone generate a jar of dirt hehe

[severe citrus]

me opening the cloud in a bottle in terraria

[sleek gale]

Yea i got similar results with soda can in shape of a brick but dalle seems struggling

[severe citrus]

do you not have access?

[hollow dune]

i have

my computer is 5 meters away and im sitting in a comfortable couch

[severe citrus]

use it on your phone

[hollow dune]

oh right

bruh

prompt is “A jar of dirt, cinematic shot”

but looks more like advertisement of DRKT than dirt

[severe citrus]

https://labs.openai.com/s/dQNKxvg5Iv3L8uZXNst9anoP

[hollow dune]

damn

this is beautiful

[sleek gale]

Amazing

[hollow dune]

Impact vibes

[hollow dune]

I have an idea:

[steep ginkgo]

ooo i love ideas

[hollow dune]

Every time someone joins and presents themselves in #introductions someone should say “Hello there” to which they should respond with “General Kenobi”

[steep ginkgo]

anatoly i just read your intro how are you finding a.i?

[hollow dune]

i don’t understand it, it’s something like a black box

you have input, you get output, you can train this box to get certain output for certain inputs, and it finds the patterns

i find it ‘hard to debug’

you can’t say “oh this neuron has this importance” like you can say “oh this line stores this variable into here”

[steep ginkgo]

yeah it can feel like that sometimes

[scenic linden]

all we know is it grouped data into high dimensional space where similar things are closer together

[hollow dune]

i see you work with ai and robotics as a hobby

do you know any good tutorials with tensorflow since i want to stick with js?

[steep ginkgo]

with tesorflow.js

?

[hollow dune]

did you work with tensorflow?

[steep ginkgo]

yeah

[hollow dune]

:o

i tried making a short tool that recognise wether my message is a goodbye or hello message

[steep ginkgo]

nice nice

[hollow dune]

trained it with ca. 20 phrases and around 500 cycles and it seemed to work

when i had something like "hello and bye" it would mark the sentence as both which is accurate XD

the universal sentence encoder is very powerful, but as i mentioned previously, a black box that i don't comprehend

[steep ginkgo]

dude ai is alot of math and statistics

[hollow dune]

yeah

but there's a difference seeing 6 neurons and working them on paper and seeing a multi billion network like dall-e on screen

[steep ginkgo]

yeah for sure

[hollow dune]

i was at a critical thinking course parallel with ai, and we had learned things as backpropagate and feedforward, which i partially forgot and on paper created a xor ai

[steep ginkgo]

nice nice

here in rural aus its hard to do things like those courses

[hollow dune]

i see

wait

are you the dylan i think of?

[fair gull]

hi

[hollow dune]

oh hi

[fair gull]

i dont think so

lol

[hollow dune]

you are another dylan XD

[fair gull]

haha

[hollow dune]

i thought you are the dylan from the rust server

[fair gull]

i do play rust lol

[hollow dune]

no no

not the game

[fair gull]

ahhh

[hollow dune]

this maths stackexchange just gave me an idea on what we can try dall e to generate

[fair gull]

i didnt know the best place to share this but i was just coming on here to say I and a couple other Dalle2 users put together a public discord to share creations as well as give people without dalle a place where they can request prompts/send in ideas.. it grew to well over 1,4k members in a week and steadily growing. We'd love to welcome new faces in! We are getting out numbered and overwhelmed by non dalle users 😅

[hollow dune]

😅

[hollow dune]

that sounds cool! i myself am already overwhelmed slowly by the dalle2 subreddit, i wonder if you can share it though, let's wait for the mods to answer i suppose ^^

[fair gull]

yeah it would feel wrong for me to drop the link in here without permission but if you want me to send it to you dm me

[lost wadi]

did anyone else burn through all of their prompts in about an hour? this AI is incredible

[hollow dune]

i haven't yet

[teal flame]

hey yall . just got my invite yesterday and im still navigating dalle. can someone please explain to me how to use the inpainting tool to increase the size of my images?? im struggling to figure it out and im on my last few generations for the day hehehe🫣

[steep ginkgo]

yes

[obtuse maple]

You’re going to have to wait until tommorow then, inpainting can use up quite a few prompts. Check out the tips and tricks section for more details.

[grand onyx]

I highly recommend digging into some philosophical or metaphysical concepts and asking Dall-E to generate stuff based on that. You get some really interesting visuals.

[hollow dune]

That's beautiful!

[burnt mango]

hey guysssss !! whatsup

[hollow dune]

hi!

[burnt mango]

Anatoly! How's you

[hollow dune]

having fun with Ryan in memes

how are you ^^

[burnt mango]

Planning to join you guys

[hollow dune]

XD

welcome!

[zenith crater]

@pale parcel omg welcome!

i remember talking in the midjourney discord

[median wigeon]

anyne else having issues with cellphone verification

[patent grail]

@uncut wadi your art is amazing

[uncut wadi]

aw thank you, i appreciate that 💗

[hybrid yew]

Just a tip for those interested

Add your Twitter link or any other links

[finite ridge]

But I like my about me section

[pale parcel]

Yess!!

I can’t believe I blew through all of my dall e 2 prompts last night I gotta make a video today 💀

[zenith crater]

i did my first night too 😭

and my third

and fourth i think

[pulsar topaz]

https://tenor.com/view/panda-office-rampage-mad-angry-gif-12918172

When you still have 14 minutes until your next query ^

[patent grail]

LOL

[patent grail]

yeah it feels much nicer to space them ou t

the day i got in i did them all in an hr

and then the next 23 hr were this lol

[zenith crater]

you're literally insane

it took me three hours my first night 😭

[pulsar topaz]

Yeah, my first took 3-4 hours too. I've used every single query though to date.

First night I tried all kinds of different things. Now I'll occasionally try to dial something in with 3-5 slightly tweaked queries which burns through them quickly.

I wasted 5 on the theme today, another 5 on Thomas the train, then went through 4 variations of a Halloween query tuning in what I wanted.

[zenith dome]

Please, OpenAI, delete my daily cap

[pulsar topaz]

I wonder what their actual cost is. I'd probably be willing to pay $15 a month for 100 queries a day.

I'd probably go $10 for 50 a day

[zenith dome]

I'd rather pay by the prompt

With no limit but my bank account

[pulsar topaz]

Paying by the prompt I'd never want to query anything

[zenith dome]

What if it was only 1¢ per prompt though

[pulsar topaz]

"is this query really worth x units of money?"

With a monthly fee though I'm more like "ah screw it, send it"

[pulsar topaz]

That would be double my $15 a month for 100 a day

[pale parcel]

It’s probably only going to be like a few cents per prompt

[pulsar topaz]

the cost per query would be double

3000 a month at $15 would be half a cent a query, I wonder what the actual cost is though.

[zenith dome]

Still, I'd rather not have a ceiling. I just want to explore Dall-E 2 without worrying about "wasting prompts"

[pulsar topaz]

At a few cents per query I'd want a larger canvas and a semi-reliable way to make sure it doesn't add gibberish to images

[zenith dome]

I think I would still feel limited by 100 per day

[patent grail]

pay by prompt would be stressful ya

i like what that one person said like if u get 300 prompts a day it feels better

[zenith dome]

I would be fine with anything over 200 per day tbh

I feel like that gives enough room for experimentation

[patent grail]

yeah

and it almost encourages u to use it more

[pulsar topaz]

Exactly. I'll be less likely to stress about wasting a query, but I'll be more likely to consistently use the service

[patent grail]

yeah 300 a day is a full hour of dalle lol

if ur doing nonstop prompts

[pulsar topaz]

And for my personal use case, I'd mostly just use the images for blog posts and my social media feeds. A set number a day would let me sit down and run with a few ideas each day and build a catalog of images I want to use that day, or in the future.

Which is basically what I'm doing now. I save about 80% of my queries and catalog them into their own folders for future use

[zenith dome]

But also, this is how GPT-3 is set up currently

[pulsar topaz]

Yeah but that's something where you're less likely to want to generate lots of something.

With images you're going to be making a bunch of tweaks to get what you want, and may want to create multiple sets of the same query

[patent grail]

yeah

also gpt is like insanely cheap

i doubt dalle will be that cheap

[pulsar topaz]

Yeah, I feel like text generation is several orders of magnitude easier to do

[zenith dome]

Maybe they should make 2 payment options.

1. Monthly premium, where you pay $15 or $20 per month to raise your image cap to 100 or 150
   or
2. Unlimited premium, either a monthly subscription that costs more than option 1 or pay by prompt, either with no limit on prompts per day

The problem I have with a monthly subscription is I would feel obligated to use Dall-E 2 as much as possible even when I don't want to. Paying by prompt would just let me pay for what I use

[patent grail]

Btw I noticed something similar, I was speaking to someone who does hair for a living and asked Dall-E for some examples of dreadlocks.

it seems to struggle w the idea of dreadlocks too if you zoom in its kinda weird straight hair 🤔

[zenith dome]

Also there would be days I wouldn't want to generate 100 prompts and other days where I might want to generate a lot more than 100

[patent grail]

Even after I specified where hes from, it wasnt a great result. Maybe its a limitation in training data D:

they werent all bad tho this was a good one

[hybrid granite]

I'm personally fine with the 50 per day cap

If the payment system is something where you choose how many generations you want, I can't imagine paying for more than that

I might even pay for less

[zenith dome]

Personally, I would be satisfied with 5 cents per prompt, but I'm guessing I'm at the higher end of the demand spectrum. 1 cent per prompt would be an absolute dream

Just as long as I'm not limited by a cap

at least, not a cap anywhere below 200 per day

[void glade]

You seriously expect to pay less for dalle then what Midjourney costs? Midjourney charges $30 a month for 15h of GPU use, which translates to about 900 images . You want to create 600 images a day (100 prompts) for $15 a month?

[zenith dome]

midjourney also isn't a very big company. Economies of scale must be considered

[void glade]

I expect it will cost several cents for one image

[zenith dome]

I seriously doubt that

[jovial lava]

I have been doing prompt for hours now, and it doenst seems to stop at 50, i am the only one ?

[void glade]

There was a discussion about pricing with OpenAi team here on this discord some time ago #openai-chatter message

[hybrid granite]

I think that's a very fair point

And one that deserves discussion

[void glade]

On night Cafe it costs 20 cents to create one image, the quality is much lower, the resolution is 400x400 pixels, and there are plenty of people willing to pay for this

[zenith dome]

I don't know how much electricity a single prompt takes, but I would imagine it would be power of the GPU times the time it is used. So in this case, maybe 1,000 Watts * 10 seconds, and at california's energy cost, that's only 0.05 cents

[hybrid granite]

What I consider a fair market value for a generation isn't what I'd personally pay

I'm not paying anyone just to make funny Totoro pics

Or generate yummy-looking food pictures

[zenith dome]

that's not 5 cents, that's 1/20th of a cent

then again, maybe I'm underestimating how much energy it takes

[void glade]

it's not about electricity usage, check out the price for A100 GPU. If you use it on google cloud it costs $3 an hour

assuming it creates a single image in 20 seconds, that 180 images for $3

[zenith dome]

so pretty cheep

10 cents per prompt

at a significant mark-up

whatever google's markup cost is

Lamdalab's A100 cost is 1.25 per hour

so less than 5 cents per prompt

actually, wait no, 1.10 per hour

so like 4 cents per prompt, assuming it takes 20 seconds to generate, using the entire A100 for that time

2 cents for 10 seconds

[void glade]

Ok, this is all just wild speculation , but assumig 4 cents a prompt x 100 prompts a day = $4 . For a whole month thats $120 for mere computing costs. Plus salaries for employees, cost of creating and training the model (100 000 hours if I'm correct) plus some profit for OpenAI

[zenith dome]

profit is included in the 4 cents per prompt

that's what Lambda cloud charges

and hence, I wouldn't want to pay $120 per month. I'd rather just pay for what I use

so yeah, I would pay 5 cents per prompt to use Dall-E 2. That's only $5 for 100 prompts, which I think is a bargain

[uncut wadi]

is it possible to refer people for access?

[zenith dome]

Put that in #community-help , it's more likely to be seen by someone at OpenAI there

[finite ridge]

Excuse me, but none of these bones are broken

[pale parcel]

Keep in mind everyone that $30 a month for midjourney is for unlimited use

You get 15 gpu hours of “fast generation” for a month and after that you switch to a free slower mode which is only about 2x slower still generates in under a minute

[patent grail]

I wonder if most of their money will come from corporate contracts

Like companies would pay staggering amounts for this

Imagine how much they'd save

[pale parcel]

Microsoft has a huge investment in openai

[trim whale]

i know myself too well to sign up for a per-prompt plan

[thorny ether]

Art is dead. Long live art!

https://twitter.com/TimSuzman/status/1540392596911947776?s=20&t=7Ye2Ikfjvr4jBawXlyqITA

Tim Suzman (@TimSuzman)

Art is dead, long live art!

#dalle

1/

Twitter · Jun 24, 2022, 5:53 PM

[finite ridge]

I thought it was an endowment, not an investment?

[pale parcel]

Tbh I dont know the difference ur probably correct

Most of openais money comes from Microsoft I believe

[finite ridge]

An investment would mean that ms would have access to their results.

[pale parcel]

Orangutan watching the sun rise eating a bagel

[zenith dome]

I wonder when the next Dall-E 2 update will be

[zenith crater]

You mean DALL•E 3?

[patent grail]

a dalle changelog would be so funny lol

all the weird specific things theyd have to patch

[fair gull]

Hey all, I just got approved from the staff to share this!

I, as well as a couple of other Dalle2 users, put together a public discord to share creations and give people who don't have dalle a place where they can request prompts/send in ideas. We want non-dalle users to feel included while they wait on the waitlist. As well as introduce, familiarize and answer questions about dalle for users who will get it in the future. It grew to well over 1.4k members in a week and steadily growing since.

So far its gone smoothly and we've built a great community. Right now its growing so fast the dalle members are a tad overwhelmed with the amount of non-dalle users and would love to see new faces to chat with and help fulfill prompt requests. I'll drop the link below and hope to see you all there!

discord.gg/dalle

some openai staff have joined as well

[blazing atlas]

Yikes. That's surprising to me and disappointing.

[slender hornet]

hello guys! i am new here.

[slender hornet]

i do too

so im wondering: is it possible to use Dall-E on a mobile phone? tried logging in on mobile and it just gave me this:

Something went wrong
We ran into an issue while authenticating you. If this issue persists, please contact [e-mail].
so i assume its just not supported at the moment?

[blazing atlas]

It is

I used Chrome to turn it into a "web app" with a shortcut on my home screen and it works well with limited functionality (no history view)

[slender hornet]

hm. could the issue be then that I tried doing that on firefox?

maybe i should try chrome

yep. worls on chrome. thanks.
weird and unfortunate that it doesnt work on firefox, vut does on chrome, but thats fine haha. at least i can use it on mobile now.

[sterile bluff]

NightCafe has done this for a long time

they're thriving

and regarding too low quality - do you mean resolution, or composition/aesthetic?

no, it's more than style transfer

they've had text to image for both VQGAN+CLIP and CLIP Guided Diffusion for a long time now.

I disagree whole-heartedly. 🙂

I'm talking about the "not worth saving"

Humor me, what do you think of this?

Neither of us know what DALL-E 2 is run on, hardware wise. We can both think what we want; we both might be wrong.

One good output, huh?

thanks, what part describes the hardware it runs on?

cxmu -

1/1

8 Acknowledgements
We’d like to thank Jong Wook Kim, Hyeonwoo Noh, Alec Radford, Pranav Shyam, and Ilya Sutskever for
helpful discussions and contributions to our work. We’d also like to thank Yunxin Jiao for creating several
figures used in the paper. We are grateful to the Acceleration and Supercomputing teams at OpenAI for their
work on software and hardware infrastructure this project used.

that's the only thing I found.

I'm trying to engage with you in a way that is honest.

I'd like the same respect in return.

You're being super dismissive of me for a reason I can't discern.

Is a GPU with 48GB of VRAM a Commercial GPU?

Because that's what I suspect they run it on.

That's an assumption, and I'm sure one day we'll find out.

[blazing atlas]

Per-use model is not really relevant anyway. As a tool for ideation it already has value as a cloud-based subscription service like Adobe's suites

[sterile bluff]

In the mean time, there is no reason to be so dismissive of fellow members of this community.

That's multiple GPUs, isn't it?

And it doesn't state what kind of GPU is it

or if it is serialized, parallelized or any specific type of pattern

Mj, NightCafe, these also run in server farms with hundreds of GPUs.

Most people wouldn't call A100s, A5000s off the shelf components.

Neither you or I know what it runs on.

[patent grail]

I'm never going back to craiyon Lol

The quality here is absolutely insane

Id pay like $500 a month for this Lol

and i have no usecase or income 😂

[sterile bluff]

@lilac pond I believe that there are hundreds, possibly thousands of DALL-E 2 runtimes living in a huge cluster. Each of them has a dedicated GPU that the model is loaded to at runtime.

They have a queue based system where the request is sent to an available runtime, or pends until one is available.

Nothing in the documentation you linked earlier discusses or references serial or parallel processing.

So we simply don't know.

[grand onyx]

I realized prompts that I run yesterday give different results today, even different style as well. Its quite interesting

[sterile bluff]

yeah, we don't get to see the "seed" so it is hard to know how to replicate results

[grand onyx]

every day the seed changes or is it with every prompt?

[sterile bluff]

@grand onyx further, even with seed I've seen in other tools that to date there is no deterministic way to control results

I'd assume every prompt

[zenith dome]

wdym by seed? You mean the initial state of the diffusion map?

or do you mean its coordinate in the latent space

[sterile bluff]

yes, a majority of these tools have some sort of "seed" which governs those attributes

could be both.

[zenith dome]

that's impressive

[grand onyx]

I feel like theres a seed for the seed generation, and maybe that seed changes daily and with it how seed generation is made

[sterile bluff]

@lilac pond go on - that sounds like it would be helpful for folks to know.

all right, I'll ask differently - what do you mean by prematurely?

[blazing atlas]

It's re-running an incredibly complex algorithm/model each time, where random noise is afaik part of the process. I don't think you should expect to ever replicate results identically

[grand onyx]

dall-e 2 is only 3.5 billion parameters

[sterile bluff]

right, and in many similar tools that randomization goes down to to the hardware level e.g. even factors like temperatures could play a part

help me understand how stating that gets you to a raw state, as opposed to just it giving you an artistic rendering of what what CLIP visualizes "encoded" to be

and sorry, I said CLIP but OAI refers to it as unCLIP. 😄

you've said intermediate and premature, but these are all part of the final six (was 10) outputs, correct? there isn't any type of in progress view is there?

'cause if there is, I'd love to see it.

if this is from a specific webpage on how it renders please link it because this is interesting.

[blazing atlas]

I think this is an interesting leap of logic. These colours are used in print medium to blend and create colours as process, but that wouldn't be required by a digital tool - it would be a waste of resources to try to "layer" colour this way to get the desired colour. This particular image seems like an emulation of traditional half-print

[sterile bluff]

yeah, it is interesting when we see what appears to be part of the diffusion noise as a final result.

[blazing atlas]

So it's creating a very vibrant base diffusion pattern with those colours for aesthetic preference reasons

[sterile bluff]

it might even be that it's been determined that is an ideal diffusion noise pattern for iterations on

I use a number of noise patterns for DD and other colab notebooks

some of them aren't so different from that, especially with low skip steps and init scale settings.

Mj absolutely juices generations to give them a very stylized look and to improve composition. It is possible DALLE 2 has some of the same.

by incapable of transforming it further, couldn't it also just be that the generation abends early, and instead of throwing an error because it didn't hit X number of steps/iterations

it drops the currently rendered "frame" to the output process like any other generation?

[blazing atlas]

I guess we have e.g. CMYK in printing because it's a base that can be tweaked most efficiently to create a full spectrum of powerful colour

[pale parcel]

Yeah its like trying to upscale the ungenerated noise

[sterile bluff]

right, but why couldn't it throw an error? I feel it's safe to assume that watermelon tea isn't represented in captions with that image above

so that is interesting to me - do you have the exact prompts for those two images?

hmmm. I don't know if color array is decisive enough where we wouldn't see an image that was in a similar state as part of training images

but I get your point, and it's very interesting and worth researching if we had a way to see in-progress images.

[blazing atlas]

the "watermelon tea" one is a bit more convincing

[sterile bluff]

yes

https://labs.openai.com/s/w0Kc8SIPjYKZUGi8LOhnecsX

the other 5 were what I expected; watermelon in tea, watermelon next to tea

[blazing atlas]

Different colours though. So either is just has a weird, esoteric idea of what "watermelon tea" is, or we're seeing that the colours themselves are arbitrary and generated early in the process with strong palette contrast and saturation

[sterile bluff]

what is just as interesting to me is

why would this one phrase cause so many "abends"

[blazing atlas]

looks as expected

te cat?

[sterile bluff]

or, the training data for that type of caption is poor

not representative

seems unlikely, but

lends a little credence to what I just mused upon

[pale parcel]

I think that’s a coincidence

I think when you get those colored outputs there isnt a gpu available to generate that noise, so it never generates then gets upscaled with the rest of the images

[sterile bluff]

oh, interesting

[pale parcel]

Just a hunch

[sterile bluff]

so if I understand

you're saying that for UX and SLA

it will always try to get your result out within a specific interval

whether there are 6/6 done or less

[pale parcel]

I could be wrong that’s just my guess

[sterile bluff]

@pale parcel which then goes back to what I mentioned earlier, that Mj "juices" images

and it is possible DALL-E 2 does the same but in a different way

e.g. having optimal initial diffusion patterns

[grand onyx]

Close up photo of a [Insert animal here], drinking water from a lake, bokeh, 100mm lens, 4K award winning nature photography.

This one is the best prompt for animals

[patent grail]

diagrams are my favorite

[pale parcel]

Yes. It is my understanding that midjourney may be more efficient, only 400m parameters and sometimes in my opinion atleast, it competes with dall e 2 which is 10x the parameters

[sterile bluff]

Mj does seem very fine tuned

[patent grail]

thank you im saving this

[grand onyx]

Tell me how it goes

[sterile bluff]

saving it, I'm doing it now but typing in Spongebob instead of an animal

[patent grail]

LOL

oh god

[grand onyx]

Spongebob what?

lol

[sterile bluff]

Squarepants, of course. 😄

more glorious than it should have been

[grand onyx]

he should be drinking the water

[sterile bluff]

I changed it

sorry

also, not sorry.

[patent grail]

Lol

[grand onyx]

i wonder how it would deal with patrict

[patent grail]

LOL

[sterile bluff]

so we maybe sorta think but also have no idea if DALL-E 2 is upscaling noise as end results because the render abended, or there are specific inits that end up never being generated but upscaled and sent

I see failures from time to time with NightCafe. Very rarely with Mj

its interesting when it happens.

switching gears - y'all ever explore the latent space within these tools? DALLE2 doesnt allow weights, but if you type a really long string of numbers with some calculation signs

you get very random results.

people, places, unusual objects combined.

[grand onyx]

Two amazing servants with long red hair using hand fan on Spongebob sleeping in his bed.

gonna try that prompt tomorrow when my limit is reset :x

[blazing atlas]

I think they would always be squares, but the upscaling makes them un-square, more naturalistic. Look closely e.g. at what it outputsfor a "waveform" - almost painterly peaks, not sharp data

by squares we mean pixels, i guess

[patent grail]

often?

do you run this daily?

[blazing atlas]

Yes. Perhaps on a whim the model decided this particular noise looked especially like a mosaic

[patent grail]

they shouldve used that example in the paper

it makes such good anime sometimes

ig theres just too much going on there ;-;

[grand onyx]

Theres something about Dall-e 2 that is a bit weird

Yeah anime are okayish

but dall-e 2 knows nothing of starcraft 2

lol

i asked it to generate a photon cannon from starcraft 2 game

it didnt

I have used dall-e mini for quite some time before Dall-e 2 and i found something dall-e mini excels at against dall-e 2 significantly

[blazing atlas]

I don't think it's copyright, I just think they were absent concepts in the data set provided

[grand onyx]

This is a chalkboard drawing of a grizzly bear from Dall-e 2

This is a chalkboard drawing of a grizzly bear from Dall-e mini

notice the difference?

[patent grail]

why is the dalle one so blurry

[grand onyx]

Thats the only prompt that seems better than dall-e 2. Everything else, Dall-e 2 is better at

except diversity

thats pretty much it

This made me realize how much dall-e 2 still has room to grow in the making of dall-e 3

its going to be very interesting

3.5 billion parameters maybe is a limitation. I could hardly imagine if they doubled this.

[blazing atlas]

Parameters and input data / training are different though, in my understanding?

[grand onyx]

If i remember correctly, dall-e 2 is trained on 650 million images with their captions. But i always know that the larger the number of parameters the better the A.i can perform

Is it possible to make the number of parameters smaller than this without sacrificing quality?

3.5 Billion parameters on 650 Million images

Theres 750 Billion images on the internet

So Dall-e 2 is trained on 0.075 percent of all images on the internet which is nothing

Get hyped for Dall-e 3

[zenith crater]

bestie!

[grand onyx]

Haha Pashahlis is in this server as well, this guy is everywhere. :D

[fading bane]

@knotty jetty would you mind reading my DM, im in contact with corridor about DALLE

[unreal yarrow]

Are we allowed to leave our instagrams where we show off Dalle art in here?

[slender hornet]

okay first day is over. spent all 50 prompts. i had so much fun. its exactly what i thought it is.

it is absolutely amazing at creating anatomically correct good looking cool characters. absolutely love it. cant wait to continue tomorrow.

also spent around 20 of those 50 images on generating images for friends and to fullfill requests.

i tried generating

• Aloy
• Korra
• Ahsoka
• Rayla
  as characters but none of them looked like their acrual counterparts. They certainly looked inspired by them though.

That confused me because I have seen so many accurate pictures of Darth Vader around here. Maybe its because those characters dont have a lot of images associated with them in the dataset. Or its because I didnt specify an art style. Or its because its not a photo as art style, while all those Darth vader pictures seem to be photos. In any case, I shall continue testing this tomorrow.

[zenith dome]

I've found that characters have to be extremely famous for them to work in Dall-E 2. Darth Vader, Homer Simpson, Batman, and Superman all seem to work pretty well

[blazing atlas]

it's about iconography. Vader and Homer Simpson work well because they 1) have S-tier media representation 2) have iconic designs that can be distinguished even in a few broad strokes

[patent grail]

Yeah I agree

avatar the last airbender is a no-go

ive wasted dozens of prompts 😦

i guess the ai really doesnt like to copy, it needs a lot of examples

probably a good thing

[unreal yarrow]

You can add “screenshot from avatar the last airbender” as a prompt and it will create something in that art style

[patent grail]

yeah

i cant get any characters tho

ill try that tomorrow though

have you tried that does it work

[unreal yarrow]

Yes, Dalle recognizes cartoons well actually. Once you format it as “Bear in a suit, screenshot from Rick and morty” etc

[patent grail]

oh shooooooooot

thank you so much

god there needs to be some database of this knowledge

[unreal yarrow]

There is actually

That’s where I learnt it from

[patent grail]

ohhhh

[unreal yarrow]

These guys created a google doc

[patent grail]

omg

[unreal yarrow]

Of prompts

[patent grail]

i have to read it

[unreal yarrow]

I’ll send it

[patent grail]

is it lucas'

i think i lost the link lolol

hes gonna roast me

[unreal yarrow]

Not sure what the server rules are on sending links so I’ll dm you it. I don’t wanna get banned

Actually I think it should be fine

Prompts
https://docs.google.com/document/d/11WlzjBT0xRpQhP9tFMtxzd0q6ANIdHPUBkMV-YB043U

[patent grail]

omg all good

yeah openai knows about the doc

they like it lolol

[unreal yarrow]

Niceee

[patent grail]

tysm

[unreal yarrow]

Gotchu

[unreal yarrow]

I just did these by using this prompt

[patent grail]

wowwowow

amazing

those are top notch

[unreal yarrow]

Ikr 🔥

[zenith crater]

you’re safe for now

[civic wadi]

How many images do you guys have saved in your collection? It’s only been two days of access for me, I can only imagine how many I’ll have over a week lol

[tardy granite]

I have 1000 and am constantly going back and deleting saved images so I can save more haha

[zenith crater]

72, been here for like 10 days

[grand onyx]

you can also bookmark your fav 6 pack generations and save them for after the 1000 limit is taken away, so you keep a polished and not full collection

[patent grail]

lmaooo

i have 12 🙂

only one made it in today

rough day tbh

[smoky crag]

seriously surprised people have hit 1000 already

I barely have like 25

[tardy granite]

I've hit a number of 'sweet spots' that have been generating just the most ideal and perfect inspirations, and that was before I started generating Lord of the Rings scenery, those've taken up a lot of saves, tbh

[proven linden]

I have like 40 something saved

[patent grail]

i like going back when im on the delay

[radiant birch]

I know this is clickbait but I cringed soooo hard https://screenrant.com/seinfeld-cast-cgi-characters-images-dall-e/

[zenith dome]

I don't get why there's so much hate for Dall-E mini/craiyon in the Dall-E 2 community. Craiyon is great

Sure, it's not as detailed or well composed as Dall-E 2, but it's been trained to generate pop culture, and it's good fun

craiyon actually makes me really hopeful (and a little scared) for the future of this technology.
Some day, soon, there will be an image generator as good as Dall-E 2 that has been trained with the same kind of data as craiyon, it will understand any character you demand of it, at a quality as good as Dall-E 2. For all the care that was taken in curating Dall-E 2's training, and the policies geared towards its safe and ethical use, I fear it's all in vain; it's just delaying the inevitable

[blazing atlas]

Just burned a couple of tries for the daily theme because the model doesn't seem to know what subtitles are...

TBH it was a terrible idea. And yet... 🥺

[fair gull]

Hi i got approved from the staff to share this!

I, as well as a couple of other Dalle2 users, put together a public discord to share creations and give people who don't have dalle a place where they can request prompts/send in ideas. It grew to well over 1.5k members in a week and steadily growing since. Right now its growing so fast the dalle members are a tad overwhelmed with the amount of non-dalle users and would love to see new faces to chat with and help fulfill prompt requests. I'll drop the link below and hope to see you all there! Some openai staff are actually in there.

discord.gg/dalle

[lost wadi]

i keep triggering the filter with stuff like "speaking to the dead" instead of ghost

[soft hull]

Me when i dall e two

[blazing atlas]

Me:

https://labs.openai.com/s/uBljpxV1pNqqySg8Etxj70OU

[slender hornet]

btw ive seen people include "star wars" in their prompts successfully, yet when i tried it, it didnt work? any idea why?

[soft hull]

"Wars"

"Wars between logic gates" didn't work for me so it's probably that

[blazing atlas]

it may have an exception for "star wars" - or users get around it by e.g. "a new hope"

[slender hornet]

@soft hull but someone else posted a picture including "wars" today https://labs.openai.com/s/ru6hIJdsEdaBSV9lhCAe58xe

[soft hull]

idk maybe it's a skill issue

[slender hornet]

haha

[blazing atlas]

i'd try: ahsoka from star wars, animated season 7

[soft hull]

Okay okay good news

DALL-E 2 knows who walter white is

[slender hornet]

alright ill try that later today thank you

[blazing atlas]

you could even throw "clone" in without the wars arbitrarily - "clone animation". it should get the gist if it's in the training data

[slender hornet]

tho i doubt ill get an Ahsoka looking like actual Ahsoka with any amount of prompts lol

been trying to generate a few characters yesterday and they all didnt look like the characters, just close to them. the weird thing is that the fact that they did look inspired by them means that the AI absolutely has images of them in the database, et somehow still foesnt get it right lol

[blazing atlas]

there's a good chance it has a couple of poor reference images to work with. you can't really get around that. i had the same issue trying to create a "diprotodon" - it obviously had a couple of weird reference points that don't align well with the common conception

[slender hornet]

so basically what youre saying is while it foes have reference images of those characters, they might all just be badly drawn fanart, which results in it knowing what an Ahsoka is, but in the wrong way?

[blazing atlas]

Yeah, that's my guess

Fewer points of reference also mean much less flexibility

Even if they're decent quality

[slender hornet]

i see

makes sense

[slender hornet]

@blazing atlas i havent tried "cosplay" yet. i wonder of that could help. there was a good "cosplay-like" generation of Princess Mononoke on the subreddit the other day

tho it didnt use the word cosplay in the prompt

[blazing atlas]

It tends to look like cosplay anyway when you represent mostly human characters in photographic contexts

[zenith crater]

someone just requested me to do "my little pony protest"

[fading bane]

Here is another mindblowing project from OpenAI: https://openai.com/blog/vpt/

[pulsar tree]

Y'all see this? "All art is borrowed." https://www.cosmopolitan.com/lifestyle/a40314356/dall-e-2-artificial-intelligence-cover/

[zenith crater]

Yup!

I really like that article

[fading bane]

Here is her video of it: https://www.tiktok.com/@karenxcheng/video/7111742225436003626?is_from_webapp=1&sender_device=pc&web_id=6927370540533712390

[fading bane]

I like the article, but i dont agree with all her statements

[trim whale]

i THINK it ought be possible to create and manipulate original characters in dalle2 as well through some variations of the mirror trick and partial inpainting, but it would require a lot of cycles so i'm gonna have to set aside a day to really sit down w it

[slender hornet]

mirror trick?

yo wait its you

from the MJ discord

[trim whale]

yes i didn't find out about this one though someone else posted it

trying to find the post

https://www.reddit.com/r/dalle2/comments/v1tz4l/a_3d_render_on_the_right_that_mirrors_the_sketch/

supposedly you upload a character in one medium, taking up half the square, and can explicitly tell dalle2 to mirror in another

from this, i assume but can't test until later today that perhaps you could have the same character in different poses or expressions

[soft hull]

Are we allowed to generate guns

it appears not from message history

[grand onyx]

Wait, I can now only generate one image every 15 minutes?

lol

even though the 24 hours are up and has been reset

1 image every 15 minutes?

ah its not a total reset, but the time per prompt when its executed?

[slender hornet]

im sorry im not sure i understand?

[trim whale]

yes its 50 in every 24 hour period

they roll off and roll on so you may have irregular waiting times if you used it throughout the day

[tawdry maple]

i feel like a junky that needs his fix, waiting for my next 50 generations :)))

[zenith crater]

you get 50 generations, and each of those has a 24 hour rolling cooldown; let's say I have 50 free ones to use, and I use them in 4 hours. I'll have to wait 20 hours to be able to do more. does that make sense?

[grand onyx]

ah thats so weird

lol

[zenith crater]

i like the system tbh

keeps me in check

[grand onyx]

so i have to place 50 prompts on a note pad

and speed run through

[zenith crater]

you don’t need to but it makes it easier to manage

[grand onyx]

speed running dall-e 2 on a notepad because of the 50 prompts per 24 hours

haha

i thought it resets back to 0

all of the 50

[trim whale]

yeah the two biggest QOL improvements to dalle2 would be a prompt counter and custom aspect ratios

(yes i know you can composite but that's a huge pain compared to --ar)

[soft hull]

Prompt counter would be good.

You know what would be better?

RESETTING THE DAMN PROMPTS AT MIDNIGHT!

It'd make things so much easier

😌

[zenith crater]

openai servers: no ❤️

the traffic would be mental

[soft hull]

Adjust it for time zones

[zenith crater]

openai servers every hour because of timezones hitting midnight

[patent grail]

dlale is amazing im literally in awe every day

😭 😭 😭

[slender hornet]

@trim whalei assume unlike Midjourney it is not possible to use a combined image + text prompt?

i know you can edit the image, which is sooort of like that, but like you cant generate completely new variations of that image with an additional text prompt supporting it, right?

[trim whale]

nope

you can't lock a seed either which really helps for fine-tuning when something is so close

[slender hornet]

still i think the initial character generations and the way Dall-E understands text are much better than MJ

but im having fun with both

[hollow dune]

aight i joined

the reactions are weird

might leave though, i have too many discord servers anyway

but seems to be an organised discord server

@versed terrace when is professional hacking ethical if i may ask? like what's the job; companies hire you to test their security?

sounds like a cool profession ^^

[grand onyx]

Is there a way we can know how many prompts we have left out of /50?

[hollow dune]

it's been suggested a lot

[grand onyx]

i dont want to go on excel and do it manually

it would waste my time for something so simple

[versed terrace]

Right you are. Companies pay me to conduct pentests on their websites/infrastructure. In the contract its defined what I can do and can't do, and from there I look for security issues. These can be as simple as a site letting a user delete information belonging to other people, to total access to their servers. Then I compile a report, tell them how they can fix these issues, and help them through the process of doing so.

[hollow dune]

wow

this sounds cool!

[versed terrace]

There's also incident response, which is when companies are hit by malware and it then is my job to understand what the malware is doing so we can either prevent it from spreading, reverse engineer it to try to undo its damage (not usually possible sadly), and to hopefully catch whomever did this

[hollow dune]

so you are more "security overseer" than "security crasher"

it's just there's the common stereotype around hacking that it is to only destroy security

[hollow dune]

isn't it possible to store information changes in a log, to try to prevent possible damage on servers?

that would be storage inefficient, but safe

[versed terrace]

Well, its through knowing how an attacker thinks that you can stop them. Cybersec is now more important than ever IMO. And sadly its not just peeps with crashers and malware that you gotta worry about. Whole governments hack each other

[hollow dune]

yeah, can't get political due to rules, but i totally understand and know what you mean

[versed terrace]

Well yes and no. Logs can help in storing changes, backups can help safeguard information, but these are not fool proof. Sometimes they're the source of bugs themselves. The infamous java log4j was a logging system that allowed people to hack java applications directly via its logging

this included minecraft

[hollow dune]

cyber warfare is a pretty new thing, but this is where most seems to be done

[hollow dune]

wasn't the problem with log4j remote access more than it being a logging API?

i must admit i haven't followed the scandal on a deep level

[versed terrace]

It was the logging itself that was the issue. It called the java naming directory insecurely

[hollow dune]

i see

[versed terrace]

this allowed remote access - through the logging

its why it was so awful. Practically all java applications that used logging used log4j

[hollow dune]

sometimes the first solution you think of can be the worst one 😅

[versed terrace]

suddenly they're all vulnerable

clever malware also looks for logs and backups to destroy them. Its why we recommend companies to keep offline backups- and why sometimes keeping old backups is important too as malware can remain undetected for a long while

its a cat and mouse game

if you'd like I can answer any questions you may have in dms so we dont clutter the general chat XD

[hollow dune]

i have once created my own JWT Authentication following Web Dev Simplified's Tutorial, and i remember asking a friend, why it wouldn't be safe to hash on the client, then pass the hash as the password rather than plaintext and hash the hash again 😅

he told me about pass the token and many flaws that my "idea" came with, at that time i thought i did something genius 😅

[hollow dune]

you are totally right, sorry >_<

[versed terrace]

no worries!

[zenith dome]

Do you still have that in your browser history? Can you share the bottom right?

[slender hornet]

whats a composite?

is "bisexual" allowed as a prompt word as in "bisexual lighting"? for those who dont know, bisexual lighting is qhen there is half red half blue light. looks really cool.

[hollow dune]

for those who dont know, or pretty much everyone

i would never have thought a lightning composition could have “-sexual” as it’a name

actually now that i think of it

red and blue sound like disco colors so maybe there’s a reason behind the name

[slender hornet]

https://en.wikipedia.org/wiki/Bisexual_lighting

neither did i know of it

[hollow dune]

now we need ‘german lighting’

the use of red, black and yellow

[slender hornet]

*gold

[patent grail]

does anyone else experience prompt paralysis sometimes

like just cant choose what to do for so long

and scared of wasting

[slender hornet]

a bit yeah

[hollow dune]

right

i’m a german, should know it haha

[blazing atlas]

I feel like there are a few game devs in here based on the intros, but has anyone ever got into creating with Dreams on PlayStation? I really love its workflow and there's some amazing stuff out there (from people with more time and skill than myself 😅)

I'm starting to think about how DALL·E can be used to quickly conceptualise and give visual stimulus for this kind of creation in an accelerated workflow:

"isometric view of a colourful landscape with a river and pine trees. craft aesthetic, 3D render"

[turbid hatch]

The breadth of multidisciplinary workflow speed and quality increases that Dalle, in congruence with other AI systems, will achieve is a revolution on par with the printing press imo. It’s awe inspiring, or as Dalle would say, it’s “EW-SPiirgN€”

[blazing atlas]

Absolutely

[slender hornet]

@hollow dunesee #art-and-illustrations message for a good example of bisexual lighting

[compact willow]

Can I use a swear word?

See #art-and-illustrations

This is breathtaking

And the bot got pissed at me because I uploaded 6 images at once

[slender hornet]

@compact willow upload 3 images at once, then the other set of 3 images 10 seconds later

[compact willow]

Ah

Thanks for confirming

OpenAI will put every industry out of a job

[slender hornet]

nah lol

[compact willow]

Are there ethical discussions about Dall-E 2 regarding artwork quality and making artists and original ideas redundant?

[proven linden]

Search Holly Herndon and Mat Dryhurst. They talk about this a lot.

A key thing though: Art is dependent upon context. New AI tools are basically the equiv of the invention of photography. Its a new medium and new context.

[tardy granite]

I truly feel that this concept that dalle will "render artists and original ideas redundant" is completely ridiculous, somehow this notion has spread like a virus and has become some kind of scare mongering idea. As an artist, and a writer, I feel only empowered by my time with dalle, and in no way do I feel threatened. Likewise, I am part of a relatively small game development team, and to be able to use dalle to actualise concepts and ideas, from meetings or just general conversation, is an exciting prospect. I am the defacto "concept artist" on the team, even though I am officially the writer and composer. So, thinking of what dalle could afford me, and the team, is so inspiring and so special. The makers of dalle2 referred to dalle2 as "a creative co-pilot" and I think that is the best way to think of it. I do hope and believe the technology will be come almost an industry standard, but in parallel to existing artists and jobs, not instead of them.

Besides, Dalle will always need a human to create what might need to be created, and there is an art to prompting, no doubt. Haha, a funny thought is a "Prompt Specialist" being hired, like a concept artist would. Someone with the language and knowledge to produce the highest quality results for specific projects... Anyway, these are my rambling thoughts.

[proven linden]

https://twitter.com/hollyherndon/status/1522341497449979906

Holly Herndon (@hollyherndon)

We have loved experimenting with @OpenAI DALL·E this year, and so @matdryhurst and I wrote a blog about what we have learned!

https://t.co/2N5PYnlBAT

We share some techniques and reflections on this remarkable tool 🥰🦾

Here's a meme I made with my embedding #dalle #dalle2

Likes

434

Twitter · May 5, 2022, 10:24 PM

[slender hornet]

ah god damn accidentally pressed the generate button too early. prompt wasted

[zenith crater]

how it started vs how it's going

[patent grail]

hitting the generate button w/ the new limits feels a bit like playing minesweeper now lol

[finite ridge]

Waiting for dalle to clean my toilet

[weary ore]

Sigma 35mm f8 photo of a pretty and clean bathroom, the floor is bright and no dirt is seen anywhere

[weary ore]

Yeah, in fact a lot of that has been discussed on this server before and in the AI field as whole. I guess the most consistent explanation I've seen about why it WON'T render human creativity useless is basically that every AI model so far needs training. And it can only train what is a 'right' or 'wrong' generation if there's some kind of example of what is expected. It may be the most sophisticated and optimized algorithm, but it needs the human input/example/guidance to get to what WE human like and consider right/wrong, to approximate that mathematically. And again, even if said model master 100% of what we currently done and like and find a way to mix and match that all, recursively, it will eventually run out of new tricks if no new information is generated by us, human, for it to learn (new music, new art, new books, ideas etc.).
Some may say that, someday, it won't need us to create 'genuinely new things' (another highly controversial terms, even for human artists but well... that's another rabbit hole heheh) that we are able to like and experience, and that may SOMEDAY becomes true. But this day is very far from now I guess, as some of the most advanced linguistic model, just to give a concrete example, struggle with some very very simple concepts for us humans like basic mathematical operations and in turn mathematical models are... well, mathematical. And these kind of things happen all over the AI field really.

That said, I DO think that it will *generate *(pun intended) changes in the creative field, specially for commercial creative activities (like branding, marketing, etc) but I see it in the mid-term more as a tool that these professionals will have to incorporate on their daily jobs (just like photoshop and computers overall were incorporated in this field) than something that will replace then.

[granite merlin]

DALL-E down for anyone else? Haven’t been able to generate variations for the last 10 mins or so

Nvm seems to just be variations when uploading for some reason

[weary ore]

That's strange. I just generated a variation from upload right now to check, and it's working fine on my side...

[blazing atlas]

I worry a bit that having the AI output train itself is a solution that circumvents some legal issues around a model trained with others' IP. It will diminish the quality - there are so many intangibles in authenticity even if it passes a human-eye test, that won't be consistently in the training. It's also a kinda disingenuous approach to the complexity of copyright here. But I'm speculating a lot.

[granite merlin]

Will give it a go again 🤔

[weary ore]

It makes sense, I guess. In my humble opinion copyright will invariably cease to exist in the future. It won't make any ethical sense and, financially, big corps will eventually find a more attractive solution. With the speed we will 'consume' art/creative content (I hate the wording here, but well...) and the access to art being primarily virtual (no physical medium needed to consume it), the whole copyright thing will probably become obsolete. What will replace it, and how much time it will take, I guess we will have to wait, to see.

But as an artist that entered the AI field in the last year, I have thought about that a lot, lately.

[patent grail]

You make a good point, but what you described is also how babies learn everything. We're not that different from the AI.

I learned everything just from seeing things around me and getting positive or negative feedback

I think eventually the AI really is going to make every job redundant, and I'm looking forward to it. People will look back and be terrified at how much time their ancestors wasted creating spreadsheets, driving cars around, and typing up reports.

I agree with you though, I don't see how Dall-E could replace a full time artist. It creates striking images but it's not consistent between generations and there are many prompts it just ignores (putting rings on fingers, describing more than two objects, etc).

I do see it having an impact on people working gigs on Fiverr though, I haven't really seen a lot of discussion about that

[blazing atlas]

The human / AI analogy is not very good in this case. The scale of human intelligence is very different in practice to this clever illusion

[patent grail]

In his comment he said a limitation of AI models is that they require training and human guidance on what is right and wrong

Which to be fair is also what babies get

I see your point though, Dall-E is not even sentient, there's a huge line

I bet if the people who worked on the paper saw this chat, they'd think we're insane lol

I read the paper for the 1st time today, puts things into a bit of perspective lol

[blazing atlas]

I mean no doubt machine learning is more like human intelligence than traditional programming, but to me it's like saying an orange is more like the sun than a grape. (I have very high esteem for the humain brain 😛)

[patent grail]

LOL, that's a good way to put it

That stuff gives me a headache, I can't handle the problem of consciousness

[blazing atlas]

If you want an even bigger headache, read the recent Scientific American blog post on quantum theory, and how nothing in the universe that has not been observed is actually physical (at least within our current idea of physics)

https://blogs.scientificamerican.com/observations/coming-to-grips-with-the-implications-of-quantum-mechanics/

I'm no Everything's a Simulation, Man! type but this sounds eerily similar to the way video games don't bother to render or physically emulate anything that's not within the camera.

[patent grail]

oh no

I'm going to read it, I googled scientific american and I also found this https://www.scientificamerican.com/article/does-quantum-mechanics-rule-out-free-will/

which is giving me a headache already 😭

I will read that after hahahaha

[lofty birch]

Re: ai automating away artists discussion, I thought this guys video had some good takes in it’s ~30 min runtime. https://youtu.be/oqamdXxdfSA

[patent grail]

This article....

🤯

I was gonna generate one last kitten and go to sleep

now im having an existential crisis 😭

[blazing atlas]

Yeah had to nope intellectually out of that idea pretty fast. I have kids to take care of

[proven linden]

On quantum stuff; Sabine is one of my go-to's when it comes to cutting through overconjecture on quantum topics, which is annoyingly frequent https://youtu.be/Wsjgtp9XZxo

[hollow dune]

ahh i see

[blazing atlas]

https://twitter.com/jrohlback/status/1540959442971541505?t=bS9IXkb6NcTIcArFSH6YKg&s=19

Jeremy-O (@jrohlback)

Here's how @OpenAI #dalle2 was used to turbocharge the creative process with @mediamolecule #MadeInDreams, using my phone and #PS5

4:50pm: Start a very basic scene

4:56pm: Send a screenshot of the scene - a few loosened sculpts, grass, clones trees, a river - to my phone. 1/3

Twitter · Jun 26, 2022, 7:25 AM

[slender hornet]

i am so excited to try a large project with Dall-E 2 tomorrow when my promots are back. ill try out inpainting. ill start with describing a location. then inpaint one or multiple objects into it. then inpaint a character into it. then inpaint the characters face and other small details.

i have high hopes that this will result in a great detailed fantasy painting. made entirely via AI.

[slender hornet]

but first i wanna try outcropping via inpainting.

i generated this image using Dall-E:

doubled its size and added transparency everywhere

and now i am excited to see how well the inpainting will work to fill in the gaps

actually wait i am stupid, dall e only works with 1024x1024, ofc doubling size wont work. hold on

there this makes more sense lol

Dall-E's image generation is already amazing but the editing/inpainting feature is the cherry on top

just 7 more hours

[slender hornet]

i have an artist on instagram who i regularly order commissions from and he views image generatoon AI quite negatively, because he thinks art is a very human thing

[manic dragon]

That's my video. Thanks 🙂

People said the exact same thing about the invention of books. Everyone was worried that, because everything was written and recorded, we'd never have to remember anything, and referencing a book was "cheating". I know that AI is fundamentally different from other tech, because of its ability to adapt. I think it's simply too early to say whether the average artist will be out of a job or not

Maybe it will put production artists out of a job, maybe it won't.

Maybe it will make teams much smaller, therefore eliminating jobs....or maybe it will be a bunch of small teams doing several productions, rather than a few huge teams doing huge productions?

With the supply of good art skyrocketing as a result of AI, could there be a newfound demand for custom, bespoke art or tv shows or whatever else?

I design physical products as my primary discipline. 20 years ago, the only way to sell your product was to get a distribution detail with a big box retailer. With the advent of the internet, you can be a small product company with a tiny niche that other big companies completely ignore. Maybe the same thing will happen with media/entertainment/art?

Once again, hard to say for sure, but I think it's very presumptuous to say that "Creative employment is dead. We're all screwed."

One thing I can say with certainty is that creative employment as we know it is certainly dead.

But with every end there is a new beginning.

/end rant.

[lofty birch]

OpenAI hurry up and create a post scarcity utopia so AI can free humanity from drudgery please ^_^

@manic dragon your video is definitely one I am pointing people at for those who have these concerns because I think you outlined a lot of them well. The other one i point people to is the Holly Herndon / Mat Dryhurst article about DALL•E 2, have you read that as well? Would be rad to get them to talk to you lol do another vid ^_^

[manic dragon]

I haven't read it, please send the link

[lofty birch]

https://mirror.xyz/herndondryhurst.eth/eZG6mucl9fqU897XvJs0vUUMnm5OITpSWN8S-6KWamY

Gets a little technical and speculative, very interesting, helpful too I think for situating this moment in the context of previous effects of technology on art history development as your video touched on as well. Good read

[manic dragon]

I'll read it right now, thanks for sharing

[uncut wadi]

As a production artist, there is a lot of fear and uncertainty in my circles. Personally, I can see certain art jobs being gone soon. Primarily ones focused on "popcorn art". For example, kids coloring books. They demand a ton of content but those images are usually very cookie cutter with slight variations and a range of quality that don't really matter to the consumer (kids).

[lone quarry]

Hi

I have access but literally don't have time to use it properly and i feel too guilty about it

Can i give the access to someone else who can use it better

If there is please let me know

[slender hornet]

ehhh i think AI alone will not help us end up in a post scarcity utopia. we also need to fix our societal problems.

thats the big theme of cyberpunk. humanitys technological progress outpacing its societal development.

still, these AI developments are inevitable and I dont think they will suddenly make artists obsolete anytime soon.

[slender hornet]

i would have said just fill peoples prompt requests but like that also requires time...

unfortunately account trading or giving your account to someone else and such things are banned from what I know.

[lone quarry]

😔

[slender hornet]

also when people say this AI is about to replace artists jobs... have they seen what capable artists can produce? Dall-E 2 is very good but it is still far away from reaching the quality levels of a lot of artwork ive seen online (i would post examples but i think that would look like me promoting their work here so ill not do it)

[warm widget]

yes! it might replace basic artists

[uncut wadi]

Yeah it's a spectrum. Like you're a professional artist when someone decides to pay you but that doesn't mean every artist is capable of the same output

There's plenty of businesses that have artists on staff or deal with freelancers as a means to an end; to generate profit.. Not necessarily because they hold some value on artists themselves

[patent grail]

theres no right or wrong way to use it, ik people who only use it once a week to make something funny

i think openai uses the usage data to learn more anyway

and theyre adding like 20,000 ppl a week now i wouldnt break the tos by giving away your account and getting banned for life

no need to feel guilty

[patent grail]

Probably stuff like fiverr, when people want stickers, profile pictures, thumbnails for youtube videos, that kinda thing, it will be relevant

Although I saw a fiverr gig earlier this week where they guy was offering to copy and paste your prompt into Craiyon for $60 😂

[zenith crater]

[civic wadi]

This reminds me of the default prompt given when you open the Disco Diffusion notebook, someone turned it into an NFT 😂

It’s the “hello world” of disco

[slender hornet]

oh man i hate NFTs

[zenith crater]

same

[strong silo]

I don't know if I'm getting old, but what’s nft so cool?

[grand onyx]

Every day i get one warning from dall-e 2 out of 50 prompts

the warning is because of keywords

like i had no idea the white house digital art would give me a warning

How am I supposed to know what is prohibited and what is not?

Do warnings reset after a period of time?

[slender hornet]

i have not seen any official information yet about how many warnings it takes to get banned, or if its manual review and you may be banned after 1 warning already or only after 50 (like depending on what you got the warning for), or what words are banned, or if they reset after a time

but generally they are playing it very safe and for example White House can very easily be used in a political context hence its banned

[soft hull]

I'm not sure, but I've been warned before, it is clear "furry art" is not accepted. Other times I have accidentally included the word 'war'

[patent grail]

#community-help message

[soft hull]

yo hi smtz

is dalle2.app frozen?

like forever

[patent grail]

nah, im working with some ppl on it

new vision

[soft hull]

Cool! Excited to see what you come out with, as the site currently is already epic :D

[patent grail]

but yeah, from this it sounds like you have to really mess up to get banned

[soft hull]

oh

god

I just read that message, I guarantee the only furry-aligned content I try to produce are normal images unlike the ones mentioned in that message

[slender hornet]

thank you! this is actually really valuable information. so no one sjould have any worries.
@grand onyx see above

[slender hornet]

yeah i think normal (e.g. not naked or not characters engaging in sexual acitvity or poses and such) furry characters should be fine.

but thats just my guess

[soft hull]

I mean, things like my profile picture

Not weird creepy shit

[slender hornet]

thats not a furry character

thats just a wolf

[soft hull]

It's a fox

More like images of this calibre, that are easier to look at and say "Yup that's a furry"

[slender hornet]

yeah rhats more like it

also sidenote but i like that design

nice colours

[slender hornet]

i have created my first outcropped image!

though the results are... well... lets just say that using a character that misses part of her head as a base maybe lead to some issues lol

[patent grail]

top right is an awesome effect

ive never seen a full body + face that looks good

[slender hornet]

yeah thats not what i meant but the 2nd head placed on top the old one in some of the images lol

[patent grail]

ohhh i see now

thats weird.....

i feel like each layer in this network could be released as its own billion dollar company

just the ability to upscale from 64x64 to 1024x1024 is insane

[blazing atlas]

this is not true; most people care about quality. what they can't do is fault quality without a technical understanding or language of criticism, in art or any other domain. so if you slap a shitty AI photo up in a commercial context most people won't "care" until someone points out e.g. a fault in the light refraction, a clearly glitchy detail, other mishits on verisimilitude. at which point they'll be up in arms, and the brand or whatever suffers

one of the reasons I'm in this community and not GPT3 or whatever is that I'm a writer and teacher of writing, not an artist. when i play with language-based AI i get the shits because i have a lot of metalanguage and structural understanding to break it down. whereas with DALL-E i love it because it looks fine if ya squint

[steep ginkgo]

i think of ai more as a tool as many of my artist friends ask me to make references, im a programmer and when copilot came out people where horrified they would loose there jobs but now i can produce work to my understanding 10x faster

huh didn't see this

there where a few talks here and there at the first annoucments

especially from nocode enthusiast

wait i think ive you around the sever are you the software dev?

[patent grail]

I've tried to use copilot, I find it kind of interferes a bit

But OpenAI's model itself is insane, it can write websites and stuff on its own

I'm sure within a few versions I too will be out of a job 🤣

[patent grail]

!watermark

[muted sedge]

Cropped Image!

Here's your cropped image. Remember to credit OpenAI when sharing it.

[warm flax]

!watermark-small

[muted sedge]

Cropped Image!

Here's your cropped image. Remember to credit OpenAI when sharing it.

[warm flax]

yayyy thank you @patent grail for the bot and @wicked kiln for the templates!! ❤️

[patent grail]

!watermark-small

[muted sedge]

Cropped Image!

Here's your watermarked image. You can now use it as a profile picture!

[warm flax]

(just created #dall-e-bot for anyone who wants to use it!)

[patent grail]

(I changed the description a bit)

tysm for adding it!!! yay 😄

[slender hornet]

i do. as evident by me paying a lot for quality art.

[slender hornet]

oh wow that is amazing.

[hollow dune]

eh? excuse?

if i didn't care about quality i'd use craiyon, because it is way less restricted and more knowing than dall-e, but craiyon lacks the hd quality

[zenith crater]

this bot is amazing

@patent grail @wicked kiln my two besties

[hollow dune]

#dall-e-bot

[slender hornet]

i wonder if it would be possible to generate a whole high-resolution character using composites?

[patent grail]

thats what im thinking

[slender hornet]

@patent grail like we know that it can create very detailed high resolution pictures of faces. but once its full length character design it all becomes amlittle bit incoherent and mushy. i wonder if you could prevent this by just doing a composite?

start with the head, then do a half body, then mid body, lower body, legs and feet, etc?

[patent grail]

its a good idea

i tried a bit today

but dalle is kind of psycho it just adds random things

i guess i should be padding the bottom only

to work down from the head

the first full body, high quality dalle photo 😳

[slender hornet]

pls keep me informed on the progress! i have run out of generations so i need you to serve as a replacement until i can try it myself tomorrow lol

and yeah probably better to only pad the bottom and not any additional stuff in the already generated parts or else it will just confuse Dall-E probably. afterall it uses the already generated stuff as a base to figure out the correct proportions and anatomy

[charred pivot]

Hi all! I'm an artist with aphantasia and getting the chance to work with dalle is something that I treasure greatly. Anyone here has aphantasia or is on the spectrum?

[soft hull]

aw hell nah one of the nft cryptobros got access

[scenic linden]

bruh

[soft hull]

ikr

[weary ore]

Yeah, but I guess the general idea is IF we should do things without it in the first place, if the tool makes the process yield better results, with less struggle and is accessible enough to most of the artists. The final decision is always on the artist, either way. Sure there are certain artists who are used to and prefer (for ethical, practical or artistical reasons) to not use it, and that's ok also. They develop their process and follow with it, and yet are able to generate amazing results doing the things the way they believe.

In my way of seeing it, this discussion is somehow similar to Autotune discussion in the music industry. Pitch correction software, as far as I know, cannot be considered ML or DL like DALL-E, but is certainly an AI tool. A lot of people are against its usage, but most of pop/rock/dance music nowadays use it (in a smaller or bigger degree). There will always be artists (and segments), which prioritize live 'raw' performances, that won't use it as they feel it will take away the 'energy' of the original performance, and that's ok. And there will always have artists whose aesthetic depend heavily on the 'artifacts' that the tool produce as a sound signature, which is also ok. I can see the 3 scenarios applied to basically any type of artistic AI/ML/DL tool so far. They are all operated by humans, are completely optional, but they eventually yield results that are so good and valuable (if used correctly) to the process that most artists tend to use it, if it's available to them, to some degree.

[weary ore]

Welllll..... metaphorically yes. But on the technicalities they are very different, really. One thing that haunts the AI field is that the models that are developed have results that, to some degrees, are very alike the ones that human intelligence produces. But at the same time the process is so distant and of such a different nature that some in the field even go the extent to kinda question if calling it 'Intelligence' or using terms like 'dreaming', 'hallucinating' and similar wording when dealing with AI does more harm than good.

I guess the fundamental difference between humans and mathematical models is the generalization, really. Humans can become specialists on physics, psychology, great athletes, artists, gamers, doctors and etc. all with the same 'brain architecture', using the same body, and with the same 'hardware and configuration'. When you are dealing with AI the things are very different, at least structurally, the most general models that have seen the light of day until now are, well..., limited to say the least in their generalization. And even in domains that they are able to work on, their results stay very close to average human results. That will certainly change with time and whatnot but the point is that the path that we will take when dealing with AI is more on the mimetization of human actions than of learning, really. That's basically why they need the examples. Human babies also need them, but once they start developing language and comprehension of causality and a whole set of other skills, they can start interact with the world in a different manner and build and evaluate their own example/results. AI models, at least for now (and the foreseeable future) cannot do that in an effective way. That's why they still and will stay dependent on human input for a long long time. And that's a good thing if you ask me. Humans are incredible and we have yet to fully understand our potential. #teamhumanftw hahaha

[burnt mountain]

Awesome bot! @patent grail

[quick plaza]

I have found a killer prompt. It's a bit random but "Mickey Mouse portrait out of ketchup at a bookstore" generate random outputs that did not include anything named in the prompt. Have tried it multiple times (and flaged it once)

[weary ore]

Awesome job there @patent grail! The padding function is just amazing!

[soft hull]

22:15 i get prompts lol

[burnt mountain]

Too many flagged words?

[patent grail]

How many violations did u have?

like strikes

[burnt mountain]

False positives? I remember in OpenAI's AMA, when your account gets falsely suspended, just email support. Someone will take a look at it.

Whaaat

[patent grail]

Ya you should email them 😭

Joanne said people really only get perma banned for awful awful stuff

Don't have the link to the message rn but yeah

[burnt mountain]

I did too, a few hrs ago

Idk where'd you get that info from. They don't comment about the datasets they may/may have not used

Well there's your reason 🥲

[patent grail]

Maybe not a good idea 😭

[burnt mountain]

Policy literally says:

[zenith crater]

i mean... that DEFINITELY implies something of an r-rated nature

[blazing atlas]

If you're curious about HTML and CSS layout and had respect for the rules, you wouldn't have chosen thst website. Don't be disingenuous.

[burnt mountain]

Says who

[blazing atlas]

It's pretty unambiguous that it's not within your role as a current beta user to be troubleshooting that problem. Email support, but be contrite.

[burnt mountain]

And that lead you far. Congrats.

[zenith crater]

If it was truly unintentional wait until a staff member comes online and reviews your prompts

You should be unbanned if that’s the case

[uncut wadi]

You may be jumping to conclusions. No need to write a manifesto when it could just be that office hours are over or something equally reasonable / likely 🎱

It's also entirely necessary in this case.. Its a relatively small team trying to monitor 50~ entries per day for 30k~ people. Just saying it might be explained by simple issues of manpower / workload and to wait on a response before reacting to anything

[lofty birch]

I admire your courage lol I’m just like “ok please don’t ban me guys thanks for letting me play with this cool toy for free” lmao

Lord knows i’d love to make stuff that could be used to do political commentary but I am refraining until the service becomes more productized and terms become more explicitly clear

[blazing atlas]

Publishers legally responsible for content they disseminate moderating that content is not related to "free speech", but it is complex. That's all I'll say because it is inherently political.

[zenith crater]

@warm flax thank you for the change

[patent grail]

@shy sphinx 🥳

[zenith crater]

and juicestie

[sterile bluff]

at least you have a sense of humor in the end about it.

there is going to be a continuing structural tension within research groups and companies who want use of the applications and products they are developing, but also want to protect their organization and brand from harm.

the scope of what harm entails and the way it manifests is becoming broader as the means these applications and products can be used and abused also broaden.

private organizations, especially for profit corporations (like Reddit, Facebook, Twitter) certainly muddy the waters as they become an incumbent way for people to interact and engage with each other. they are not truly free and open.

Government interference with the internet is also problematic.

What's the famous quote? The internet should interpret censorship as an outage and route around it?

hopefully people remember that and work towards building platforms and tools that meet their needs if the information or works they want to create and share contribute to a stronger, more loving society.

[edgy bluff]

Damn. That sounds very cool :O would be Nice to have DALL-E on demand

[haughty zealot]

i imagine all these advances with text to image AI will prompt the engineering of new analog systems which can run it better

i have no idea what im talking about

[edgy bluff]

Would be hella cool though. Analog computing is the future

[haughty zealot]

interesting

[patent grail]

dalle will be running on calculators one day lol

[edgy bluff]

XD

[haughty zealot]

how would it run on a desktop computer though

[edgy bluff]

I Will run dall-e on my texas

[haughty zealot]

np

i will watch

🤔

watching

[edgy bluff]

https://tenor.com/view/yoda-sad-upset-tired-star-wars-gif-15907225

So you are saying. That the diffusion algo could be run on a pc

Is it as good?

So. If you create a good clip method You could make a program as good as dall-e basicly

Oof

Yeah Well ey. You’ll make it work. I believe ya.

[haughty zealot]

wow

[edgy bluff]

:O

Okey :O

I won’t make any learning algoritms. I’m just an artist. That’s what you do. You’re the medical programmer making programs for NASA

[haughty zealot]

its quite goood

[edgy bluff]

I believe in you Though. You’ll make it work 🤌🏻

Don’t think you should use that word

Cool 😎

Cool. Almost like You’re an artist with the code

[sterile bluff]

That comment wasn't about you. It was about society.

But, if you identify yourself in it, then you're already better than most people.

[edgy bluff]

Cool 😎

[sterile bluff]

It's more about what is the size of the model after it has been trained. Yes, it's likely in the tens or hundreds of terabytes of image files

but when diffused and mathematically visualized

well, let's just say if a Mj runtime can be executed on a single A100

I'm confident that other current gen models can do the same.

[edgy bluff]

@sterile bluff love your title. Ai art whisperer xD

[sterile bluff]

@edgy bluff self anointed, but I do get around. Thanks. 🙂

[edgy bluff]

Hehe. I have given myself the title machine whisperer on my instagram xD

[sterile bluff]

I love it. What's your IG?

[haughty zealot]

great video! thanks for sharing!

had to rewatch some parts

[blazing atlas]

Warning - a little bit of an essay incoming. Apologies in advance. Feel free to scroll on 😅

There are a couple of things I've been starting to think about that I haven't heard discussed yet but I think is important; I hope it's not taken as political - I never think about this stuff to be woke or whatever but because I have academic experience with critical theory, sociology and philosophy, and it's hard for me not to.

Firstly, this type of technology is, in many ways, perfect objectification: that is, because it has no relation to specific humanity, and yet represents surface level qualities of human beings, its results are perfect "objects." A lot of times that's not a problem, I can create a perfect abstraction of a tree and I won't then carry guilt that I have objectified the planet. But if a man perfectly objectifies women - if I do it with a simply typed, innocuous phrase and a 20 second wait - is that truly of no relation to specific human stories, histories and relationships? As usual, it's about power. I'm going to avoid the word appropriation because that is hotly politicised, but if I treat indigenous or cultural histories as objects in my little digital diorama - as I have done in the last few days - it's a shirking of responsibility, in my opinion. A white male like myself using AI software to produce images shares the same ethical responsibility, ultimately, that any other white male visual artist or writer does when it comes to representing others' and using others' traditions. The problem is amplified by the ease and lack of expertise with which an AI user can create their "perfect objects," and their disjunct from a social moderation based on relationships and community - ingrained through growing art theory and culture.

The second issue is more interesting, more technical, and potentially more of a problem; that is, that DALL·E seems to be guilty alongside its other biases of a kind of Orientalism. In my experience it has treated ethnic / non-white (and especially indigenous) cultural cues as generic ones, and will often confuse indigenous art-forms, styles and subjects. A prompt to represent a "Gamilaraay woman" manifests as generically brown faces; a "Wiradjuri woman wearing eagle inspired dress" took colour cues from bald eagles and presented more North American First Nations visual coding. Perhaps this can be 'explained' by a lack of representation from darker skin-tones in the training data. More interesting, a "Mirdidingkingathi Juwarnda Sally Gabori" painting cue presented distinctly African masks and some generically "ethnic" images - perplexing because I don't understand how this tokenistic, generic treatment of very distinct Black cultures (cued by a specific Kaiadilit woman's name only!) actually functions, in terms of the model's logic... is logic even the right word?

Part of the thought process behind dealing with bias in AI imaging is that it's still the tool of an artist, and that artist is responsible alongside the AI for diverse representation. If you take the responsibility off the artist and erase all bias from the model pro-actively, systematically, in some sense we lose a valuable tool for understanding the pervasiveness of bias in our own media landscape. That may be a price we need to pay. But if the technology is actually incapable of creating specific diversity; if already marginalised cultures are not just relegated to the margins but the same, inaccurate margin; and if it becomes impossible as an artist to produce meaningful diversity - or, particularly, self-representation - with intent, that is a bigger problem. That's a problem with the technology.

Because of the first point, I feel a little 'icky' accruing and presenting examples of the second point; so I'm not going to do that. What I would say is that to me it's a case for more diversity in the tapestry of users who are contributing to the early days of these discussions! It's a radical new medium, and such questions are going to be asked sooner or later.

[haughty zealot]

i have tried to generate some more indian-specific cultural objects which DallE gets mixed up with other similar objects

same thing when i do stuff relating to jewish or indian-jewish culture.

but i'm not paticularly concerned about it, because i understand that the training set is more broad and the limitations of latent space on current hardware might cause some of these concepts to be categorised as indistinguashably similar

as a person of colour who belongs to a culture of less than 67,000 people, i think its important to understand how the model works, and that my own culture might not be documented or represented because it would take crazy amounts of money and computational power to run a model with a latent space which represented the sub-aesthetics of cultural artifacts from every culture from around the world. right now for better or (mostly) for worse western american culture is the global culture, having the advantage of years of technology which the global south where i live does not. it's incredibly difficult to build an algorithm to sift through images online and categorise them correctly to remove all bias from a training set, the task is basically the same as building DallE2. i am confident in OpenAI's efforts in reducing bias, but i also think that its pretty much an impossible task to accurately represent millions of cultures around the world.

imagine taking every unique aesthetic movement from your own culture, then multiply that by a million. thats roughly how many dimensions would have to be added to DallE2's latent space to accurately represent every unique culture on earth accurately.

i cant even begin to imagine how many petabytes the weights would be for that model once fully trained, let alone the amount of money it would cost to train such a thing

but i think it could be possible in the future with analog computing, but i dont know what i'm talking about because i dotn understand analog computing in the slightest bit.

but there is a definite case for a better and larger training set here, at least to the level (not in all aspects, like the more vulgar or tasteless) that dalle mini has.

completely agree about your point about the artist being partially responsible for correct representation, or at least having a disclaimer about representation

i would argue though that all art is inherently objectification at some level, it just feels more icky because its a machine doing it and its more detached.

[blazing atlas]

Yep I did consider that! Thanks for your response btw @haughty zealot, I was going to do this as a Twitter thing but figured this community is pretty trusty. Your experience is really appreciated (to me, anyway)

[haughty zealot]

totally agree with your concerns that AI in art and design might kill off diversity in aesthetics, a trend which has been happening for a while because of pinterest. eventually itll be up to us to train our own models with the influences we want, which will reverse this trend.

hopefully OpenAI will in the future create interfaces which could help unseasoned designers and artists with this

like people who dont want to suffer through days of google collab errors lol