openai-chatter

[sterile bluff]

sparking, please consider this a warning. you have an opinion, you've voiced it in the appropriate channel.

disengage from arguing with people about this.

[left pecan]

Ik I'm not doing anything he said something so I responded to him

[sterile bluff]

and I'm asking you to not do that

and move on from the current issue.

[raven obsidian]

Imagine if dalle could generate gifs

O.o

um

[sturdy river]

Next ai idea: Upload an image, a trendy caption is generated, basically the opposite of dalle which it already does behind the scenes afaik

[raven obsidian]

is that a million dollar idea?

[left pecan]

Anyway dall e is getting worse yall 😭

[raven obsidian]

an AI that generates a GIF

[left pecan]

My friend generated me thus and it looks bad

[sturdy river]

There are already some out there, although not too great yet and not beginner friendly like myself

[left pecan]

The quality went down

[raven obsidian]

why is it every time I come up with a good idea

[left pecan]

Help me

[raven obsidian]

someone else has already thought of it

[left pecan]

Her face is eughh

[raven obsidian]

so entirely random out of the blue question, is there an age limit for this server?

[obtuse zealot]

Idk

[wraith juniper]

temporal coherence on AI generated images is very hard
there are system that are capable of that already, but I don't think that dalle was made with this goal in mind

would be cool tho

maybe on dalle 3?

[raven obsidian]

im sure its easier said than done

[raven obsidian]

lugui

your profile is amazeballs

[left pecan]

See what I mean

[wraith juniper]

my profile picture? thanks =)
i've made it myself xD

[obtuse zealot]

https://labs.openai.com/s/zY0z4AxlWb89vfLTCeYD4H3x

[raven obsidian]

that and your profile itself

[left pecan]

Look at the body

[raven obsidian]

"i'll do anything for a price ||i'm a software developer||"

[left pecan]

It's slightly Decreased

[raven obsidian]

what are you talking about?

[wraith juniper]

hahahaha thanks =P

[left pecan]

His picture

[wraith juniper]

my profile picture was inspired by this: https://www.extremelyfungible.com/default-cubes

[raven obsidian]

it looks perfect

[left pecan]

Not his pfp

[left pecan]

Thia

[raven obsidian]

yes

that

[wraith juniper]

btw, now i got an idea.. maybe make a collection out of Dalle's outputs for EFT

[left pecan]

Look closely

At the head and legs

[raven obsidian]

So AI is often times not going to be able to get something perfect

[sterile bluff]

sparking - every generative tool has strengths and weaknesses. some are overcome by better prompting or, where applicable configuration

[left pecan]

No sloth I'm not advertising or protesting

[raven obsidian]

it may have extra or missing limbs

it may not have PERFECT dimensions

[left pecan]

I'm just point stuff out

[raven obsidian]

but overall there is nothing wrong with that image

[sterile bluff]

further, this is a beta

[raven obsidian]

if youre looking for perfect images, you could take a course in graphic design

[sterile bluff]

you're welcome to point stuff out, but it feels like all you want to do at the moment is complain.

[left pecan]

Sorta is but stop tagging me and mentioning me we are in another server rioting over rhis rn lol

[sterile bluff]

and I'm wondering what value that brings to you, in your life.

[hoary leaf]

already exists, CogVideo

[raven obsidian]

what?

I havent mentioned you

[sturdy river]

I wonder if dalle will implement a tool to make those zoom-out videos, or maybe a third party thing

[left pecan]

Not u

[raven obsidian]

you just replied to...

huh?

[left pecan]

I'm saying to u all in general

I can't respond in 2 servers at once

Lol

[raven obsidian]

I'm still lost

[left pecan]

It's ok. Don't worry about it

Lol

[wheat storm]

The vibes of this server has gone down since the masses were let in

[wraith juniper]

you can always refine the results by in-painting it
some times, the AI's model just ins't trained enough on the style of your request, you can see as it generates stunning images of faces, it absolutely nails styles like paintings and get extremely cool drawings because the imperfections are disguised as "the style"
but sometimes if you want things on styles that it has less data, it will surely not get a good result on first try

[gloomy nimbus]

Anyone feel the model has been nerfed a bit ? Redoing some of my older prompts and getting way worse results... not even close :/

[left pecan]

I just say don't mention me!

[sterile bluff]

@wheat storm fair, more people = more opinions

more noise.

[raven obsidian]

I'll try and re-vibe us. This one was pretty cool.

[wheat storm]

Yeah with everything haha

[wraith juniper]

I've been testing this last few days about combining dalle withother programs, photoshop and I just got a nice result with Blender

[raven obsidian]

Closest I could get to a pirate with my description without getting flagged lol

[wraith juniper]

looks like Kunka from Dota 2

[wheat storm]

Thats really detailed

[raven obsidian]

I did a series of trying to get dalle to invent new techy stuff. I should try ask for something more specific, like a memory wiping device.

[obtuse zealot]

She/ her lol

[rocky burrow]

Great prompts!

[left pecan]

These are my closest results

[raven obsidian]

Honestly some of my favorite images dalle has generated for me were ones that didnt fit the prompt I asked for

Like this was supposed to be a 3 legged spider like vehicle

but look at how cool that is

[brisk anchor]

Serious Black Ops Mystery Box vibes!

[left pecan]

Sorry lol

The low one

[raven obsidian]

KITTYS

[left pecan]

This is the worst one

[obtuse zealot]

Haha it’s okay

[left pecan]

The others got close

[obtuse zealot]

https://labs.openai.com/s/P6BfDMBX1erz1kWJZYxCE8FE

[brisk anchor]

I feel like it is selection bias

[raven obsidian]

god i really shouldve spent more time on other art styles @obtuse zealot

I got stuck on digital art and oil painting

[brisk anchor]

I had that feeling too, but in retrospect it was just the wonder decreasing

[wheat storm]

I always get good prompts with cats/dogs, better training data I imagine.

[raven obsidian]

there were so many choices

I get choice paralysis

[obtuse zealot]

The cat looks depressed ☹️

[sterile bluff]

my choice paralysis is in what tool I generative with

[sterile bluff]

am I gonna do some NightCafe, Mj, Dall-E 2, or Disco Diffusion?

[left pecan]

It is

[sterile bluff]

heck, I might even make something in WomboDREAM.

[obtuse zealot]

Hey as a volunteer do you get more credits

[left pecan]

Then there was these that didn't even look like the prompt I gave

I SAID CAT NOT DOG

[sterile bluff]

I received some credits for running #daily-theme.

[raven obsidian]

That's so cool

Well deserved, I'm sure ❤️

[sterile bluff]

it's a fun channel, I encourage folks to participate if the theme appeals to them that day.

[obtuse zealot]

Nice, I saw some of your art on your insta they’re really cool!

[obtuse zealot]

I’ve been doing th daily challenges

[raven obsidian]

I'm really hoping I can find a way to break past my self doubt and anxiety to start doing my own artwork again

baby steps

This is another one of my favorites

[sterile bluff]

I anticipate the volume will ebb and flow for #daily-theme now that generations aren't free.

bound to happen.

[obtuse zealot]

Yes I think so too

[raven obsidian]

So hypothetically now I could make picture books from my generations and sell them?

[wraith juniper]

damn, looks pretty good

[raven obsidian]

it really pops, it feels very unique

[sterile bluff]

yes!

[raven obsidian]

kewl

[obtuse zealot]

Cool thought

[raven obsidian]

Workshopping titles in my head rn

[brazen sierra]

do people win more dall-e credits by participating in the challenge channels?

[rocky burrow]

Use GPT3 to generate them ;)

[raven obsidian]

I think challenge isnt meant to be taken literally

I did that for one of my artworks

[sterile bluff]

just for fun.

[raven obsidian]

These were all generated from a GPT 3 prompt

[sterile bluff]

dang, those are cool.

[raven obsidian]

The prompt was incredible

[fast island]

I lost some that I thought I had saved. Is there a way to see all our prompts, even if we can't see all the images?

[raven obsidian]

Create a piece that celebrates humanity and its boundless potential. Think about what makes us special as a species and what we can achieve when we work together. Use bright colors and patterns to depict energy and movement, and consider adding elements from nature, science, and pop culture to create a rich and complex image.

--GPT-3

[muted osprey]

you can get these from your browser's history if nowhere else

[raven obsidian]

I asked it to generate some more

and it gave me stuff like "an animal running through a field"

I wish I were joking

[muted osprey]

well, it got Nature and Energy and Movement at least, lol

[raven obsidian]

Also check this one out, this is one of the only times I've seen DALL E choose to use Black & White

The prompt was very much not specifying it

A ship captain imbued with powers of The Great Leviathan Uk’otoa manning the helm of his large ship, digital art

[violet wigeon]

himwa

someone send me a ai generation of a face

I am going to blow your mind

[fast island]

That totally worked! Images and all! 🎁 Thank you!!!

Phew

[left pecan]

I got these results from
young African American boy with dreads around the age of 15 wearing a baggy shirt with a hip hop symbol on it , digital art , anime art , well lit , well done , beautiful background, highly detailed , detailed face , award winning

[violet wigeon]

alright lemme blow ur mind

[left pecan]

Lamo

Lmao*

[violet wigeon]

fixed face

[left pecan]

HOWD YOU DO THAT LOL

[obtuse zealot]

Did you use some high res fixer

[agile jungle]

He's probably using the ARC colab

[civic wadi]

Ooo got me thinking. I should upload DALLE images to Disco diffusion. I have a good upscaling method for that program

[brisk anchor]

do share?

[civic wadi]

Set your init scale / clip guidance to 5000 and 7000 respectively. Skip 150 out of 250 steps, turn on VITL14 model, eta to 1.0, clamp Max to .09, all else kept default

The relationship between init scale and clip guidance will determine how much detail it will add. Too high of clip guidance will start to add new stuff to your image. Too low of init scale will have the AI take over the image, there is a balance between to the two that spits out your same image but enhanced

[brisk anchor]

aside: do you go to georgia tech? you look like the inbetween of 2 people I already mix up that I know

[civic wadi]

I should make a video demonstration as disco has just gotten WAY bigger, a lot of extra settings that can be ignored

[brisk anchor]

Thanks!

[civic wadi]

Ahhh close, over at UF in Gainesville

[brisk anchor]

thats a nice one, yeah?

[civic wadi]

It’s beautiful here, but I am ready to get out of a college town haha. I’ve lived my whole life in Florida, but took a couple months last year to travel the country and really settled the west coast is for me in the future!

[brisk anchor]

Good luck!

I hope to be a remote worker and play the credit card travel game

[civic wadi]

Thank you! Let me know if you have any issues with disco, it’s a pretty daunting notebook at first glance

[civic wadi]

It’s the future haha!

[brisk anchor]

ive run 2 images on a disco implementation, they werent great and I gave up lol

better things to do

[civic wadi]

I’ve yet to mess around with portrait mode recently added. It’s got a midjourney feel to the results so far which is awesome. Overall my best results have been landscapes. With prompts, you are able to add weights which helps a lot. “A lemon:10 on a beach: 5” “clouds:-5” etc.

Takes way too much time per generation imo, but that’s just a constraint of colab

[stuck nebula]

👀

[elfin jungle]

I am interested in this aspect aswell. I am super behind with a project, could use some help integrating some results into a larger composition

[stuck nebula]

yes it's allowed

[elfin jungle]

oh okay, thank you so much. this just saved me 2 days of modelling and rendering.

[remote bough]

Apparently it's all good with the new terms 🙂

[elfin jungle]

tbh I have mixed feelings about it , like I am happy it's gonna speed up my work, but also OMG all the nfts are going to be just dall-e stuff from now on ?!?!

like fck my 10 years of training and education LOL

[frosty quiver]

Essential reading (Open AI) : https://www.reddit.com/r/dalle2/comments/w3qr3x/comment/igy8jgt/?utm_source=share&utm_medium=web2x&context=3

[remote bough]

true that! I reckon is unavoidable, but Im sure a lot of people will use it to their benefits and create incredible things... If you just use it as a tool rather than minting final just the prompts, that can be very powerfull

[elfin jungle]

I am curently looking into extending my current artworks, or creating some new items to put in there so I don;t render and model a bunch of stuff

[remote bough]

Like everything in life, lots of lazy people will try to exploit this amazing tool to make some wuick buck, but in the end, the real ones will remain, and these others will fade away I believe

[regal path]

Anyone else notice in the terms the section about Ownership of Generations, and that you won't own the copyright, I wonder if this is an oversight since it conflicts with the commercial use terms.

[remote bough]

That sounds very interesting

[elfin jungle]

can;t say I would mind the quick buck, but I never end up on that side of life somehow lol :))))

also really like the closed beta, the responsibility of it, the initial call, not into gatekeeping, but maybe I liked gatekeeping this one

[remote bough]

Yeah of course, but I don't think this is a long lasting mindset, NFT people are getting savvyer each day and noticing these things so I reckon it shouldn't be a problem in the long term for the ones who really use the tool properly

[civic wadi]

This plays into the images being able to be used by OpenAI still to train their models

[raven obsidian]

Really well worded Reddit post. Thanks for sharing.

[regal path]

But it says "will not resell Generations that you have created"...this is clear as mud.

[stuck nebula]

how is that reselling

[regal path]

Read section 6 of the terms: https://labs.openai.com/policies/terms To me it seems like this section needs to be re-written because it goes against commercial use.

[sterile bluff]

hiya, @regal path. hope you've been well.

it does sound like there is a conflict between the two.

[lofty birch]

These AI engineers need to get some more lawyers on this lol

[regal path]

Hey there, been well. I wonder if it's an oversight. I'm not creating any NFTs until this is cleared up.

[sterile bluff]

but I am not a lawyer, nor do I represent OAI.

[brazen sierra]

the animals it draws are too cute

[sterile bluff]

@regal path can you help further clarify your concern?

It's very similar to the earlier Mj licensing where Mj claimed ownership and gave you a license to use the images through CC

[regal path]

This line "OpenAI grants you the exclusive rights to reproduce and display such Generations and will not resell Generations that you have created".

[sterile bluff]

all right. what is your interpretation of it?

[regal path]

The "will not resell Generations you have created" sounds to me that you can't sell your generations as NFTs. I do agree that they are probably trying to have terms like MJ, but this is poorly worded or is old text they used when they weren't allowing commerical use.

[sterile bluff]

OpenAI << this is the subject of this sentence

grants you the exclusive rights to reproduce and display such generations << this is what they grant through the terms of service

and will not resell Generations that you have created << this is what they, OpenAI won't do

do you agree that in "grants you the.." they mean the creator/user?

[regal path]

Ah...I see now. I hate legalese, sometimes it's so poorly worded. I misread it.

[sterile bluff]

it could be worded in a different way, perhaps even bulleted.

[hybrid granite]

Besides NFTs, what good ways are there to monetize DALLE? I can't see too many people paying a premium for something they theoretically also could do with 13 cents

With things like watercolor paintings, nobody would ever pay me for a picture of one when they could instead pay for a real painting

[sterile bluff]

I've made a couple hundreds bucks this year in various commissions with other generative tools.

Minimal amount of post work on them.

[hybrid granite]

Is this something you'd still be able to do now with broader access?

[charred orchid]

cover art - t-shirt merch - any kind of merch tbh - concept art for whatever - illustration for visual novels - assets for video games or other visual media

[hybrid granite]

I'm certain, for instance, I could have sold images for cheap last month when I had something unique

Because it's not just "does this have value," it's "who would ask you to do this for them versus generating things themselves"

[raven obsidian]

Is it not possible to view your prompt/generation history on mobile?

[sterile bluff]

There are millions of people in this world

Billions

And maybe a million plus people who have made generative ai art

[charred orchid]

tilt your smartphone sideways, that way I can see mine

[raven obsidian]

Never would’ve figured that out thanks

[sturdy river]

im curious, how many credits do u guys have? Just click ur profile in the top right

[charred orchid]

141 left

[sturdy river]

ye im at about the same

[charred orchid]

I'm already regret using up one to complete an uploaded image of mine

cause the results were terrible

[sturdy river]

I wonder if the 100 we got for being in the closed beta is removed in a month like the 15 free prompts or if it lasts a year like the purchased credits

[charred orchid]

instead of "huh, I wonder how I can make the prompt better" I am now like "yup, fuck that - not gonna use up another prompt to tinker with"

prompt was
"Anime Key Visual of an egyptian Woman in a blue coat looking down in amusement, kneeling from the top of a balcony in a city with arabic architecture. night time lighting. stunning vfx."

[idle jolt]

Welcome to the world, DALL-E ❤️ ❤️ just got the birth announcement on email

[worn forge]

My biggest concern is that most of the results you get from DALL-E 2 are crap

So, they're charging you - for crap

They take a lot of tinkering

Really want to be able to use a quarter credit and generate only one image

[cerulean coral]

the 'appreciation' credits were counted as paid, with 12 months lifetime. Check on your profile on the website

[calm parcel]

https://tenor.com/view/sleepy-sleep-tv-land-bedtime-lie-down-gif-17912510

I used ~400 credits today...

[charred orchid]

show top 3 generations

[zenith crater]

iiii

how 😭

[calm parcel]

spent $

[idle jolt]

I think it might be DALL-E

's BIRTHDAY

🎂

[sterile bluff]

https://www.npr.org/2022/07/20/1112331013/dall-e-ai-art-beta-test

[zenith crater]

i mean how as in how do you even think of 400 things to do

[sterile bluff]

cat. cat in a box. cat wearing a hat in a box. cat in a box with a hat.

😉

[calm parcel]

i have an active imagination lol

also... I think I generated 1000+ teddy bears

[zenith crater]

mine is dormant rn

[raven obsidian]

Prompt: I'm going to bed

Result:

https://tenor.com/view/looney-tunes-quarantine-sleeping-sleeping-beauty-beauty-sleep-gif-16980796

goodnight everyone

[violet wigeon]

Gn

[rocky burrow]

Love the prompt! What style did you use?

[violet wigeon]

🤍

[calm parcel]

i need to continue tweeking my superbear

😄

[violet wigeon]

C:

[rocky burrow]

lovely results! What tags are you using?

[brazen sierra]

digital art

[violet wigeon]

I am from meow meow world

[rocky burrow]

Really? Thats the only modifier? Digital art is definitely inconsistent with its style lol

[brazen sierra]

prompt was something like 'digital art of cute bunny on car, sunset in background'

[zenith crater]

that one is so good

[calm parcel]

I literally started out earlier with "Space bear does spacey things in space while exploring space"... and evolved it from there lol

[strong valve]

Random question, is 15 credits next "something" means 15 credits every month or day?

[violet wigeon]

Months

[strong valve]

That's a pain, Thanks for your help

I feel like they can lower $ per promt by doing simple model optimization

Ow well now, There's always Opensource models

[violet wigeon]

dress flower food_themed_clothes fruit green_eyes hair_flower hair_ornament hair_ribbon kasumi_miruchi kneeling long_hair mary_janes original pantyhose pink pink_hair pink_legwear ribbon shoes sitting solo strawberry twintails wariza

[violet wigeon]

Someone put that in dalle2

I'm on https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2Fknn5.laion.ai&index=laion5B&useMclip=false&query=dress+flower+food+food_themed_clothes+fruit+green_eyes+hair_flower+hair_ornament+hair_ribbon+kasumi_miruchi+kneeling+long_hair+mary_janes+original+pantyhose+pink+pink_hair+pink_legwear+ribbon+shoes+sitting+solo+strawberry+twintails+wariza

This is where dalle2 gets its captions from

It may generate something super similar to the source material

[rocky burrow]

Not entirely where it gets all of its source material captions from, but its a good implementation of the type of tagging that CLIP can do

[violet wigeon]

Well does anyone want to try it?

[violet wigeon]

We cracked the code

[lofty birch]

Looking at that CLIP browser has given me some good ideas for why certain words were bringing up certain results before. It can be pretty useful for fine tuning prompt ideas

[strong valve]

The CLIP model that DALL-E 2 uses, uses a completely license datasets and different overall data than public clip models

[violet wigeon]

Oh :c

[broken spade]

Omg those penguins

[violet wigeon]

Yesssy

[strong valve]

Also considering Aesthetic ratings is taking a play, so overall results will be different

And the DALL-E 2 Filter

You get the point

[rocky burrow]

A few more penguins

Based off an original artowrk I did a few years ago lol

[brazen sierra]

don't know why it drew a little red hat on one of them 🙂

[rocky burrow]

adorable

love it

[violet wigeon]

Kitttysss

Abavabava

Someone send me a dalle2 face and I will make it look human

[civic wadi]

Dadaism paired with faces always gives me a gold / green color combo

[slate wing]

wow

[olive surge]

looks like florida man lol

[slate wing]

I am trying to transform my drawings. Although it is a 2D picture, it is not perfect, but it seems to be recognized to some extent.

재밌네

artificial intelligence is really cool

lol

[violet wigeon]

oh wow so it does know japanese

[slate wing]

the second image is really cool

cirno prompt?

I do not speak English well. I am using a translator.

I am Korean

[violet wigeon]

hmmm nvm then

[left pecan]

Prompt

[zenith crater]

[slate wing]

wow

OMG

[slate wing]

lol

[zenith crater]

oh that’s good…

[violet wigeon]

upscaled

[slate wing]

this is great

[violet wigeon]

that is amazing

[slate wing]

lol

[violet wigeon]

too bad I only have like 125 credits now...

I cant be wasting it no more

I really wish dalle2 could understand japanese because I bet alot of the source material from pixiv is in japanese

[slate wing]

haha

[hoary leaf]

It should understand other languages

[brazen sierra]

not sure what all those circle buildings in mine are useful for

[tardy granite]

blowing really really big soap bubbles for the locals

[brazen sierra]

haha yeah

maybe portals for traveling

[slate wing]

This is probably a dimensional gate.

[sage trail]

I was going to say portals, but bubble blowing is way better.

[slate wing]

Dall E who draws monsters well

wow

Game illustrations are created in 20 seconds. It's very surprising.

[violet wigeon]

T^T I blew through 30 generations today

[broken spade]

Did you have fun though

[violet wigeon]

nope :C

no fun only trying to get good generations

[rocky burrow]

@brazen sierra I really liked the idea of a city whos main export and production was bubbles

[brazen sierra]

nice

asked it to blow me some bubbles too 🙂

[tardy granite]

bubblecore™

[violet wigeon]

Pls dont hurt me

[brazen sierra]

it does a nice job with bubbles at sunset

wasting all my dall-e credits to see more bubbles haha

[civic wadi]

These bubbles are amazing

[dusky leaf]

is it just me, or after putting the words "fantasy painting" , it almost doesn't matter at all which artist you use

[opaque patrol]

Question regarding credits:
Do early testers (users before beta phase starts) will get 150 credits per month?

[cerulean coral]

Remove that zero, also, non stackable

The only pre beta perk is additional 1 time 100 credits

[opaque patrol]

Ah

I see

Wait, you mean stackable, the credits will reset?

[autumn token]

Imagine when they make Diffusion model using bloom

https://www.youtube.com/watch?v=8ozeo91_iJ0

hahaha

👀 Guess the market will decide the price of DALLE

[patent grail]

Do y’all have any feedback on the discord itself

[hybrid granite]

The Discord itself is fine

[brazen sierra]

it's odd the only roles are 'server booster' and 'online'

usually discords have admins, moderators, etc

[cerulean coral]

That is correct

[opaque patrol]

Oof

Gotta waste all my 149 credits (one for this authentication to get in this server lol)

[cerulean coral]

But, wait

The 100 "appreciation" token were treated as paid tokens, thus can be hold for 12 months

[brazen sierra]

i used my 100 tokens already 😦

[cerulean coral]

Ouchie

[opaque patrol]

You can choose which tokens you will use right?

[cerulean coral]

But again, 50/day to 50/month is absurd, the community has protested this

I hope they might revisit their policy for the limits, but won't hold my breath on that

[cerulean coral]

Free used first

[opaque patrol]

Ah

So it means i have used 1 out of 50 free tokens today

[cerulean coral]

You can check this on your profile, in case you missed that

[opaque patrol]

I mean tbh it's okay, I just using it for fun but not too much usage into it
I find it okay but might be useful for future project where I find good art for my music project in future

Cover art I think

[scarlet maple]

I was having fun making interesting things to test the model.

Now its only 15/month? lmao no I won't be doing that

Even the 150 we got for being beta boys is only 3 days of use before the pricing so like, bruh

I'll just hope to god I get what I want for TTRPG tokens on my first spin of the model and if not then xd rip I guess

(realistically it takes 5-20 spins to get anything usable too...)

[cerulean coral]

The problem of this newly imposed limits are on the way user interacts with the system. While abstract art and style / impression art is fairly a one prompt generation, many cool stuffs, like in/outpainting, fine tuning, and even edit, tooks multiple dice-roll prompts to satisfy user needs. Even the 50 prompts / day were already limiting the exploration (but still very doable), and now here we are, 15/month, which is basically the dead end on any prompt query exploration

[scarlet maple]

Its essentially impossible to make any of the really cool stuff without paying now.

And I can guarantee 90% of people that were using it before won't pay for it now.

[cerulean coral]

Quoting a sentence from a Twitter Space held earlier today, we're now limited to adding Homer Simpson on Star Wars level now

[scarlet maple]

But I suppose the crazy whales who dropped $750 on it for 5000 gens the instant they could are the real targets for the monetization so

[lofty birch]

Seems like it favors professional usages over hobbyists

[cerulean coral]

Also, they called it beta, but what part of this is beta? Some company pays people to test their beta, ffs

[scarlet maple]

The problem with that is it sucks for professional use tho. It doesn't give you what you want a lot of the time, its got a bunch of blacklisted words, the weird injected diversity stuff and its still limited to 1024x1024

[patent grail]

Any reason you didn’t do manual verification

[scarlet maple]

It says it'll take up to 72 hours for that iirc

[opaque patrol]

^^

This

[patent grail]

Okay

[opaque patrol]

I can do manual if I can

[patent grail]

I’ll see if it can be relaxed

[opaque patrol]

Just making sure

[edgy basalt]

In a few months there will be so much competition prices will have to fall.

There are so many other models that already do other things better

But I'm not mad at the pricing for now

[scarlet maple]

I was annoyed at having to use one of my 50/day tokens to get in here. I can't imagine how frustrating it'd be to use one of the 15/MONTH tokens would be.

or a paid one

[azure arrow]

Definitely agree that some kind of preview and select where to go deeper would be great. Wonder if there was never the option of 50/day for free if $0.13 per generation would be as polarizing?

[patent grail]

Tbf manual verification used to be the only option

[cerulean coral]

neither do I. I never complained about the pricing, since OpenAI said themselves that they will try various pricing model in the future. This is their first one

[patent grail]

The automatic one was added for people who didn’t want to wait

[edgy basalt]

DallE does the best cartoons and digital art style stuff, imo. Have you guys checked all of the other models though?

[autumn token]

remember ...... 20% of your generations are tampered because their "diversity"

[opaque patrol]

I already sent my email

[cerulean coral]

The problem is the trial period, really. Imagine you just joined right now. How can you even evaluate the usage if you can only test it 15 times a month?

[autumn token]

This is not totally true

[patent grail]

But I asked OpenAI if it can be made easier I will lyk what they say

Don’t want the server to become biased towards paying people

[brazen sierra]

there's only diversity added if your image has people, and you don't specify gender or race in the prompt

[autumn token]

Not true at all

[edgy basalt]

Which are your favorites?

[autumn token]

Disco Diffusion is my favorite

and it can be heavily customized

[edgy basalt]

For sure

[cerulean coral]

I remember one hilarious prompt that only have the word 'elves' that indicate a person, in a bedroom, and one particular image stuck out

[brazen sierra]

it describes in this blog when it's applied https://openai.com/blog/reducing-bias-and-improving-safety-in-dall-e-2/

[autumn token]

First there have been lots of post on this

and some did no include person

or even words

I tested with numbers

[scarlet maple]

Probably, but the problem is everyone got access to it for free with so many generations per day that 15/month seems (and is, honestly) extremely stingy in comparison. Like, it'd take $200/month at the current pricing model to get the same throughput as before. I don't think many people are down to spend $2400 a year on this.

[edgy basalt]

Post image I the art section, I'm waiting lol

[brazen sierra]

it didn't add any diverse people to all those images i made earlier of bubbles and futuristic cities

[autumn token]

ALSO .... tampering with prompt in mathematical term is inducing bias in a model

basic statistics

[azure arrow]

Yep, the anchoring is real and definitely not the best user experience move, if they kept it to even 100 per month for free with the same pricing the backlash would probably be much lower

[worn forge]

OpenAI will need to tinker with their prices soon

Midjourney is steep but much cheaper

I think the way they do their $30 unlimited sub is better

[autumn token]

Give it 2 years and DALLE will be behind

[worn forge]

Unlimited gens, but limited priority time

[scarlet maple]

I woulda been annoyed at something like 10/15 free per day with the current pricing model but the model still would have been usable for me.

[autumn token]

BLOOM is open source

[worn forge]

DALLE is already behind

Parti is like...
Way better

Too bad Google will never make it public

[scarlet maple]

15/month is basically completely unusable.

[worn forge]

What the hell do they plan to use that damn thing for

[autumn token]

Parti is not open source

[worn forge]

It's not accessible, period

[autumn token]

No worries, there will be accessible ones better

give some time

Huggingface great community

[scarlet maple]

10-15 per day woulda been like "Okay you can get 1-2 good results per day or make a cool outpainted image. Pay for more."

[azure arrow]

Yeah seems like it's only a matter of time before there's a race to the bottom on pricing - might be ill-informed but it seems like the real factors to make better output is more data, more compute, more training time, none of which are entirely defensible moats in the long-term

[autumn token]

DALLE is really porweful when prompt chaining

but at this price is not possible

[scarlet maple]

Yeah Dalle-2 is something like 1-3B references, wheras the Google ones are 20+ billion

20 billion is where it gives you readable text too

[cerulean coral]

Gottem
#openai-chatter message

[autumn token]

there are some research that at 170 billion parameters is when it starts understand deep context

[azure arrow]

Which to be clear is entirely nuts - somehow this stuff is still super niche but feels like it quickly will change everything? Wild times

[devout epoch]

did outputs suddenly get worse in quality today or am i going crazy?? i'm using the same prompts to compare and the quality is much worse

[autumn token]

imagine to give it a story and it understands how to build frames of it

[cerulean coral]

Tbh, it's probably dice roll (if not the diversity bias thingy)

[devout epoch]

i tried several prompts that i've already used and they were so much worse that i decided to stop wasting my credits

[autumn token]

their diversity bias thing destroys any attempt of making prompt chainng

[cerulean coral]

Ouchie, mind to share your prompt?

[autumn token]

it induces random bias into the generation

[devout epoch]

never in my life have i heard a legitimate reason to complain about "forced diversity" but i think this might be one

[cerulean coral]

iirc, diversity bias only triggered if there were any word that refer to human

alho, back to this (personally hilarious) attempt here :
#openai-chatter message

[scarlet maple]

I used my first token to see if mine were worse. They seem pretty comparable tbh.

[autumn token]

There is ppl saying I am racist know when I say these things

[scarlet maple]

It still doesn't know what a cutter is and continues to give me frigates

[autumn token]

but its an Autoregressive model, its the bread and butter of NLP

its like cosine similarity

[blazing atlas]

Not if you're specific to begin with.

[autumn token]

not true

[cerulean coral]

and the training data were just biased in nature, nothing we can do for those unless you go and curate 3B params, have fun

[autumn token]

but adding tokens is bad for chaining

more tokens not necessarily better result

[devout epoch]

https://media.discordapp.net/attachments/954859649226641408/999539817706692780/unknown.png?width=1920&height=481

today vs two days ago, maybe its just bad luck

[blazing atlas]

Those all look great

[devout epoch]

hopefully its all in my head

[blazing atlas]

I couldn't tell you which set is meant to be worse

[devout epoch]

hm ok

[cerulean coral]

without the prompt, we have no idea what to expect

[devout epoch]

“sonichu: the electric hedgehog pokemon, digital art” 😅

[cerulean coral]

Ah, I gotchu, first one seems a bit rounded

but still, I'd say dice roll

[devout epoch]

i hope youre right!

[autumn token]

well, the other point of forced diversity and filter is:
It is not open on github

[devout epoch]

i had a bunch of bad outputs that i didnt like so i took a step back and stopped using my credits

[cerulean coral]

I remember generated chimera of pikachu x charizard, lemme see if I can find it

[autumn token]

we dont know their list of forbidden words and we dont know how their diversity code workds

[scarlet maple]

I like the top set more than the bottom set imo

[autumn token]

we cant even make GPT-3 generate prompts.

[cerulean coral]

yeah, one of the few things they can do to increase transparency is just to list out what we can and cannot do, release the clip text query list

[devout epoch]

same, that was the two days ago one

[scarlet maple]

If you want something similar to the person who made Sonichu's stuff try adding ",traditional art, deviantart" to the end, maybe

[cerulean coral]

wait, is sonichu a thing?

[humble dagger]

hi everyone

[devout epoch]

thats a whole rabbithole you probably dont want to go into @cerulean coral

[scarlet maple]

Don't look it up you'll get internet poisoning

[devout epoch]

its cursed knowledge

[humble dagger]

how yall doing

[blazing atlas]

Pretty well 😅 fortunately I'm way too busy at work to worry about limited DALL·E prompts

[humble dagger]

ive been having so much fun

and now this

[blazing atlas]

60 booklets of poetry analysis to mark

[humble dagger]

honestly kinda disappointing

[humble dagger]

oof goodluck

[scarlet maple]

Yeah thats pretty much everyones reaction to the pricing model

[humble dagger]

maybe don't assign them next time

now im just like

save it save it save

[blazing atlas]

If only

[humble dagger]

and not gonna use

until now i was running prompts for friends and doing 50 each day within minutes

[bronze dagger]

oh god... it's him

[humble dagger]

i hate monetization

good night yall, take care of urselves

[cerulean coral]

monetization is inevitable, it is the limitation that makes no sense on beta version

[humble dagger]

yep

[scarlet maple]

Their shareholders likely got involved and demanded monetization.

Since everyone was like "aw but okay" about reducing to 4 prompts returned

then rapidly after it was "Okay now pay for it"

[humble dagger]

im still a student and cant actually do monetized stuff so while its inevitable i cannot like it ig

[humble dagger]

yeah fr

not cool

[cerulean coral]

I read before that OpenAI described themselves as "limited for-profit company", bunch of bs

[humble dagger]

as a dc fan i already had enough reasons to hate execs man

[bronze dagger]

people that ended up having to pay for OpenAI's GPT access: first time?

[humble dagger]

what is that

gpt?

[bronze dagger]

like DALL-E, but for text

[scarlet maple]

You probably heard about AI Dungeon, its that.

[bronze dagger]

you can generate all the fanfiction/articles/research papers/literally anything you want with it!

[bronze dagger]

🇫

I miss when AI Dungeon had access to GPT-3....

[humble dagger]

oh sexy

[scarlet maple]

Isn't their best model GPT-3? Dragon-something-or-other

[cerulean coral]

do they ever tho?

[bronze dagger]

unfortunately yes you will get banned for that

[humble dagger]

(i use "sexy" for every context except what its supposed to be btw)

im aroace, sex is a joke to me so i joke about sex

[bronze dagger]

imagine getting banned from using GPT-3 by exploiting it for fanfiction. couldn't be me!

[humble dagger]

lmao

[cerulean coral]

I remember the first days of AI Dungeon 2, we use torrents to download the pretrained model, should not be GPT-3, since that is proprietary thingy, right?

[humble dagger]

i just don't update my fanfics for two years like a real legend

[bronze dagger]

AI Dungeon did have access to GPT-3 for a very short time, historically known as "Dragon Summer"

it was godly. unparalleled. quite possibly the best AI-assisted fiction writing era in history!

Unfortunately, people used it to create contentious content and OpenAI taketh away

Now we're all stuck using NovelAI like Neanderthals. Truly the dark ages we live in!

[blazing atlas]

Gpt-3 isn't good enough for me, but I'm a writer and teacher of writing, so, makes sense

[autumn token]

I made GPT-3 write college level essays

[bronze dagger]

Oh! Also if you want full context, here's all the details https://www.wired.com/story/ai-fueled-dungeon-game-got-much-darker/

[blazing atlas]

It can do it, but they're shit

[autumn token]

haha what did u do ?

[blazing atlas]

Impressive sense of verisimilitude but terrible in the grain

[autumn token]

Did u just try to use a few prompts ?

no semantic embedding

[bronze dagger]

Oh! Speaking of which - you guys all have OpenAI accounts, so you can play with GPT-3 too! https://beta.openai.com/playground

[blazing atlas]

Admittedly I haven't had as much experience stretching GPT-3 for its best results like I have with DALL·E

[rocky burrow]

much more cohearant bubbles than mine!

[bronze dagger]

I think it got updated (relatively) recently. Sudowrite uses an older model of GPT-3 from my understanding, and it sucks. The version on the playground is pretty good at the moment!

OH!

Here's a thing I made it do

https://pastebin.com/D3wBvVc5

I suppose this is only funny if you know both shows but it manages to make some impressive, deep cut references to Perfect Strangers despite there (presumably) being no fanfiction or wiki knowledge at all to pull this information from

[blazing atlas]

e.g. this to me is so frustratingly vapid. it has the sense that some sentences were just auto-completed for grammar and cliche, not meaning. it's impressive when taken at a glance, but without cohesion or meaning that would make it useful

[bronze dagger]

ah, well that's the problem!

[blazing atlas]

on e.g. literary analysis, it's usually just wrong. it's very happy to say wrong things

[bronze dagger]

you told it to write an essay....

...you didn't tell it to write a good essay!

[/joke]

[blazing atlas]

tbh does that work? it seems to work with DALL-E

[rocky burrow]

The average reader only reads on a 9-11th grader level and has little reading comprehension. GPT3 is perfect for blog posts and social media bots and just general storywriting prompts

[sterile bluff]

would be funny if you prompted gpt3 like you did in CLIP

[bronze dagger]

I have no clue honestly, but it would be interesting to see if modifiers like that could change much!

[sterile bluff]

"a very very very good science fiction novella written in the style of Charles Dickens, elaborate words, trending on the New York Times Bestseller List"

[bronze dagger]

I do know you can make GPT-3 write like a child

OH!

[rocky burrow]

You can always use GPT3 to make DALLE prompts for you :)

[bronze dagger]

AI Dungeon actually had something like DALL-E 2 adjectives!

They have a whole catalogue of good terms to use and whatnot

just like DALLE!

[blazing atlas]

I tried it with "the style of Hemingway" and it definitely did NOT work

[sterile bluff]

sad

[bronze dagger]

[blazing atlas]

unless Hemingway often writes about "the hustle and bustle" 🥴

[sterile bluff]

We need a theme, folks.

Letter is "D"

[bronze dagger]

did you specify which Hemingway

I know DALL-E 2 will give me profoundly bad results if I use an old European artist and don't give them the full name

(sometimes)

[blazing atlas]

Dragons?

[sterile bluff]

did we do dragons? I know we did dinos

what about diorama?

[blazing atlas]

love dioramas

[sterile bluff]

we did dragon.

[bronze dagger]

I think diorama is a good one!

[sterile bluff]

glad you all think so, testing it now.

oh yeah, it's nice.

[rocky burrow]

Heres some diorama examples I did with a crochet astronaut

[bronze dagger]

🧑‍🚀

neat!

[tough fossil]

imo they should've gone with a monthly subscription model with a basic and standard tier for individuals with unlimited prompts but throttled speeds for the basic tier, and a separate tier for businesses with more than 10 employees or something. Knowing every prompt costs money kills your desire to experiment with random stuff

[buoyant wren]

June 30th vs today, not sure if theres a noticeable difference

[delicate shoal]

how's the promptin everyone?

[dusky leaf]

I had the brilliant idea of making Hitman with the Dall-E logo on the back of his neck, but to no avail.

Somebody wants to take over from here?

[blazing atlas]

Hard to know without seeing the prompt. Did you use the word "tattoo"?

[brazen sierra]

can't you just use paint to paste the dall-e logo over the neck?

it's a rectangle

[dusky leaf]

Yeah but the point is to make it through Dall-E

[dusky leaf]

Digital painting of A bald man with a colorful Dall-E barcode tattoo on the back of his neck

[brazen sierra]

i would try but i already wasted too many credits today

[autumn token]

thats a terrible example. Thats not how I do it

Huggingface is the about the become the new OpenAI 👀 there are some really nice stuff showing up that outperforms OpenAI

[blazing atlas]

So how should I do it

[autumn token]

can start simple

https://www.youtube.com/watch?v=HkJuFQmkQXM

it gets complex when you add more to. Like vector embedding, vector databases

Models to increase performance

and so on...

[blazing atlas]

fuck no lol. i spend my days teaching students to write critical arguments, i'm not going to spend more time teaching a model that's supposed to learn by itself. i mean, i appreciate that it's powerful and can be done, but if it requires that kind of explicit instruction the utility starts to dwindle

[autumn token]

well, I am pointing out it is possible

has been done

and soon will be a product

Chegg for essays

and it will outperform human performance

[blazing atlas]

perhaps, but in terms of depth of knowledge the proof will be in the pudding. until it's powerful, immediate, pre-trained and accessible, it can't have an impact even for a generally tech-savvy, AI-aware general user

[autumn token]

Oh

thats not true

depth in knowledge is added with part whole hierarchies

and being zero shot or few shot

means you can finetune it on the fly

The only thing I quite cant figure is mathematical reasoning

Generally I think that if you cant deploy a Hugging Face model, train and push to inference, then you are not tech-savvy

Dunning-Krugger Effect at its best

[blazing atlas]

That's totally relative. You're narrowing down the definition of "tech savvy" to a dwindling number of people that doesn't have relevance to general demographics. I can't do that and I'm in the top percentile or two of technical competence within my discipline (education)

I mean, I COULD if I wanted to, probably, because I have the capacity to learn and I'm not scared of code

I just have neither the time nor inclination

[autumn token]

Everyone can, thats the beauty of todays world

its getting more and more accessible

[elfin forge]

You could do what exactly?

[blazing atlas]

Well maybe. But I work with primarily middle-aged women who teach literacy and literature. "can" is irrelevant. if you asked them to "deploy a Huggingface model and push it to inference" they'd tell you to get fucked, in no more ambiguous terms

[blazing atlas]

run an ai model with open-source code I guess

[elfin forge]

This is a pretty bold claim.

[autumn token]

Hugging Face became the git of AI models

its getting easier and easier to run open source pre trained models

[blazing atlas]

what aspect?

[elfin forge]

What do you mean by "what aspect?"? I said that labeling yourself as among the top 1-2% is a claim you'll have to back up with something if asked..

[blazing atlas]

well i work in a school of 100+ teachers / administrators and my expertise granted me leadership in rolling out new LMS for example

I've presented on digital classrooms, was early in gamification and have used php for online portals back in the day before Canvas etc made it a bit of a waste of time

[elfin forge]

LMS is learning management system, correct?

[blazing atlas]

yeah

[elfin forge]

I thought you meant programming at first

[blazing atlas]

well i've dabbled in javacript, php, that kinda stuff, python and basic back in the olden days but not really much any more

which is why, as i say, "tech savvy" is relative

[elfin forge]

Oh, I see..

Well, to me, a tech-savvy person is someone who has regular encounters or interests in technology

[brazen sierra]

prompt: 'photo of a man dabbling in python in the olden days'

[autumn token]

genius

[elfin forge]

A credit gone to waste :(

[brazen sierra]

38 left before i have to stay paying the 13 cents

[autumn token]

so you dont spent per generations. You buy blocks of 115

[elfin forge]

Wow, look at this comic book villain I got..

[blazing atlas]

i love the way comic art sometimes does the thing where it extends beyond the frame

[autumn token]

BTW, you can remove DALLE thing on bottom

click inspect image

the DALLE label is added on browser, client side

[elfin forge]

Haha, cool

[blissful summit]

Here's the math for how much I think we made openAI spend: DALL-E 2 was announced on april 6, and May 18 there was an announcement saying there was over 3M images generated and I know there was 1,200 users invited by then (a post said 1K as of may 8, and they were adding 200 or so each week from april 6). So then it took about 42 days with on average 600 users to get 3M images made. And each image costs 0.03261USD (15USD for 460 images). So those 3M images spent 97,830USD ! Of course as always this prob costs openAI actually less since they Have to (should?) make money more than costs them to run the DALL-Es. So as low as 10,000USD. Now there is probably like 400,000 users or more. OpenAI I think has a few billion USD, but still is interesting lol.

[autumn token]

They spent millions training and fine tuning

[brazen sierra]

with 1M users i think they'll be able to get much better because of the feedback signals from the users on which images are good vs bad

[autumn token]

100k is change money

[brazen sierra]

e.g., if you save or download an image its good

[autumn token]

not true

quality breakpoint is model size

[blazing atlas]

They definitely have quality data through collection and shared images

[autumn token]

OpenAI is actively deleting images from the dataset that they consider bad

and messing with other stuff that creates bias within the model

Its possible it might get worse

[brazen sierra]

it can get better at generating what people want

[autumn token]

or very good at very specific things

[blazing atlas]

i think it will only get intrinsically worse at this stage if they feed AI images back into the data set

[autumn token]

Thats not true

synthetic data increases performance

[blazing atlas]

i'll believe it when i see it (that is, a functioning AI model working with majority AI generated data)

(with improved quality to current models)

[brazen sierra]

in a month they'll have a billion labels on dall-e generated images + whether each one was saved/downloaded or not

[autumn token]

https://towardsdatascience.com/the-magic-of-synthetic-data-using-artificial-intelligence-to-train-artificial-intelligence-with-cebe0d50a411

I barely searched it

its old actually

lots of research being done on it

[blissful summit]

as odd as it seems I was working on some older AI but had to stop for a bit, and I was thinking feeding AI its own thoughts to train on doesn't seem to improve its prediction ability, yet, somethign tells me an AI has to save its thoughts to improve

[autumn token]

It can increase performance

depends how u do it

[blissful summit]

if you take for example a markov chain, and make it train on some outputs, I had a hard time pinning how to make it improve the little model

[blazing atlas]

i understand that it has allure in the industry, i just don't agree that it will have good results. i can't quantify that, but again. personally the proof will be in the pudding

[autumn token]

markov chain doesnt work well with synthetic data

[blissful summit]

perhaps, when I think in my brain, and make progress on problems, over some hours, maybe I am really just training but in a long term way...? just an idea.

[autumn token]

markov chain is considered slow for standards today

[blissful summit]

(that is, training on the data my brain does have....however that would work)

[blissful summit]

ya I know, I was looking at it anyway though earlier, along with a host of other methods like related words trees etc, and may look a bit more into it later also

[autumn token]

Rafael Valle pointed out lots of interesting points on his research with TTS generation

early tts had markov chain

https://rafaelvalle.github.io/

[brazen sierra]

i wonder what the demographics of this discord are. Maybe 50% AI researchers and 50% digital artists?

[autumn token]

fewer percentage of AI researcher by the day

[blazing atlas]

unlikely. i am neither. i am sure that early onboarding valued diverse experiences with some preference to art and academia

(i was surprised i got access so early after clicking "other" on the profession list for the waitlist 😅)

[autumn token]

I will probably do it next time

[hybrid tartan]

I mean if you think about it, a GAN setup is mostly synthetic data

[autumn token]

GAN is so inefficient

when research trend moves into diffusion models

and other little similar things

things will get really interesting

[blissful summit]

i was thinking diffusion is best but then Parti came out, not sure why it looks so powerful comapred to DALL-E 2, seems as good, if not maybe better who knows

[autumn token]

its a 20 billion parameters model

has a better dataset

[blissful summit]

oh ya, i had that in my head but not this minute lol

still seems close though, or same then

[autumn token]

the gan or diffusion is only making the image

the secret sauce is the transformer

[elfin forge]

Wait, there is something better than dalle?

[autumn token]

https://github.com/lucidrains/parti-pytorch

needs training LOL

[blissful summit]

and imagen, also jsut made by Google

[elfin forge]

Cool, where can I get imagen?

[autumn token]

if 1000 ppl came together to make bloom which is pretrained and better than GPT-3

I am sure ppl will make a free better than DALLE solution

[blissful summit]

no access they said they won't but thye were letting jsut a few give requests weeks ago

[autumn token]

that github has it

not trained

if you push a pretrained model like that into hugging face you will start your own cult

hahhaha

ppl will love u bro

[runic ibex]

hey guys.. so.. im not sure i understand it completely, but does it mean 15 prompts PER WHOLE MONTH?

[blazing atlas]

Correct

[runic ibex]

that's so few! 😦

[autumn token]

then u buy in bundles of 15 usd

[runic ibex]

115 for 15 bucks seems quite pricey

[blazing atlas]

It was always going to be a paid service

[runic ibex]

yeah, but I expected something more reasonable

[blazing atlas]

I guess time will tell on the price point

[runic ibex]

but maybe that's how pricey it is for them, so they adapt it for users, dunno

and since when this limit will apply?

[blissful summit]

this one i wanted to make just a realistic photo, idk, came really nice

[runic ibex]

does the limit apply from today? and is there a way to see how many prompts we have left now?

[burnt mountain]

DALL∙E has just entered Beta. All changes starts today.

[blissful summit]

yep see sidebar, it shows 50 last promtps

[blazing atlas]

I think what they're probably monitoring are patterns of use amongst the longer-term beta users. I think it feels acutely painful for the recent influx users that their free toy was taken away within a few days, but I've seen a pretty high turnover in the Discord which suggests people get really enthralled for a while and then stop using DALL-E. They need to make money out of users who know how it works and need to buy prompts for a particular use-case, but aren't going to be using it on a daily basis, because that's more realistic.

[burnt mountain]

Yes! Click on your profile at the top right and you'll see your current credits.

[runic ibex]

😦 😦 😦 😦 😦 😦

[runic ibex]

ah, yeah, I see. thanks

[sinful stirrup]

:(

how much do credits cost?

I've had dalle beta for a while and have just noticed that there was a discord for it

[runic ibex]

what does this mean? that 19th of august is the month for me I suppose? that's when I'll get free credits once again?

[runic ibex]

15bucks for 115 prompts

[sinful stirrup]

is that good?

[runic ibex]

i think that's too pricey

for a few prompts

[burnt mountain]

Please do note that changes you see today may not be final as per Joanne. They plan to explore more options with regards to the pricing.
Found in #announcements :

Hi everyone, thanks for your feedback so far! Please keep it coming. This is the first pricing model and we plan to explore other pricing options that will align with users’ creative processes as we gather more data.

[sinful stirrup]

by 115 prompts do u mean you can type something and generate it 115 times or 115 images

[burnt mountain]

That would be 115 generations.

[verbal snow]

Thank you for coming!

[runic ibex]

I feel like for a priced model, something like $15 for 1000 generations would be more welcomed by people.. 😄 but I suppose that would be too much from OpenAI's POV

[fervent dome]

What went wrong lol

[blazing atlas]

prompt was probably a bit too much, especially with unnecessary specifics like "cereal" and "caught"

so DALL-E went rogue 😅 i'd probably rephrase it as "tiger in a grocery store aisle, CCTV footage" and if you're still having problems an inpaint might be needed

[brazen sierra]

hmm

it probably takes some iteration, first making a real tiger walking down the cereal aisle, then inpainting that into a cctv footage

a real tiger, but the wrong aisle

[edgy basalt]

Holy shit lmfao

[burnt hornet]

Has anyone noticed that some of their prompt tags that used to get them a consistent theme/ result no longer return that theme??

[edgy basalt]

No, mine are tremendously more consistent but I have become much more precise in my descriptions

My advice to everyone is to go read thousands of prompts that made great art

[burnt hornet]

Fair enough

[edgy basalt]

And develop your art vocabulary

[burnt hornet]

Yea I had an artist I like which always returned his theme but after the last 2 updates it’s stopped giving me his art style

[edgy basalt]

Ah, yeah, I can see that happening in some cases

[burnt hornet]

Idk if it’s because of these hidden prompts they are adding messing it up

[edgy basalt]

I think we are training the model in real time so expect change

[blazing atlas]

This won't work - start with the style and the setting (cctv of a cereal aisle), then inpaint the dissonant element (a tiger)

[burnt hornet]

Sad face 😦

[edgy basalt]

When you love something, really crank it out while you have a hold of it, so to speak

[blazing atlas]

Unlikely unless you're worried about different genders and races popping up in your white, male art. It's probably to do with the claustrophobic feeling of only getting 4 outputs

[burnt hornet]

Bruh

[blazing atlas]

Yeah?

[edgy basalt]

How many other models have you people tried?

I've messed with Craiyon (Dall E 1) and I'm on Midjourney but I want to explore more

[blazing atlas]

Only free stuff. Craiyon and Wombo Dream or whatever

[runic ibex]

any idea for the best place to find great art with prompts to it? or you mean just something like here on discord and comb through great and not-as-great arts just to see what's up?

[blazing atlas]

(Craiyon is not DALLE 1 btw, DALLE 1 was a previous OpenAI model. DALLE MINI was the name for Craiyon but it didn't have any link to DALLE other than maybe the model design)

[blazing atlas]

Local library, art history section 😎

[edgy basalt]

Specifically Reddit and Twitter. The actual prompt is in the file of the pictures when you download from, Reddit. You are going to pick up on all kinds of things the top 1 percent does that others don't. And yes, library art history section.

[edgy basalt]

Thanks

[burnt hornet]

tbh that is what im worried about. If i paid to see a white dude or an animal in clothing i want there to be 0% chance of seeing anyone else. Adding hidden human prompts could mess that up

[runic ibex]

i dont have twitter, but for reddit, you mean Dalle subreddit?

[blazing atlas]

If you're paying to see a white male you add "caucasian male" and you will get one

[fervent dome]

Thank y’all for the help! I still was unable to get CCTV or any other security footage style to work. I was able to get a fuzzy result though that looked like it was taken from a phone that was out of focus.

[edgy basalt]

Get Twitter. And yes, along with other art sections.

[shadow vine]

telling someone to use twitter? hm

[burnt hornet]

but with added hidden prompts it not guaranteed to favour your input over theres surely?

theirs*

[edgy basalt]

Weird when they come out blurry

[blazing atlas]

Read the blog post. It only adds specificity to counter bias when the prompt is ambiguous on race or gender

If you want white men exclusively, demand white men and you'll get them

[shadow vine]

"if you want a specific race or gender you can specify and you'll get them"

[blazing atlas]

Exactly

[shadow vine]

point being your replies are coming off rude

[blazing atlas]

How so?

[shadow vine]

You are assuming he wants white men

[blazing atlas]

He said directly he does

[burnt hornet]

no i didnt

[blazing atlas]

Well, regardless. You'll see a range of men and women and a range of races unless you add specifics, in which case you'll get specifics

[edgy basalt]

Anyone know why they just randomly come out blurry sometimes like his tiger? And why it won't let you rub out the whole thing in an edit to make it sharper?

[shadow vine]

If you want a very detailed face for instance it's easier if it's zoomed up. Otherwise can be blurry and same with other detailed objects.

And sometimes it just chooses an art style that's blurry lol

[burnt hornet]

you can try specifying HD but that doesnt always guarantee it will unblur

also i think if you rub everything out in in painting you are basically doing the same as just regular prompt generation

[edgy basalt]

It will be cool when tools like this advance to let you favorite perfered styles and prompt ammendments so you can say keep the framing of something and dump it into other styles you've developed very quickly

[rough lava]

Does the free 100 credits carry over if they arent used within a month? Anyone know?

[burnt hornet]

yea

paid credits carry

[shadow vine]

Yeah I also want a tool to find the art styles of other images so you can use it in your prompts

like if i get a random cool art style and want to replicate it in more prompts

[burnt hornet]

i agree we need access to a reverse dall-e, i.e. you give image it gives you the propmts it associates with that image

[edgy basalt]

I saw something by Google Imagen that looks like it may be in the near future. Looks really flexible

[edgy basalt]

Absolutely. It would be great to be able to just feed in 2 pics and say I want pic A in the style of pic B

[shadow vine]

Yep I'm sure there will be many cool features to make getting the right style you want in the future. This is only the beginning of AI art

[edgy basalt]

And even better to see the branch that got a prompt to where it is. How weighted some words were. The secret sauce

Absolutely

[burnt hornet]

add this to suggestions

[shadow vine]

I made a similar suggestion not long ago but more is always nice

[runic ibex]

just a random pic that I've managed to get through variations of variations.. which will not be possible with the limited prompts now 😦

[burnt hornet]

sweet pic tbf

[runic ibex]

the original prompt was "hyper detailed stylized portrait of knight in golden ornamental armor sitting on a toilet, scrolling through phone, bored, digital art, award winning, trending on arstation, tarot card" 😄

[burnt hornet]

hahahaha not what i would have imaged

to me that says like some sort of sci-fi fleet commander or ship captain

[runic ibex]

im on a different computer, so cant post the OG pics of it, but it looked interesting. and then through variations, this came out to be. in general it is similar.. someone sitting and scrolling 😄 but not a knight and a toilet anymore 😄

[burnt hornet]

hahahaha, it was a solid prompt idea tbf

[runic ibex]

I have two more pics here from those variations.. I made plenty of them and some were realaly interesting, but this seemed to be the most nicely looking one 😄

https://www.instagram.com/p/CgIHHursLK2/

Instagram

[burnt hornet]

third one is the best!

it really went for that sci-fi feel

[runic ibex]

yeah, there are some sweet sweet results when doing variations 😄

[scarlet crypt]

What im curious about is how much computing power is used by the generator and how much by the upscaler, if there was a way to just do one and not the other that might save a lot of resources

Perhaps an option to just generate one image for fewer tokens could be nice too

Just my 15 cents

[burnt hornet]

i think generator is majority of it

[pliant shard]

Don't take this the wrong way but... how would I make money?

am I not understanding something.

[thin comet]

fuck, I lost one prompt in verification

[thin comet]

in my experience variations lost quality

[opaque patrol]

Welcome to Cotton Candyland!

Every "houses" are theme park you can ride!

[buoyant wave]

sell images

[compact musk]

In my experience variants were never at the same quality level as the the original as well

[opaque patrol]

lol as i go deeper variants i found such a cool darker side of cotton candyland!

creepy but it really do feel like darkside version vibe

[runic ibex]

they do, but it depends on what you do vary. if some picture that looks really cool but has not many details, variation can bring the details in (although it will make it look more painterly). But yeah, generally too many variations can lead to a simplistic picture. not always
thats why 15 prompts is too little.. you cant even experiment with it anymore and you better pray the first prompt will give good results

[thin comet]

the first two are variations and the last is the original

it lost details that were given with the paramters:
godless heathen fantasy art painting, 8k, ultrarrealistic, perspective, high definition, enhanced image, award-winning, ultra crisp, 8k, unreal engine 5, trending on artstation, daily deviation, vibrant super detailed

[obtuse zealot]

I get images like this: https://labs.openai.com/s/oc21IUm7ZhJEt66Rds4HUIRi https://labs.openai.com/s/7GpL4bQCP09GKTUJBBWhCeQ2

https://labs.openai.com/s/o0GTHPB0lqF4a6nUbKao08rB

[thin comet]

but those are with the prompt

no?

not "variations"

[obtuse zealot]

Yes not variations

[thin comet]

yes, doing the same prompt again gets you good results

and also, you get 4 images when with variations you get 3. And the prompt is consumed anyway

[obtuse zealot]

I didn’t know you only get three variations but I have noticed from the beginning that variations don’t look as good as the original

[thin comet]

yes, is like it were "variation from this" without any parameters, when you have added all the parameters for better looking appareance

[obtuse zealot]

Mhm

[soft hull]

so you can sell the rights to any image that you make?

idk who would want to do that

but if it means money

[last hazel]

Getting a tattoo of one of my prompts!

[scarlet maple]

Man do people really just fill the prompt with the "please make this look good" words

[burnt mountain]

Sometimes, I literally do, with this modifier: "very beautiful picture". With that alone, it will produce dismal results. But with "very, very, very, very, very, very, very, beautiful picture" It will literally take your generation to the next level.

More info here: https://twitter.com/phillip_isola/status/1532189616106881027?s=20&t=pOIguwOJ5RDoYBj9zeEoVQ

Phillip Isola (@phillip_isola)

Language-conditional models can act a bit like decision transformers, in that you can prompt them with a desired level of "reward".

E.g., want prettier #dalle creations? "Just ask" by adding "[very]^n beautiful":

n=0: "A beautiful painting of a mountain next to a waterfall."

Likes

1158

Retweets

197

Twitter · Jun 2, 2022, 2:37 AM

[sage trail]

I'm not a lawyer, but my current reading of the TOU is that OpenAI keeps the copyright and you get a license to use it commercially. So you can't sell ownership of it, but maybe you could transfer commercial use of it? But not sure how that would hold up in court.

There seems to be some ambiguity in the wording currently though

[left pecan]

MY BDAY IS TDY

[subtle isle]

https://labs.openai.com/s/YuUsdJC2jTIW0DgCsv2Boeck
From a reply to someone on Reddit writing

"It's 2030, programming, painting, design and writing have all become dominated by the elite few that know exactly how to add the perfect amount of "verys" in CODEX, DALLE and GPT respectively."

Using the very trick quite a lot since then :D

[burnt mountain]

Man of culture! How many "verys" have you tried? 😄

[subtle isle]

For the linked one 12 verys and I think that also was the first version I did 👍 :D

[wraith juniper]

LMAO

[burnt mountain]

I've yet to try more than 10 I think! Would be interesting to max out the character limit with "verys" lol

[wraith juniper]

any adjective that you add to a prompt can be used as a "dial" that increases by 10% for each "very" you add

[zenith crater]

happy bday

[rich spire]

Hi

[zenith crater]

Hi

[next linden]

I got a bit crazy and went through 300 credits 😭

[hollow dune]

is it wise to give up your rights of the images and gift it to those who requested the prompt?

[next linden]

Wow love this!

[sterile bluff]

I wouldn't say it is either wise or unwise. Every situation may be different.

[hollow dune]

well when i generated a prompt for someone not like i care to own the image anyway

[scarlet maple]

adding enough very seems to destroy the original prompt lol

[civic wadi]

One thing we are not able to transfer is the share link. That original link will always link back to the original users account. This has me thinking..if I sell an image to someone as an NFT, can't I just be an ass and claim I still own it due to that original link?

[scarlet maple]

Yeah idk about spamming very tbh

[civic wadi]

I bet "extremely fractured" could add some cool effects in place of the "verys"

[violet wigeon]

Helloo

[violet wigeon]

Try adding cirno x20 times to see what it gives

I can't read every message from when I was gone my brain is going to explode

[scarlet maple]

https://labs.openai.com/s/ZraxlTMq59cKZ3Z5tXCH8Q3D
https://labs.openai.com/s/gcwql9rjBr2PF0EjpNVJIsqn

This is the kinda stuff I was liking from it before.

[violet wigeon]

You all are wasting your tokens :<

Don't do it like me I already have like 120 left

Oh okay good xd

[loud steeple]

have u tried "vray render"/"octane render" 🤔

i wonder how it compares

[violet wigeon]

Have you tried my prompt list

[scarlet maple]

I haven't and I'm unwilling to spend the tokens anyhow since I already have good results from the prompt. Plus like, I wanna use my prompts to make more TTRPG tokens since now that its paid I have to use them in useful ways instead of just having fun or testing the model.

[loud steeple]

epic, add octane render? 😂

or

[silver hazel]

“I’m a professional prompt engineer” Me:

[loud steeple]

"anime key visual by makoto shinkai"

[violet wigeon]

Try out my prompt list with those

[scarlet maple]

the day before it went paid I spent 30 tries trying to get a cutter or a sloop flying in the sky with a sail on the bottom and didn't get a single one, even with inpainting

[loud steeple]

i did a bunch of comparison

definitely mix and matching

[violet wigeon]

Wait where

[loud steeple]

between the websites u listed

changes stuff

[violet wigeon]

I can't wait to see!

[scarlet maple]

Like, that'd be exceptionally fucking frustrating if I'd paid for those rolls and got exactly 0 results.

[loud steeple]

yeah totally diff results from diff sites

[loud steeple]

did u try crayion?

it commonly does great when dall-e 2 fails

[violet wigeon]

Did you post them somewhere?

[loud steeple]

wut, no?

why

ok wait let me dm u everything

[robust bluff]

https://labs.openai.com/s/m7gypQCLOZYtgEy2wsi46YoS

[scarlet maple]

Eeh. It at least makes schooners but still no sail on the underside

They're all honestly fairly good but still not perfect

It couldn't make a manhack which Dalle made for me in like 3 prompts.
https://labs.openai.com/s/DZWMTmtfqoPhkSTtPtaDVeKE

Like, Caiyon only gets sawblade or drone,

[verbal snow]

Thank you!! 😀

[violet wigeon]

Using my method, generated by @loud steeple

[loud steeple]

apparently anime mage girl birds exist!

yes, epic stuff!

[violet wigeon]

https://labs.openai.com/s/MtWQfuKvhlsY6r2qDyp8f8Tr

Anime key visual of a white-haired magical girl with bird wings, cinematic advanced lighting, 4k konachan wallpaper, sharp image, highest quality image, trending on pixiv, trending on safebooru, 1px lines, advanced clip studio art settings, [incredibly detailed] [extremely consistent] professional anime fanart

Maybe the word clip studio art is throwing it off

[mild laurel]

I'm so sad... Was busy the last few weeks and haven't been using Dall-E. Get back on it today to find the new payment model... I really wish we'd been given a heads up this was gonna happen. I really regret not using Dall-E while I still could 😦

[violet wigeon]

:(((