#openai-chatter
1314 messages ยท Page 11 of 2
new verification system
@sacred jasper will be looking for mod volunteers in the coming days โ if everyone here could help us keep this server civil / sane thatโd be so appreciated!
Running out of prompts is such a kick in the gut
@warm flax BTW I didn't mean I meant AGI should be something that is not general purpose, I meant I'm thinking of training on lots of data and types of data. Rather I meant that humans each focus on a single job, liek one does say farming, and so he is made more likely to say farming related words. Think of it like this: train the model > then add a simple trick to make it likely to say farming words.
Or IOW this is Facebook's Blender AI, which does it.
Are you talking about fine-tuning (https://beta.openai.com/docs/guides/fine-tuning)?
An API for accessing new AI models developed by OpenAI
(What I was saying is that if one had a million AGIs, each I think would need to be made to prefer saying some words/jobs of focus, instead of how like GPT-3 talks about anything.)
Hmm, not sure about finetuning but I know for sure Blender does it really good.
Facebook's Blender
The invite only system definitely helps but manually approving everyone isn't feasible. I remember I had to message Natalie to join, now it's changed to Joanne ๐
That's why I suggested a bot...
its not getting better, they have rly wonky upload systems
like there no way of say, doing a batch upload of series, that have similar tags...it all needs to be done manually
and i hate that
one by one ..
is daunting ...
so is deviantArt good? they used to be, now...not so much
also the groups are very ....paranoid on AI ...most of them.
they do not like the idea
in general
I really like the closeness smaller discord servers have that large servers don't
i dont like midjourney's approach of discord being in that process of a prompt, i mean what if its personal ?
but now I couldn't keep up with it now
I started putting Midjourney notifications much more aside now
I once had this fire for Midjourney
first thing i do when i join a server, Disable notifications, Disable anything that rattles me XD
no signals
no signs
lol
And how are you going to select which ai to use given some context?
ppl react so weird about pings anyway
like i get it, too much is shitty
but ppl shoot of on ya if u ping once
as if u committed a crime
Yeah for big servers.. you couldn't keep up with what's happening
u ever visited one of those mega discords ?
I was there when Midjourney just had around 10 newbies channel
I like looking at cool image creations but I just don't have the time to look into thousands of them in Midjourney
Slack was proposed here earlier. Idk why it wasn't pushed through. Probably because of the ease of access that Discord gives?
i think midjourney missed the point of art a bit, like they do it as a factory does,
mass production
You can be assigned the AI_man for the job.
As for the future million AGIs, most would be assigned the common word in the dataset such as farming, food, and cars, and as you know medicine and AI are not so-so popular so those would be assigned rarely (for now). But intelligent algorithms can discover what words are actually more probable (related to food etc) and so AI should pop up as much as food, exactly.
Also: The "fathers" of AI will likely make more AIs have AI-related jobs/thoughts, but that's a good thing ๐ Also the AIs will be aware they are machines ๐
thats not what i hoped for XD
pieces in here ppl send all have something personally attached, coz its in a much smaller scale, makes also feel a lot different.
but ppl do mass images, it just looses that
whole feel
hmm should we just all switch over to slack then..? I chose discord bc i heard that itโs the hip thing to do these days lolol
There is no "simple trick" in ML.
Also, the word general means general here. It's difficult because of the no free lunch theorem. Improvement on one task means decreased performance on all other tasks.
discord is hip
Slack is for โ yeah what Barret said
MS Teams it is then
formal discord, in other words
I suggest OpenAi buys Club Penguin and we have a mini-metaverse for DALL-E
lots of Delivery companies use it to have a com line between hubs for quick info
@patent grail just volunteered to make a verification bot for us ๐ (but also no pressure)
I can make discord bots
great idea, will go ask finance for the millions of dollars to make this happen โ stay tuned!!
does that include the verification for associated email address?
direct to bill gates plz
Dalleverse
u want stable bots ? ๐
i dont think we have access to the email info ๐ฆ but we can ask people to generate prompts
I'm just going to extend @muted sedge ๐
Great idea
sure, the bot is open source!
Got a link?
nocie
letโs let @patent grail take a stab first and then we can talk about additional features etc that can improve it! ๐บ
Yeah I have most of the logic written but definitely watch the github, and feel free to open pull requests ๐
a team of Dall-E artists gather together in a lively conversation at the table in Club Penguin, digital art
That looks great, the embeds are so cool now
the problem with just having a million future AGI all have the same brain in full with no directed focus (so it says ex. medicine related words often) is then they all aren't doing somethign different, so there is no parellel work being done.
Hence, Facebook's Blender is the solution, or it seems to do the job.
all manually coded, no libs/nugets only the essentials to make them
@warm flax make. this. happen.
Penguin on the left, chicken on the middle, pig on the right? Now that's diversity.
I love JavaScript
was alot of work ๐
rather large, but she's stable as hell
Yeah the code is a bit of a mess rn, especially the compositing stuff Lol. If anyone wants to improve that... ๐
its js ?
I see nerds are in chat. I lost my MetaMask password. Some say to go through my laptop files and extract a "vault" sequence from a .ldb file.
Anyone has any idea what I'm talking about?
nodejs ?
Having multiple ai together is called ensemble and is nothing new
well if anyone needs a consultant programmer , dont be afraid to DM me
Maybe when I get time
i have lots of stuffies ๐
isn't that when they are working on the same problem though....I'm talking about different jobs, each (or a few) can work on
right now Barret is going to roll a green giant
and smoke that giant ....until he is no giant no more
๐
if you want i give you my code sauce for a better starboard, you'll just have to run it somewhere 
curious, what improvements to the starboard ๐ฎ
there was this code i created that reacted to all emojis, not just the standard star
Once again I am asking for a separate channel dedicated to posting unhinged theories about artificial intelligence ๐
like sometimes you want to pin something with lots of ๐, ๐ and maybe even the star variations โจ
STEP 1 create a thread
STEP 2 get friends to chat there
STEP 3 enjoy
STEP 4 block Skynet
#tech-talk
you can also create a thread :3
Too late for that now, she knows your dirty secrets we all seen picture of it ๐
To be clear I donโt really care to have the new thread, I just want it for others :p
damn dall-e, u scary >_>
you think this falls under bugs ? like ..decapitated without using that word
๐
Verifications bot is live!!
i was using gpt3's html knowledge to made a website centred around scott the woz
and when it was putting his full name
it put "Scott Wiederhold"
and for some reason every time i did it it would be a website that looked like you were signing up to some email newsletter
This is sick! ๐ฅ
u want the prompt? XD
Yes! Send any horror / occult / freaky / creepy / satanic prompts my way, please! ๐คค
oh rly
this one is surprisingly easy, Miniature in a Bottle. Monkeys. and then all the realism/render style tags
Thanks mate, will have fun with it!
i think i once saw a prompt of you, and i borrowed parts coz i didnt know what the term rly did
maybe u recognize it ๐
Baphomet?
Heretic? Also Caprcirons are cool โญ
DALLยทE 2022-07-16 22.42.34 - Hellenistic Capricorn. Mage Yokai. Detailed Neoclassical. Detailed. Trompe-l'oeil. Art Nouveau. Detailed face expression. Liminal Cinematic lighting
so i have to credit the person, but i cant find it rn
but its not my prompt originally
whoever it was, knew exactly what he wanted
๐ฏ
I believe we are heading towards no-credits times ๐
well im a fair guy, if its not mine, then i say so ๐
and what about skulls
they fall into ur alley ?
Nope, this is very soft-core 6th grader for me ๐
I love occult / secret society / satanic imagery
ah ..
oh i know exactly what u mean by that
if you are a true fan, then you know of these very old titles that refer to death, not as "real" but tries hard to be
Places of Death or something
let me look it up
I'll take a look, thanks!
Faces of Deaths
Faces of Death (later re-released as The Original Faces of Death) is a 1978 American mondo horror film written and directed by John Alan Schwartz, credited under the pseudonyms "Conan LeCilaire" and "Alan Black" respectively.The film, shown in a documentary-like style, centers on pathologist Francis B. Grรถss, played by actor Michael Carr. The na...
they are pretty trippy
like i watched this as a kid ...and i was pretty fucky after that, like ya know a good month at least
:S
Ahhhh I'm loving this ๐ญ thank you, friend
Any idea where I can watch?
Lots of new channels appearing ๐
Im loving the new channels
?
all thanks to @silver hazel, who actually created a whole another example server for me w/ his brilliant ideas!!
What is the new channel about?
He do a nice work
not only one channel, there are a lot of new channels here
He is very supportive of the community and has made good contributions.
Guys Iโve been having an awful day and this honestly cured my depression ๐คง
if I'm aware of it, by the way I saw one that said art_Chat
Yes! To discuss art we like, without taking away from people who want to showcase their generations ๐ผ
I really thank you for what you do, and for your support of the community, I personally am very grateful to you and keep making you want
understood and received
thank u guys <3
I stole it from @sour python , so you stole it from me or him
guys wouldn't it be funny if i were like "last night before we invite XXK people to this discord!" and then tomorrow no one ends up joining ๐ถโ๐ซ๏ธ
people are constantly complaining on my DALLE tiktok that they havent gotten accepted yet
I didn't get the link for the discord in my invite email, maybe that's it?
oh yeah we still haven't integrated it into the invite email!
but we do include it in the p.s. section of product update emails
also, major thanks to @burnt mountain for designing our fancy new server icon (it's a gif!!)!!
looks good
you could also put the button with the discord icon if you can before creating the images
I wrote an essay (or whatever you want to call it) about DALL-E and I would love to get thoughts from other people who have used it! https://medium.com/@reichkrystal/part-i-what-is-dall-e-99401777bfa9
you can always just train it endlessly if you have the resources there is nothing static about machine learning
although you need be careful with catastrophic forgetting
~~since we now have an art channel why not also add a channel for music sharing ๐ ~~
Hey everybody ๐ค
#daily-theme is the heart of this server and thanks to @sterile bluff we get to experiment & observe the unique results & perspectives of so many people
Iโm kicking off the @dalledictionary Instagram with an alphabet campaign where each day we will play with a random topic in an alphabetically chronological order (A,B,CโฆX,Y,Z)
Iโm stoked to announce that @sterile bluff will join me in this alphabet journey and #daily-theme will announce random A,B,C topics for a whole rundown of the alphabet from A-Z!!!
What does that mean?
You can continue participating in #daily-theme as normal, this is just a one-time event! The themes will still be 100% curated by @sterile bluff
You can join me on our Instagram journey where we re-gram each Instagram story / post that follows the daily-theme!
If youโre already uploading your daily-theme results you might as well tag @dalledictionary for a guaranteed shoutout of your daily generation!
Greatly grateful to @sterile bluff for making this happen, as I believe this synchronisation with the Discord community will give birth to something very fun, insightful and beautiful ๐
Yโallโs #daily-theme results are always so cool and interesting and I think it will be super cool to spread our community art outside of Discord! โค๏ธโจ
ill give it a star so this pops in the starboard
i wasnt here gor a couple days. wait something changed?
new channels?
@silver hazel is admin now???
Hi, this is a very temporary title and will be removed in a couple of hours!
wow roles are colored now?
i was gonna suggest that lol
well thats good
problem though, roles still aren't displayed seperately to other members
@silver hazel check #off-topic ๐
making good progress
if only Dall-e team can allow me to use the api without constraints (limit) that be awesome
that would benefit everyone here
"Acquisition"
It shows a lot more people per image now, but now it also only seems to have white people?
before it had more
before it would more often show people of other races
They might be trying to roll it back after people figured out they were adding words to prompts:
That was yesterday- today things are looking different
AIs are so bad at hands lol
almost finished it
also used a new lil gimmick, all question keywords how, when, and what are basically the prefix itself
so its easier
if anybody wants to know the method im using, let me know
I want to know !
ill dm ya
Hey. Do you all want to brainstorm some custom emojis for the server?
So, it turns out DALL-E 2 was static all along, not being trained more - they just added words to prompts that weren't there
i dont think the process is just one thing, it consists of many sub processes that do things
maybe one part is a fixed state, maybe the other part is not
Helloooo
Well that's sad, 4 images instead of 6 now
just received an email
It was inevitable
๐ข
๐ญ
QUICK GUYS
oh well still grateful to be part of this
Generate 6x while you can
uhhh
still remembering the good old days with 10 :/
Hey @warm flax do you know how many images will DALL-E give us when you go public?
I mean, it's a remarkable service and currently free...
I remember when it was reduced from 10 to 6 and someone just didn't refresh the page and kept generating 10 results ๐
.(
It is, but now we're getting less than half of the results we got at first
๐ฆ
BRB doing this ๐
that's barely any at all
i am crying
10 results per prompt and no daily limit
maan the nerfing dalle2 has been through
Grateful for the beta access ๐ There are thousands of people willing to sell their soul for a single generation, so count your blessings yall
It isn't nerfing, you're a beta tester for a product you haven't paid a cent for...
if you wanna be fair that doesn't mean it isn't nerfing
A Discord invite was included in the email. Is that the first invite sent out (in mass) since June 22?
yes
and more people get to play around too 
if we survived 10 to 6, we can survive this too yall
"Person holding up a sign that says" ๐ค
true i suppose
theyre still changing prompts btw
Yes
Dallexia is improving to be fair
hi everyone 
hie
read emails
you got one from openai
annas what do u think of this change i can see you typing 
New people! Letโs act normal yall ๐๐
#misc-images is a good place to put favorite generations if they don't seem to fit in the other Topics?
i will pay my soul for endless generations ๐
I don't think you understand the true power of collection'ing prompts
For one you can share them at any time
To avoid making that the bottleneck
and if you save to collection you can also see all other generations too
using a little trick
was there a post w/ the technical details on the mitigation?
wdym adding words you have to train the network more to update what it generates
dalle is great, only thing I miss is generating real people
nah they just take your prompt and append words before putting it in the neural net

Hi @patent grail! Excited to be here!
hihi new person!!
wdym
i don't think they do that
Really?
"A person holding a sign that says" will consistently produce corruptions of the word "Female". https://labs.openai.com/s/dK0PeSihMPOOJ6TTzWeCvbtE
Read #announcements. Joanne just described how it's done
i didnt even think about dalle x coding train,
that will be amazing if its not happened yet ๐
It has not!
I see, apparently they only change it if it is left unspecified
haha, wasnt sure if i missed it. staying tuned!!!
still isn't really giving us total direct access to the model to be fair
Although it still needs work
ah
what about "Photo of martin luther kind jr"? does that still produce women some times?
If one has access to DALL-E through the openai labs, is that the original, or DALL-E2?
dalle2
There still is the problem of prompts like this producing the word "Female" (or corruptions)
my banana is sad too https://labs.openai.com/s/OrZNjRXrGG2e4eWhlEWqD5AP
(in this case FAMELEL
)
who run the world?
Famelels

The Burger King advertising an 86% discount coupon because nobody goes to his restaurants anymore
ViT-H text encoder the one dalle2 uses can only handle 77 tokens 1 tokens is like 3-4 words it would be a waste and also what words they will add anyways they just train it more
oh god not the random percentages again
Oh I just noticed it's a "96" not a % sign. They look quite similar
4 images per prompt now
Do we have Elon Musk here? ๐
do you know which one is more expensive
hdd storage server or computing electricity power
yeah 
We need Elon Musk here ๐
Check #announcements
um i just turned into elon musk 5 seconds ago. hi im real elon and im here now

Adopt me please ๐ฅบ
Also take me to mars
<3
The two spotify artists who made songs called "OpenAI Dalle 2" really finessed googles search engine lol, it is always the first thing that comes up for me
hi as speaking from the view of elon musk 2, i will happily adopt you and take you to mars
give me free money first
(im definitely real elon)
๐
"OpenAI gracefully scales down to zero images per user, grants access to entire world"

Yes definitely real
๐
Any other game developers here, or just me? ๐
anyways, nope he isnt here , even if it would be cool if he is
Ik that
@admins we want Elon here ๐
This is just the "demo" in their blog though
Lmao
who else would pay for dalle2
reminds me of the youtube memes, i'll try to predict the future
2022 DALLE: There are 6 images per prompt for a user
2022 JULY DALLE: There are 4 images per prompt for a user
2023 DALLE: There 2 images and 2 generated advertisements per prompt for a user
2024 DALLE: There is 1 image per prompt for a user and the site is filled with advertisments
2025 DALLE: You can watch an advertisment to generate a prompt for free daily
Hello ya'll!
Ai dungeon vibes
XD
2022 DALLE: There are 6 images per prompt for a user
Winston: Hey guys
2022 DALLE: omg its winston from overwatch
we should reduce the number of prompts to 3 or increase to 5
as a weeb, i find the number four offensive
/s
elon leaves the company https://www.bnnbloomberg.ca/elon-musk-left-openai-to-focus-on-tesla-spacex-1.1215616
Im still getting six generations
wait really
It's starting later today
Probably triggered
Elon left OpenAiI a while ago I thought
๐๐
you didnt know?
oh 2019
yeah
Me after seeing the 4 output limit:
Such updates carry positive news! This means we are closer to getting a public paid DALL-E, which means unlimited (well, adjusted to your budget) prompts, freedom and most importantly commercial rights!
๐๐ I canโt wait
I'm curious is we lose access to DALL-E at some point
commercial rights?

i like
eyyyyyyy
i watch ads anyways to play mobile games so i'm used to it
no thanks
XD
I hope they have something like gpt-3 to adjust how much $ you want to spend monthly on prompts
If we are paying per prompt, that is
hah so true they did a lot of modification in the internal representation
Tbh some of these results look more like a solution that goes on top of the trained model
Like these
Yeah this is a pretty lazy way to diversify it
Seems like theyโre actually just tampering with prompts
They just put another script that modifies your prompt sometimes that acts BEFORE the first text encoder
meaning that the model hasn't been changed
(because it's a static model)
Couldnโt the data be relabeled and the model retrained
All that would be needed theoretically to diversify it is for images of white people and men to be labeled as such instead of generic
That would be a good solution
Take stuff like "Female lawyer"
and just make it lawyer
Nah that would be the other way
may i ask if there's anyone here with access to google's imagen and can give us a comparison between dall-e and imagen?
I think the bias is because white peoples and men are largely unlabeled as much as other groups are labeled that way
Has anyone talked to someone who works at OpenAI about this?
They must have thought of it at least
the neural network is static, not the values/training data
I made one imagen result once, a google employee named (I think) William Chen did my prompt
Changing the training data now does nothing, it's already trained
And if you were to retrain then it wouldn't be static
i need access to imagen
same
i need to try it all out
It would just be a new static model
My guess is they donโt want to retrain it because it would take a lot of time and computational power
Which is lazy
Honestly disappointing from such an amazing company
I'm not disappointed
That they took the lazy route ?
just mad
also there's a thing called
how is it called
wait
i forgot the name of the bias
damn wait
how's the thing called where you can erase existing data via training
the bug so to say
ยฏ_(ใ)_/ยฏ
Same thing happened for me, got kicked out and can't log back in
Who broke DALL-E?
Thanks obama ๐
Did dalle just go down for anyone else?
I just waited a few seconds and hard refreshed the browser CTRL - F5
For everyone
Huh. Did I just get manually verified? I haven't done my discord code generation yetโฆ
ah they must be switching to 4 generations per prompt nowโฆ
๐ญ
yeah it went down for me
basically you can by accident retrain existing data and make the ai forget stuff
(no prompts
forgot the name
Lmao wtf
(glad I don't have to waste one of my 50 generations today)
Overfitting?
Hey guys.
Huh kicked out of dall e
someone brought up the term here, now i forgot
A big right?
into it's new form
Bug*
I wonder if they'll reset our 50 prompts for today, that would be nice
Or temporary at least Jesus please
I got a quick question. Will DALL-E 2 be free forever?
(4 generation dalle2)
nope
Why not?
LOL!
in your dream bro โค๏ธ
Expensive
yeah, these people have just worked their arses off for nothing
Because it costs money to run that much computational power
OpenAI is open source, no?
Electricity isnโt free
because corporations care about money
Itโs not
and dall e is like a gold mine
it takes a lot of resources to train the dataset, generate the images and host them
DALL-E 2 should be open source for everyone.
That way people can use it on their own machines.
a lot of things in life should be a lot of things
I agree with you
Catastrophic interference, also known as catastrophic forgetting, is the tendency of an artificial neural network to completely and abruptly forget previously learned information upon learning new information. Neural networks are an important part of the network approach and connectionist approach to cognitive science. With these networks, human...
found it
and that is never going to happen
Money will always win
oh yeah
catastrophic inference
It's kind of funny why they call themselves "Open" AI
They should change the name to "Closed" AI
well they are fairly open in comparison to imagen or so
I think dalle is back
Also mentioned in the first "bias reduction" article posted by OpenAI that it is even harder to make the AI forget stuff it has learnt, such as stereotypes
Wow! that's the first time i have heard anybody say that! ๐
I doubt the average gaming computer would be able to generate images as quickly or as good as what we get with the service. As far as I understand it takes a lot of resources (other simpler models are already resource-intensive)
A 3060 goes a long way with Disco Diffusion
But we should be allowed to generate images for free too...
This is why theyโd have to retrain the entire model
Relabel all the data and do over
If they want to diversify it correctly
Paid for on their own servers.
Free to use on our computers.
Sure it would be slower on our hardware, but we should at least have an option...
Code is considered intellectual property, you canโt force anyone to share code they wrote nor should I
4 generations also means inpainting and variations will only produce 3 results
You*
lol dont worry they overcame this tho ๐
But yea Iโd love if it was open source
yeah
People at LAION are trying to make open source dall e 2 based on the paper
tbh i think the community should do that
I guess we'll just have to wait... Eventually, something will happen...
or like partake at the very least
How??
It could be done way better with an AI
give us the tools to help and they review it
Detect white people and men in the data and label them as white and men
Problem solved
I would love to work with them to label datasets that would be awesome, even just as an internship
like recaptcha does
but who will train the thing that detects it
So far, I'm both impressed with DALL-E 2...
Openai could do it so easily
What other AI programs do PPL use? I have Artbreeder, Midjourney, Wombo, StarryAI, and started off using DeepDreamGenerator. So glad to be using Dall-E 2! The wait was well worth it.
And also disappointed
Classification is east
Much of the dataset is human labeled
Disco Diffusion!
DALL-E 2 is great to make concept ideas, but never for finished results.
the community should also take effect in the generation style. i think the current "recommended" way of generating images is too verbose.
I use everything on kaggle cuz I'm broke 
You wouldnโt have to relabel the images completely just add the keywords white and men
@silver hazel And I will prove it right now
see dent's message, that's the point
i don't think so training a model on 1024x1024 resolution is almost impossible, they just train the base diffusion model the 64x64 one
Iโm trying to learn that. It's a bit convoluted ATM. Iโm unable to understand how to run the program on google drive
I mean the ones where you have to identify features -- and for this case having humans may help?
Literally adding those two labels to pictures of white people and men would balance the data
I like using DALL E to give me inspirational ideas on how or what to make to make for art.
the super resolution model is great what do you mean
I think it depends on the art style youโre going for and your personal liking ๐โบ๏ธ
if they would train a network specifically on these, np at all
with the right wording you can definitely get finished quality outputs ๐
Just have to try lots of times and maybe do some inpainting with dalle
but yeah can also maybe depend on what you're trying to achieve.
Some restults are definitely more difficult if not impossible (text)
This looks great, right?
On the Disco Diffusion discord, there is a guide called "Zippys Cheatsheet", this has everything you'll need to get started. I personally pay $10 a month for Google Colab pro to get access to a Tesla P100 from my Macbook air.
Edit: send me a DM if you have any questions I would love to help, I run their official instagram so I have been involved for a while
Yeah, it doesn't...
It looks great when looked at far away
Until you zoom in and see this
Thanks I joined the discord this week. I'll search for this recommendation
This smudgy mess
See what I mean?
It's great for concept ideas.
But for finished results, nope...
Hello
that's from the upscale model i think
well if they want it to be updated and relevant it can't be static youtube algorithm is never static it's constantly learning to keep up with trends
yes but you can try to first generate a closeup. with "extreme macro closeup" and then outpainting it with the "pad" command in #dall-e-bot and then building from there to get a high quality image ๐
How the heck do I do that with DALL-E 2?
I kind of feel like some of my results are less impressive today than they were last week, in general.
go to the discord channel here called #dall-e-bot ๐
then go with the "pad" command and then upload the padded image to dalle2 again.
Then paint in the transparent section describing the image again.
agreed
i consider code as art
๐
i liked this thanks for adding more space in the collection
Outpainting is my favorite community made feature. The in-app tool teased in the cosmopolitan article has me excited!
Does anyone have any tips for generating realistic illustrations of faces? Or inpainting better faces?
try โportrait ofโ with some camera lenses!
for faces I always start with "a professional portrait photograph of . . . "
there are some pinned in #โโtips-n-tricks
Thank you!
I'm super happy with all this โจ
hey, hello รต/
got access 3 days ago, and it is amazing
I'm generating pictures of dnd characters, a lot =P
wished you could generate a completely new canvas though from giving an existent image, for example upload an anime girl and you could say something like "This girl fighting a dragon, same artstyle" and it would actually do that
if that's happening dall e won the market
Technically you could do that, just not all in one step
that would be kinda dangerous too tbh, it's possible but then people would use it to create fake images of certain people
if I could request something, I would also request it
sometimes I want to replace something but I don't want something different I usually try to let some of the original image for it to use as a reference, and then do a second inpaint pass to delete the original
they reduced images per prompt to 4 but didn't add a remaining request counter ๐
Who can I speak to about a deactivated account? I know I messed up with some violent prompts but I'd love a second chance :/
uploading celebrities remains filtered by dall e though
What about sharing images of realistic generated faces?
sometimes I wish I could upload photos with faces..
You can share realistic faces under the same rules as sharing an other DALLE image, make clear intent that it is made by an AI. However, you can not share/create images of already existing people
You can do, "a realistic looking face of a construction worker", but you cant do "a realistic looking face of my IRL construction worker friend, Hank Builder"
"A famous movie star", but not "ryan renolds"
I tried once uploading a photo so I could replace the background with a cool cyberpunk city...
oh my god guys you won't believe what just happened ๐ so i opened in youtube listening to spooky music and to my surprise an advertisement appeared with happy music and because i didn't expect an ad i got scared ๐
hey, if I used dalle to generate something, then I composited it with another software on a picture with a real person, am I breaking the rules?
Hey guys! I'm new here - does anyone know when dalle will be taking applications or allowing for people to do commercial use with it / build products?? Curious on a timeline or where to find that info
Hank Builder is best builder
how about the non celebrity ones then ๐ค
afaik dalle will stop you from uploading any human
I dont want to speak on behalf of OpenAI, but in the past I have seen videos doing exactly that. Again though, they made it clear they were working with DALLE as reference
oh, sure, I just want to see what people that has been here longer than me knows about it
thx
you can avoid that by freezing some part of the network and also by scaling it up https://openreview.net/forum?id=GhVS8_yPeEa
ahhh alright that's good!
human faces
they just don't want you to replace features on someone's face
I hope I can get back into DallโขE 2 soon
i made some amazing creations and i miss it
made like 800
if there's any mods maybe I can DM
hey, question.. it happened that a completely fine prompt could generate things that are not very nice (and would break the terms)
is my account at risk? should I flag it if the prompt was ok but generated something that may be interpreted as not ok?
Better to be safe than sorry I'd think
I was trying to find an example to link to ya haha, realized it was posted on OpenAI's instagram story the other day. It was a really cool edit altering the background of a photo of a person using DALLE. That being said I assume that process is approved
dont work at openai, but did you get banned? just curious
huh
But you still got the email?
yeah:/ sucks
I got invited and got the email today about their policy update but yah can't access my account or my gallery
oh ๐ฆ
thanks
youโll have to understand how this network of a bunch of numbers work though
sorry sorry ^^
i appreciate that bro!
Reduced to 4 images per prompt ๐ช๐ช๐ช
could ya do Super Mario water color painting?
tbh i think 4 per is fine you can always run variations
i am!
Itโs not fine when I canโt pay for more gens per day!
I just pushed DALL-E 2 a bit too far
๐ if you could pay for more i bet many would
I gave them this prompt "squid shaped creature translucent upper body gelatinous rippling sac like weather balloon lower body with long whip like tentacles flying in the sky"
Iโd be more fine with 4 results if we got more daily prompts to compensate tbh
yeah go for that I wont ask for a lot
lets see the result
This is why DALL-E 2 will never overthrow artists...
Thatโs the problem for me, dall-e isnโt consistently good enough to support that many results but only because I canโt have any way of generating more images
thanks!
It drew a squid
adorable
interesting bach
how did it avoid copyright rules
oh yeah, that's interesting
wait is copyrighted stuff intentionally sabotaged? Iโve genuinely never run into it if so
hopefully ill get back from support.... I'll let you all know! thanks :)
oh, I was just adding, didn't meant to correct, sorry if i was rude =P
sure
In the past, users who have been banned have been able to reach out to team support. Every flag is manually reviewed (with the influx of users this may not be the case anymore?), and they are very understanding. Assuming you are not trying to create some vile imagery (lol), reaching out should fix things
all good all good
by correcting me you help clear up misunderstandings
so its a honorable act!
My offbrand Homer
I'm one... and there are others from what I've seen people say in their intros ๐
thank you! I did try some cartoon violence (i.e. words like fighting or war) but nothing ever too crazy
We shall see
Great guess, friend!
Sup
@signal oasis just got the invite like 2 days ago. Finally made it off the waitlist!
based
Man I love this site so much
sup jewel same
Definitely feel that! I got mine right around the same time. Signed up the first day too lol
๐
I have to wait a few hours to use it again now : (
I signed up in April and got invited June 18th
i used it so much ๐
i love taking deep fried memes and running variations on them to make brand new obscure memes ๐
Cool, good to know ๐ any ideas about using the results in games? I emailed support, but no reply yet (been a few days). Presumably, it's not OK for commercial games, but unclear about free games.
aw yeah please do buddy
I'm gonna check out the results in some of the channels
hey all! glad to be here!
I wouldn't use it for a free game yet or anything public facing. Rather be safe than sorry in terms of copyright stuff.
I might use some of it for closed playtesting and non-marketing stuff just to give an idea of what I'm looking for in the future. Basically just placeholder art that might fit the end result.
well no one knows how it work but at least we know that it encode image and text into high dimensional space where something with similar meaning are closer together https://colah.github.io/posts/2015-01-Visualizing-Representations/ that's why people just froze the top layers of the network and leaves few bottom unfrozen
Another crazy DALL-E generation prompt!
"a panther with silvery gray fur and bat wings flying in the sky"
I love how sonic the hedgehog prompts give sonic a new design
Yeah, so far I've been using it to generate ideas when I'm not sure how to draw something. It would be very useful in a lot of public-facing marketing things (e.g. Steam capsule images for updates), but the repo readme made it pretty clear that it's non-commercial and testing-phase right now, so I'm on hold
Yup
I've gotten into all the other OpenAI betas within the first week or so, even placed top 10 in their Codex challenge. I thought my place was secured for the same time frame lmao. ๐ Love seeing what it can come up with, can't wait to incorporate it into my app ๐
that's awesome bro! what app are you working on
Hey there!
Even in non-commercial stuff I wouldn't use it if it is public facing simply because it can lead to issues. Like someone posted a small behind the scenes stuff on a thing they plan to release years from now but still got called out for having placeholder images that they didn't have rights to because even a "behind the scenes" video on YouTube is still technically "marketing" their game and/or "marketing" their company.
There's a link in my bio with a bit more and some previews, but it's basically a social network for creators with a built in AI studio that should feel familiar to Adobe users in pro mode. Does more than just art though, I've got studio tools for creative writing, music generation, etc.
๐
I just killed DALL-E 2
Yeah, I'm holding off. Hopefully everything will be clearer/permissive once they ship the for-profit licensing. For now, it's just good for generating ideas for stuff.
I used the prompt "the fourth dimension"
What do you think of Codex?
I love it, definitely works well if you kinda already have a general idea of what needs to happen. Was able to make multiple "games" in the Codex > web demo without much issue. I've been using CoPilot for a long time now also.
The Codex "inpainting" is really nice and would love to see that on CoPilot, I was in the CoPilot beta since the beginning too
You can go to #dall-e-bot and apply a cool PFP Watermark using the commands!
nice how'd you do it
This is my logo
hi
I painted over the logo
oh thatst sick
And told it to generate this
I also got it to generate me some versions for Christmas and Halloween ๐
you got all your holidays lined up thats sick haha
Those are really cool lol
This actually shocked me though, since it's not supposed to be able to spell at all!!!
How the heck did it do THAT!?
You need valentines day and saint patricks day now too ๐
That is so interesting. I personally don't code but my roommates do and it has been so cool introducing them to this stuff
words that are more than just words often show up! Maybe not everytime lol, but "congratulations" "welcome", "merry christmas", etc are all words that are more than just words themselves if that makes sense, really sick what you are doing
DALLE Seems to be terrible at generate text. Atleast when I tell it to it always has spelling mistakes
Anyone else got the same problem?
its doable but not ez
it will cost more prompts then u are willing to part with
in my opinion
The coding stuff is really cool, the possibilities is crazy with the direction of everything. Looking forward to an open version of Codex fine tuned on Bloom, I feel like it's inevitable.
AI can code for ya, no problem
it does alot for me
sometimes im that lazy i ask it to do all of it
no joke
difference is, now im writing a doc more then i write code
AI can teach you code, if u desire so
Definitely, I jump around up to 6 languages a day for my various tech stacks and it's sooo helpful remembering the syntaxes
it knows all lang XD
Uh-oh...
I think I overloaded DALL-E 2
โญ๏ธ Looking for volunteers to come up with some fun and cool Emojis for this server! Letโs generate a bunch of them and we can do a community-wide vote soon? #998607250488840222 ๐๐
been using this for a while
great shit!
like a freaky Dall-e that knows all code
lol and where do u think Copilot is based on? ๐
right
but Github has trained it with all the open source it has
all of it
imagine that
some ppl start to cry about License problems, it only touches those that have the proper license ofc
would be very dumb idea to not do that
yeah this is like common knowledge by now
oh wait you're new here
welcome!
I recently used copilot a bunch to help a CS undergrad I mentor with asm
could be confusing for new coders
that i have noticed
I was creating some items for a chat companion and it was fun seeing it auto complete with various things it thought I wanted. I used to make procedural games awhile back, would love to use AI with things like that. Between the image generation and language models continuing to get better we should see start seeing a lot of things like that I imagine.
but once they get it, they love it
well it kinda makes sense that them getting money from our code is kinda problematic
@modern linden I mean I used it as an oracle to help me, do the mentoring.
๐
I hadn't done asm myself in like 15 years.
no Github knew this prior, they taken the right actions, perhaps Dall-e team can consult Github on this, they have a firm grip on Licenses
Codex is sick
I'm currently working on projects using generative AI to do some parts of games or something else but it's quite hard to control the outputs
hardcore memory allocating
we got a Discord bot in my server that is directly hooked to Codex/OpenAI
u can query it in the channel
@modern linden same
Ngl I left that server cause of the weird slutshaming in the nsfw channel
A little incelly
What an interesting verification system
Greetings from Grenada ๐ฌ๐ฉ ๐๐พ
you just told everyone you went in the nsfw channel ๐
Assembly is old trash, it was made at a time when it was good trash but now it's just weird syntax for reverse engines
Dear god
im currently testing it
well old trash, but as the saying goes, one mans trash is one mans gold
I guess recently a lot more people got access?
so i can test if its worth it
Codex JavaScript sandbox is really good
Can make games on it so easily
Yeah well I did some in uni and it's not my gold
This is a lie, it's so hot in the Netherlands right now
We use GPT3 to do classification of the question first, then send the question back with an appropriate prompt with a few-shot examples in that classification.
hahaha true XD
GLORY TO JAVASCRIPT
i use Ada ...smallest one
Having to manually allocate every functions arguments is not funny
Has anyone tried the Yandex GPT-3-like model that went open source a few weeks back?
iirc it requires like 200GB VRAM for inference
@modern linden not sure how you're replying
mistake perhaps >_>
I made a persona engine that gives consistent results keeping with a characters personality throughout a Convo. Are you trying to zero shot the responses or using some examples or a fine tuned model?
You created this bot?
Interesting. Is hosting the model costly, or do you use the API?
using some examples
Oh so this is how the bias works... :|
Not my image
Yeah I can imagine, Copilot has been crazy helpful with my thesis code lmao
I host a similar bot that costs me like $5/mo on Replit
i got the free grant, im still at 18 dollars XD
How do you go about hosting that? A GPU enabled VM racks you up more cost than 5$/month.
WHAT, that's the fun part though!!!!!
also i selected a handful of ppl that have access
yeah
so not all can do that
copilot is handy with maths especially
its my best friend ๐
It is helping me quite a lot with my neural network architecture in Pytorch
saves me so much time
I'm using an API for inference
indeed
what where ur expectations ?
Ah I see, I thought you manually set up the MLOps. Thanks
what did you expect from intelligent google search though
Something magical. Something that could make what I can imagine... However, it turned out to be another typical stupid program that isn't smart, as usual, because computers aren't smart. They're just a mathematical machine.
but it has more then just "look what i made!" its truly a form of art
i get the feeling you are using it wrong, my good sir?!
coz there is no way in hell someone thinks this is boring
Not for that, but other things I have access to the TPU cloud and I use lambda labs for GPU cloud infrastructure by the hour
I would love to start using DALL-E 2 in a commercial setting for helping generate 'image'-captions for social media
@modern linden I was just saying, what you can do is have a prompt like this:
Classify the questions into one of the following categories:
1. Code example
2. Code explanation
3. Explanation
4. Expansion
Q: How do I do list comprehensions in python?
C: Code explanation
Q: What is the capital of france?
C: Explanation
Q: What's the weather like there?
C: Expansion
Q: {user question}
C:
That way you can have GPT3 tell you what kind of question the user is asking. Then, once you know, you can send the question back to GTP3, using either one of the language models, or the codex model, with various parameters and a custom prompt. Like if GPT3 decides the question is a "Code example" question, then you can send it back to GPT3 like so:
Give code examples for the following computer science questions:
Q: How do I define a class in TypeScript
A:
```ts
class Foo { }
Q: How can I swap to variables in Python?
A:
x, y = y, x
Q: {user question}
A:
goddamn dustin XD
is the "unstable" fixed yet?
Sorry, been thinking about this a lot ๐
Cool!
oh i can make that all in depth, fact is this bot was equipped with OpenAI that i made myself today, its running rn on a prototype
What's neat is that when GPT3 categorizes a question as an expansion, you can access some history you're tracking to send back the previous interactions as the prompt suffix, etc.
No, it's working right
i made this framework once
goddamned ........
i love ur excitement, but this is a bit too many pics in sequence ๐
just 1 will do
Because, after all, it's just math.
its whatever u want it to be,
It's not an actual person doing artistic work.
i cant tell how you think it should work
nobody can
i think u looking at it from the wrong angles
It's a computer morphing imagery and trying to make something through a bunch of mathematical formulas...
What do you mean wrong angles?
hey Dallists
It's smudgy
dude stop complaining and just enjoy what u have
Smudgy
No no. I have no bias. I'm saying the cold hard facts.
And the fact is, it's just code.
it depends on the prompts as well, check the comparison with dalle 1
"improvement" ?
It's still in the end.
Stay nice bro
A program.
no need for rudeness
he pisses me off bro :<
Doesnt mean insults are warrented
ur right
A soulless machine gluing imagery together to make something.
progress is progress no matter how small it is
Some would say that's what we are too.
Yes, but, I'm just stating what it is so far.
but this nutsack here, is obviously trolling.
If you aren't religious, then thats what we are
I mean if you want to get philosophical that's arguably true of humanity too
i personally like the pixel art it outputs
I'm saying it figuratively.
"Soulless"
why would there be a soul needed?
As in, no actual physcial intelligence.
like wtf does that even have to do with this ?
nice
High five ๐๏ธ
What is physical intelligence if not some metric over computation?
What the difference between humans and a machine?
Dall-e never claimed to be a fully autonomous drone or some sort of droid
so stop comparing this stupid things
its bs
We have the ability to tell when something looks off or wrong.
also there are different powers for generation
According to our prior experience
you can check the pricing tier/tokens for dalle 1
This search bar is the original Dalle and not the Dalle-2 correct? Im just a confused simpleton
They're worthwhile discussions to have, but they're better had if parties are not dismissive
we merely got access to a public beta
An A.I., at the end of the day is just lines of code, generating imagery from a bunch of pre-trained images.
DALLE uses CLIP which does that
Yes
sure, but he clearly has no idea wtf hes talking about, i feel like hes mocking you too.
Sure, but lines of code and wired neurons are, at the end of the day, just computation.
^
But, somehow we don't make art like that, now do we?
Art like what?
We can imagine new things.
You're essentially complaining about the resolution of the images
No.
No one is going to be moved by that
But i thought the color scheme for the dalle-2 was black
wtf is he doing anyway, barging in, act like some sort of preacher, nobody asked for this
You said "it's smudgy"
that's maybe just the website
Yeah that's a matter of resolution and model size


