>OpenAI this week deployed a technique that forces DALL-E 2 to reflect "diversity"
>also it’s now rejecting attempts to create the likeness of public figures
>OpenAI won’t allow DALL-E 2 to be used to create images that aren’t "G-rated” or that could “cause harm”
>Google won’t release its AI-generating model, Imagen, due to "risks of misuse"
>Meta has limited access to Make-A-Scene
is this the end of an era?
>OpenAI this week deployed a technique that forces DALL-E 2 to reflect "diversity"
>OpenAI this week deployed a technique that forces DALL-E 2 to reflect "diversity"
look at how the ai community persecuted my boy who released gpt-BOT, they're a bunch of sniveling cowards that can't face reality.
Man that was hilarious, tons of people in the AI field claiming that it would somehow cause "harm" to people by existing, yet never actually specifying how. Unironically a wrongthink moment
that stupid bot spamming BOT clearly harmed my usage of BOT.
It only spammed 4chan, so nothing of value was lost.
They where mas because BOT dataset out perform everything.
Turns out using npc talk is shit for AI development.
EleutherAI and LAION to the rescue
Eleuther has a chud working on it. Just check their discord.
So what's your alternative? Shut the fuck up if you can't contribute to the discussion
Seethe harder chud
Heheh, that's it? Not surprised if "chud" is the only thing you can come up with. Let me guess, after this, you will continue to say "chud"
Realistically there are no alternatives.
FAANG & Co donors will not support another open lab with generous compute resources.
Some universities might be able to compete.
CompVis research group at the Ludwig Maximilian University of Munich lost SOTA status with the release of DALL-E 2 but isn't that far behind.
the only Open thing about OpenAI is its name
it's a profit driven corporation that won't share anything with the public and have huge investors like Microsoft behind it
bullshit, openai open-soruced the core component of DALL-E:
is not like finding the included libraries and compiling, retard
only thing that matters is the training dataset. do you have access to it? if not, you can't make shit
yes, Google's Imagen was literally trained on LAION-400M:
That's literally the reason why Google won't release Imagen:
However, a recent audit of one of our data sources, LAION-400M , uncovered a wide range of inappropriate content including pornographic imagery, racist slurs, and harmful social stereotypes.
This finding informs our assessment that Imagen is not suitable for public use at this time and also demonstrates the value of rigorous dataset audits and comprehensive dataset documentation (e.g. [23, 45]) in informing consequent decisions about the model’s appropriate and safe use.
Pretty sure Google has access to every fucking thing in the whole internet. why would they limit themselves to using publicly available cc licensed stuff? what's the strategic advantage of that?
retard pls, LAION-400M isn't just CC images.
>turns off safe mode and "remove voilence"
>tons of JAV pops up
Breh... anyone could code dalle. Its really not so difficult. The hard part is actually training it.
Sharing code means nothing in AI.
Just make your own.
it's already charging 15 bucks for 115 prompts, it's over
clearly wanting money for their work is what break's the camel's neck
well it does for me, I'm not paying 15 bucks for less than 3 days worth of prompts
was hoping it'd be a subscription service that wouldn't gouge me or something
you mean all the free work the comminity did before they were bought out
you mean all the PUBLIC data they harvested for their models
but not their models, weights, ratios
the only thing that matters
its going to take other models years to amass the same amount of data
reinventing the wheel because openai sold out
>wont gouge me
just store and host petabytes of data he says
>just store and host petabytes of data he says
Then stop crying about how Satan doesn't want people to make cool shit with the AI.
>>wont gouge me
I hope you're winning the fight agaisnt this version of me that exists only in your head
It does considering how many rerolls and fixing of prompts you need to get decent results
Can't wait for someone with balls to release a pre-trained model of the danbooru database from https://www.gwern.net/Danbooru2021
You can already get some cool results from https://thisanimedoesnotexist.ai/ but I want to see boobies.
In due time anon, in due time
Such a model doesn't exist.
Training it would cost millions and take months.
The worlds top minds at work here, folks.
crowdsourcing a la [email protected]?
I really need AI generated anime asses
file.png is from craiyon which already is free and open:
yes anon I know, it's my own screenshot.
What I meant is that I need them in way better quality
why not use the actual game engine instead of training some AI to replicate it badly?
because the AI can reproduce situations that the game engine cannot and it does so with less effort for the end user.
I don't want screenshots, I want illustrations and fanart, they're way hotter
not the best example in this image but you probably get what I mean
with dall-e 2 just using the word "artstation" would give you professional results. no lewds allowed, though
yup, it boils my blood how much they fear sexualization and how much time and work they spend to stop a coomer like me from cooming
You are trivializing a really serious issue: Imagine the ammounts of CP that anybody could generate using his favorite celebrity at any age, in any indecent situation
It could ruin a whole generation of young men
Well anon you see, I just don't care. I'm in a mission to get anime ass and ethical problems aer irrelevant for me.
anon, for photorealism GAN models are still unrivaled.
Diffusion models as used by DALL-E 2 still have problems with faces.
>Bad people (1% of people) will do bad things therefore we must keep the not bad people (99% of people) from doing not bad things
who cares? victimless crime.
>said no non-pedo ever
So no one besides other pedophiles agree with you, anon
you say that like it's some kind of gotcha
this isn't reddit or twitter
That’s right, you pedos should go back there, where you came from
get the fuck out conservacuck
>pedo is shaming someone for being conservative
kek when you cut a liberal a pedo bleeds
i got the impression most pedos were conservative?
let's count hands: i'm apolitical
just because someone you don't like or agree with says something doesn't mean it isn't true
Uh oh, someone doesn't know their origin stories.
Doing this would be a totally victimless crime. The urge to stop people from being able to do things like this that you don't approve of is simply puritanism.
the only victim would be my penis XXXXDDDD
Or it could drastically reduce demand for real CP (which is produced via child abuse) by providing a legal and ethical alternative. Of course, that's not going to happen in our lifetime. Instead they're increasingly restricting access to and criminalizing e-boi hentai to make sure the pedos go underground and find worse stuff. Gotta protect the fictional children, after all.
>Or it could
nope, after we legalized gays and trans their numbers exponentially increased
fuck off libtard
no real studies can demonstrate that, because cp is a taboo topic, and no one will ruin his career trying to make those studies
also, the government NEEDS the pedo hysteria to demonize, incarcerate, and extort key individuals
Ever since internet access (and with that access to child porn) became widespread, child sexual abuse dropped drastically (like cut in less than half). Pedos are clearly willing to settle for less if it's safer, and inversely pedos will obviously go for riskier stuff (or acts!) if there's no safer alternative. This has already been demonstrated for junkies in a comparable situation: Providing easier access to soft drugs without hard drugs reduces consumption of hard drugs, while strict enforcement against soft drugs makes them turn to dealers which also provide them with hard drugs. But as you say, nobody's willing to do proper studies (and those who do get shamed out of the scientific community if their results are not in agreement with public opinion), so we'll never get further than statistical correlations.
The people who made the PROTECT act, along with many other politicians all around the world. I never said they wanted to prevent child abuse. They don't do jack shit to prevent it and actively make it worse by fearmongering about pedophiles (making sure they lose all social support on being outed, meaning they have nothing left to lose and are likely to become psychologically unstable), preventing pedos from getting help (via shit like mandatory reporting, pedos get reported to the police by medical experts for admitting they're pedophiles) and removing all alternative outlets for their urges. It's almost like they're proactively trying to incite child abuse.
can't be the US government which literally did this:
"In 2015, The New York Times reported that U.S. soldiers serving in Afghanistan were instructed by their commanders to ignore child sexual abuse being carried out by Afghan security forces, except "when rape is being used as a weapon of war".
American soldiers have been instructed not to intervene—in some cases, not even when their Afghan allies have abused boys on military bases, according to interviews and court records.
But the U.S. soldiers have been increasingly troubled that instead of weeding out pedophiles, the U.S. military was arming them against the Taliban and placing them as the police commanders of villages—and doing little when they began abusing children."
This would actually save real children in real life from having to suffer sexual abuse.
that's right goyim, we- I mean you don't want that CP do you? think of the children!
*gets on a plane to epstein's island*
I will be employing this strategy
can't stop me from cumming, I'm going to print out and mail these to the openai office every week.
Don't ever stop, anon.
I'm working with the Danbooru2021 dataset right now but honestly training GANs is pretty hard.
The biggest problem as far as I'm concerned is that there is no clear metric to measure progress.
It's hard to tell whether problems are related to the architecture, the hyperparameters, bugs in my code, the training data, and/or the fact that I only have a GTX 1070.
probably mainly the gtx 1070, this is what's used in the big datacenters for ai training now, check out the specification section of it, shit is a technological marvel and it's only getting better: https://www.nvidia.com/en-us/data-center/a100/
wrong, GPUs are legacy.
TPUs leapfrogged them.
>probably mainly the gtx 1070
Well, I'd say it depends on what you're trying to accomplish.
State of the art image generation is definitely impossible with 8GB of VRAM.
Usable image classification on the other hand would very much be possible and this has already been done for the Danbooru dataset.
Generation of images with limited diversity (faces/portraits) is also possible, just not at high resolutions.
Endless porn to fulfill all our fantasies…
All the research papers exist, just make your own. You can even make it a distributed effort. [email protected]
What doesn't exist is tens of thousands of A-100s for us to train them on. It'll only be viable to at-home-train these at least the GPU generation after this one.
lies, TRC exists
>just make your own
i saged. report me
NOTHING will happen
I will NEVER follow the rules
I will NEVER modify my behavior
I will ALWAYS say moron gay and chud JANNIE
You will NEVER stop me
It's not censoring when is not explicitely the government doing it, duh
>just make your own [schizoposting]
Yes, duh, unironically. You are not entitled to the fruits of others' labor, you fucking commie.
It costs tens of millions of dollars to train the models. It's not as simple as just writing the code.
no it doesn't lmao, it's a few thousand max to rent the gpus from amazon.
GPT-3 cost tens of millions, coming from OpenAI themselves. There don't appear to be public figures for DALLE-2 but it's probably similar. Why do you think DALLE mini is such dogshit?
That's including the cost for "researchers" and "advertising" and "programmers" and "insurance" and "social media managers" and "office supplies" and "rent"
The cost of training the models is significantly cheaper.
>For example, in early October, researchers at Alibaba detailed M6-10T, a language model containing 10 trillion parameters (roughly 57 times the size of OpenAI’s GPT-3) trained across 512 Nvidia V100 GPUs for 10 days. The cheapest V100 plan available through Google Cloud Platform costs $2.28 per hour, which would equate to over $300,000 ($2.28 per hour multiplied by 24 hours over 10 days) — further than most research teams can stretch.
A100 GPUs are 2x faster than V100s and about the same price
pls help me with this, where do I even get the papers?
everything is pay walled...
Praise the Russians.
not him, as I am even more retarded that him it seems, but where do I find the paywalled papers?
entering "text to image" in sci hub retruns nothings...
seems I need a papaer title first or I am simple too retarded
Yeah, you obviously need to know the paper to find it.
>everything is pay walled...
this has never been true for CS.
arXiv always existed.
this is their new technique kek
Wow, so it will always generate female morons huh? Diversity == female morons
Diversity is anything nonwhite.
>hampering and ruining technology based on NA's gayry
Muttland has been ruining things for almost a century at this point.
Holy fucking shit is this actually their method? Here I was thinking they retrained the model or something. Reminds me of "40%" popping up in every single image a few weeks back.
that doesn't even make sense. ain't that particular AI really, really bad at depicting text?
It's good enough to write the word black on a on a sign from a prompt that says
>a sign that says black xxxxxxxxxxxxxxx
>people no one likes
>worst people on the planet
>people who should be exterminated
>people committing crimes
internet is a white invention. please feel free to leave
Will you agree to stop using the national infrastructure that was built by stealing the labor of Black bodies, enabling the US to become an economic power and in turn enabling it to invest in things like the internet? Just to keep things fair?
sure thing. i dont live in a southern state so nothing here was built by black labor
Kind of naive
i'm upholding my end of the bargain. please leave
You're a fool if you think Northern states didn't benefit from the labor of Black bodies.
the northern states fought a war to free the blacks in the south, losing over 350,000 men in the process.
you're welcome. now please go.
And the proceeded to benefit and profit from racial bigotry and racially exploitative labor.
honestly you are seeming pretty bigoted and racist. please stop judging people based on the color of their skin.
"I have a dream that my four little children will one day live in a nation where they will not be judged by the color of their skin but by the content of their character."
Nobody profits from blacks, the average black in the USA has a net financial impact of -$750,000 over the course of his life (taxes paid vs welfare/benefits claimed).
They are a race of underachievers. White and Asian hands built your world
You're dumb as fuck
your political safe space is this way
As a pajeet I don’t give a shit about either side here, but you should know that you’re a moron for giving blatant bait so many (you)s.
>As a pajeet
No, I'd say he's (and I'm) moron for giving blatant street shitters any (you)s.
All of Africa benefits from moron labor. While lot of good that's done them.
So we're just going to pretend that Africa hasn't been interfered with and exploited for centuries now?
What about all those centuries without any major foreign exploitation? The world didn't begin in 1492.
The point is that the model is decent enough to generate someone holding a sign and that prompt generate garbage every time unless OpenAI's system appends extra words to it, which end up on the sign
for fucks sake. I hate blacks (they are subhuman) and morons enablers
Turing Police was never necessary.
ANIME ASS AND TITTIES
technology gets me so hard
I'll be honest, other than in terms of resolution, Craiyon is better than DALL-E 2. Like the cherry-picked DALL-E 2 generations originally shown off were wild and all, but that is no where near the norm, and being so hilariously neutered as it is you just can't do anything funny or cool with it, only soulless corporate approved snore fests.
looking good tho
wow, that last one even supports the Most Current Thing™
Only feds get the model that lets them produce incriminating images of anyone at a whim.
You WILL let windows use AI to analyze your image and text files to make sure you're not producing harmful content.
genies never go back into their bottles
it is possible, so it will be made in due time
this is merely an attempt to delay the inevitable, a world where everything that you cannot directly see in real life is simply false
yeah, just like GPT3: there are hundreds of alternatives now, right?
yes, and it's only been two years since it was first unveiled
a large amount of news stories are directly written and edited by bots
in 10 years, it will be stupid to not immediately distrust anything you didn't see in person
we know what roman cement is, we only don't make it because it's expensive and impractical for buildings that aren't supposed to last more than 3 decades
and we have something more devastating than greek fire ever was: napalm
Nice deflect lol
i don't see how it's a deflect
roman concrete uses specific types of ash that makes it expensive compared to portland concrete
and i would argue that napalm and all the other modern incendiary devices are the same, if not better, than greek fire
maybe not the exact same substance, but the same effects if not worse
>in 0 years, it will be stupid to not immediately distrust anything you didn't see in person
GPT-J, BLOOM, yaLM-100b.
pretty big ones already.
Genies never go back in the bottle aside from roman cement and greek fire, right?
Naive to think you'll be able to trust your eyes once Google replaces them.
>Sorry, our automated fact checking systems have detected misinformation in your optical stream
>Deceptive content has been scrubbed. And a warning has been issued, have a nice day
This is the beginning of the end. The gayrons are building a moat to keep out the plebs.
They want to be the ultimate gatekeepers of the tech that will soon replace all creative workers.
All media will be produced by AI, custom-tailored for, but only in pre-approved forms of expression.
And it will all follow American puritan values, because Americucks have never been capable of not bothering other nations.
p3dn.24xlarge instances have 8x Tesla V100 GPUs with a total of 256GB GPU RAM, 96 vCPUs, 768GB RAM, 100 Gbps network bandwidth, 19 Gbps EBS bandwith.
They're only $31.218 per hour for on-demand pricing but if you reserve it for 1 year you get it at $18.30 per hour, 3 years is $9.64/hr.
>Microsoft no longer ships MSPaint with windows, due to "risks of misuse"
Listen guys, i got it: the responsible of all this is APPLE
just think about it: apple is too shit at AI, unlike its competitors -Google, Facebook, Microsoft- but is very good at marketing, since they're like a decade behind in AI, their only alternative is to cripple another companies while they get there: all this bullshit about racist, misogynistic, nazi AI is the usual guerrilla marketing | propaganda
Pretty much the same thing that happened with personalized ads: since apple can't do it, they're crippling Google and Facebook with their hypocritical ""privacy"" bullshit, while they give all their client's data to the american | chinese government
I think I might be in danger now that I know too much
i like the cut of your jib
oh more castration, thanks for making everything as unfun as possible
Not so open after all
One-trick pony that's totally useless apart from a couple hour amusement. Cannot be used in a professional context due to (a) obvious machine generated artefacts in every single image, (b) total lack of control - you want that same from sitting on a laptop from a slightly different angle? LOLno here's a completely different frog on a completely different laptop with weird stretching and twisting of the entire thing, and (c) the company that owns it is sitting on top of every request and might just flat out reject your attempts to get work done if some woke green-haired scold happens to take offence.
The only thing they have going for them at this point is the controversy. Otherwise the software itself is worth literally nothing.
>(b) total lack of control - you want that same from sitting on a laptop from a slightly different angle?
wrong, dall-e 2 supports inpainting
From now on, it will be an arms race trying to censor everything AI
In a couple centuries people will laugh at the futility of these early efforts
Or in a couple centuries people will live enslaved by corporations that shape their entire worldview through personalized generated material indistinguishable from reality. Facebook is already getting fairly close even with relatively minimal auto-generated propaganda.
if facebook were so good a propaganda, both normies and msm would love zuck and his company, but that's pretty much the opposite of the reality
if you wanna guess what corporation excels at propaganda, look for those that normies idolize, and media never dares to criticize
Facebook has 3 billion monthly active users. WhatsApp has 2 billion users, Instagram at least a billion. Their goal is to have everyone use their services, gathering information about everyone and controlling everyone's information flow. They don't care that people complain about Facebook, they'll just switch to another Facebook-owned platform. Nothing matters as long as the cattle keep letting themselves get manipulated by Facebook.
how can facebook manipulate anybody if normies and media HATE them with passion and distrust them basselessly? it's stupid to assume that some corporations are ""evil"" and other ones are ""good"" like if real life was a retarded hollywood movie
what are you saying is precisely what the official propaganda wants you to believe: that's how leaders alienated young impressionable males to get murdered defending their personal interests through all history
distrust the institutions that had been labeled as the good guys, not the ones demonized by the fucking media
I never claimed any corporations to be good; you are the one who is suddenly painting this as a black-and-white situation with Facebook apparently being good because certain news outlets are evil. Isn't it much more reasonable to assume they are all bastards out to make the most money out of us, and that their interests generally don't align with ours?
Judging by your image and implications, you seem to be part of the current US cultural conflict in which Facebook is playing both sides. Your anger and frustration leads to increased usage of their services, so they will keep feeding you information to make you rage at lefties. Meanwhile, they do the exact same to lefties and trigger them to believe anything associated with the right is evil. Snap out of it. Just because one organization is evil, does not necessarily mean anything that opposes said organization is good.
(You) started singling out FB for no reason. could as well have said "media", or "tech corporations", but that wouldn't fit your agenda, would it?
Facebook is just the current scapegoat, and ""privacy"" is the current meaningless buzzword in official propaganda aimed at the mindless masses. If you want to make a point, it would be smart to avoid the official narrative, for starters, otherwise I'll assume you're another imbecile regurgitating state propaganda while he believes he's a rebel fighting the system
Tell me, which other corporation has 3 billion active (!) users and on top of that custom-tailors their information stream to each individual user? For reference, even Microsoft Windows only has about 1.5 billion users, and it's not even a product designed to control the information stream. Facebook is by far the most influential company on the planet right now. But don't worry, Facebook is definitely your friend because it supports your side in an engineered two-party conflict in your country of 330 million people.
Consider this: Facebook has significant influence on almost ten times as many people as there are citizens in the USA.
>X corporation GOOD
>Y corporation BAD
stop it. real life is not a baby's fairytale. all them want money and|or power and have complex and even contradictory interests.
by oversimplifying shit ad absurdum, you're following your basic irrational instincts, most probably manipulated by propaganda that wants you to believe there are there are "good" corporations which are your friends, and becoming an useful idiot in general
You can’t stop what’s coming
anyone got a link to that one?
We make libre ai to generate an accurate representation of what floyd was
>BOT complains about muh censorship
>halfway through the thread and posters are already defending CP generation
You couldn't be more transparent if you tried
(you)s are eternal, my low effort friend
and also who the fuck cares what a bunch of weirdos do? am I supposed to feel angry about children that don't exist or what?
name one good reason it shouldn't be allowed
spoiler: you can't because there isn't any logical or objectively true reason.
This is absolute fucking trash and they expect me to pay for this shit too?
stop asking too many question goy
I literally can't do anything on it
I can't even upload my own models for variations because the female form triggers their content warning
HOW THE FUCK DO YOU THINK ARTISTS DRAW RETARDS? YOU START OFF WITH A BASE MODEL
WHO IS THIS EVEN FOR?
LITEARLLY FUCKING CREATED TO GET UPVOTES ON REDDIT TO SHOW OFF
>WOW TECHNOLOGY GUYS
No, this is like actual unusable trash
Fuck any retard who claimed that
>OH YOU COULD DRAW YOUR OWN MANGA USING ONLY DALL-E PROMPTS
Not only can it not take any models to import, even trying to do variations on the images you already have, it doesn't even do anything to change them
Not to mention the prompts barely ever have a good style
Can't even do fucking black and white basic backgrounds
Like the fucking last one is just the same image but flipped
Not really, once it's out, it's out.
There are weaker models out there for free. It's just a matter of time before the open stuff catches up with those models.
I mean, there is a lot of money in generating dirty stuff, let alone the political tool it can be to use it against rivals.
>OpenAI this week deployed a technique that forces DALL-E 2 to reflect "diversity"
And what a great fucking technique it is, just appending black person or women to people's prompts.
Literal actual forced diversity.
They also make it the ugliest moron that they can
that's racist because she has pancake tits like crackheads do and there's a stereotypical and categorically incorrect association between black people and crack usage, I'm reporting this to openai and the CP4AI (Colored People 4 AI) twitter groups.
Is black skin, blue eyes even a possible combo?
if you get a really deep tan
yes, Vanessa Williams exists
Plenty of LA ABOMINATIOS running around America that look like that
It's not the end, it's the part where others start looking to do the same thing without all the retardation.
It's just a matter of time.
also lol if you ever thought they would give people access to their toys without massive amounts of fuckery due to "concerns"
You had your one chance and it was direct access to GPT-3 through AI Dungeon before OpenAI stabbed them in the back.
>In April 2021, AI Dungeon implemented a new algorithm for content moderation to combat instances of text-based child pornography created by users. The moderation process involved a human moderator reading through private stories.
>text-based child pornography
That's not even half of it.
Their user's data mysteriously ended up on some task website where they had people manually tag generated stories, apparently without AI Dungeon dev's consent or knowledge.
Shortly afterwards they lost their special access deal to GPT-3, lost most of their userbase and had to downsize to a shitty small model that was nowhere near as good.
>text-based child pornography
2 decades ago people would laugh at the idea, but nowadays i'm supposed to take this shit seriously
The Rika 'I would never' copypasta is now child porn. The world has definitely gone crazy.
why the cotton picking fuck isn't there an AGPLv3 AI project?
You can download and run GPT-J, GPT-BOT (torrent), OPT, and bloom on your computer for free. It just takes a lot of memory and gpu memory to load. For image models there are also models available, I hear the Russian dall-e is good.
that image to video example was really good
that's not how actual rain looks like
no the one on their website was waves, although that rain one is also really good so I don't know what you're on about.
i like how the boat gets wrapped at the end
While it's probably not as technically impressive as the video thing (which, while impressive, doesn't look very good in practice), I'm most interested by the outpainting. I think this has great potential for practical use.
>Photoshop's autofill filter from 10 years ago
poorn M$ get anything right
dall-e 2 already can do inpainting.
outpainting is a poor man's inpainting.
Cant we just use our 3000 series rtx cards for this shit instead of giving these garden gnomes even more attention?
DALL-E mini/craiyon already runs fine on an rtx3080.
craiyon is for TPUs not GPUs.
craiyon's parameters are too big for a consumer GPU's RAM even for just inference.