Meta's new AI image generator was trained on 1.1B Instagram and FB photos

https://arstechnica.com/information-technology/2023/12/metas-new-ai-image-generator-was-trained-on-1-1-billion-instagram-and-facebook-photos/

by my12parsecs • 3 years ago

338 219 3 years ago

arstechnica.com

TheCoreh2• 3 years ago

“Not available in your location

Imagine with Meta Al isn't available in your location yet. You can learn more about Al at Meta in the meantime and try again soon.”

I wonder why it's region-locked?

philipov2• 3 years ago

Which region is locked? That might give a clue.

RowanH• 3 years ago

New Zealand is locked out. (Normally we get first dibs on things being a small test market)

avallach• 3 years ago

I got the same from the Netherlands

K5EiS• 3 years ago

Norway is blocked, so probably some GDPR issues.

fallensatan• 3 years ago

Canada seems to be locked out as well.

TheCoreh2• 3 years ago

Brazil. So it's unlikely to be GDPR-related, unless they're also treating our LGPD as a special case.

triggerhappy77• 3 years ago

Me to Meta: Please let me in(dia), we are locked out too

lxgr3• 3 years ago

Meta's AI stickers also only seem to be available in the US for now (or at least not in WhatsApp in the EU).

mvdtnz2• 3 years ago

AI stickers are in my region (not USA) but imagine is not.

xnx2• 3 years ago

Link to tool: https://imagine.meta.com/

One of many AI updates from Meta yesterday: https://about.fb.com/news/2023/12/meta-ai-updates/#:~:text=E...

floathub• 3 years ago

Note that you need to "Log On" to Facebook/Meta/WhateverTheyCallThemselvesNow to try it. Kind of curious, but not curious enough to create yet another burner Facebook account.

[edit: still learning to spell]

theonlybutlet• 3 years ago

Thanks I should've read your post before opening the link and promptly having to close it.

vsnf• 3 years ago

How are you guys even making burner facebook accounts? Every time I try (though granted, I haven't tried since 2016), I get stymied by a hard phone number requirement.

misja111• 3 years ago

"Not available in your location yet" (Switzerland)

JumpCrisscross• 3 years ago

"Not available in your location yet" (Switzerland)

Have the GDPR questions around data provenance been resolved? I thought EU/EEA is currently off limits for publicly- or user-data-trained AI.

mvdtnz2• 3 years ago

Not available in my region (New Zealand), darn.

tikkun• 3 years ago

I tried it now.

My experience:

Took 4 minutes to log in and do one generation. (Login to FB, then it took me through a process to merge accounts with Meta, which didn't sound good, so I restarted with 'sign in via email' which ended up doing the same thing anyway, I think. Then I was logged in, did the generation.)

My at a glance is that it's:

For image quality

1. Midjourney

2. Dall e 3

3. SDXL and this

For overall ease of use and convenience

1. Dall e 3

2. Midjourney

Of course, this is all biased personal opinion, and YMMV.

whywhywhywhy3• 3 years ago

Depends what you want really, Midjourney and Dall-E 3 have specific looks to them which kind of look cheap/tacky now its everywhere.

SDXL is reconfigurable and completely flexible so really its the only tool in the game for pure creativity.

Centigonal• 3 years ago

I clicked the link to try generating an image.

Several beautiful modal dialogs later, my Meta account has been linked to my Facebook account, my Oculus profile is now my Horizon profile, and I have chosen a publicly viewable(!?) display name for my Horizon profile (a profile for a game I have never played and never intend to play). I have been informed that my Oculus friends are now Horizon followers, given the chance to select "how social [I] want to be," asked to invite my Facebook friends to join Horizon -- and I still haven't generated an image. I almost feel like this image generator is somehow a long con to get people to update their Meta accounts.

I want to find the group of product managers responsible for this user journey and just... shake them out of it! The design you shipped is really dumb! None of this makes sense outside of Meta! There's a whole world out here! Nobody cares about Horizon Worlds!

Invictus0• 3 years ago

This is also called "shipping the org chart".

esafak• 3 years ago

It's almost beginning to resemble the UX of an enterprise application!

dzink• 3 years ago

Bet Horizon worlds will get great engagement metrics for the next earnings release on the topic.

cowboyscott• 3 years ago

Is training with user-generated content a way to launder copyrighted images? That is, if I upload an image of Ironman or whatever to my Facebook or Instagram page as a public post and Meta trains their model on that data, is there wording in my user agreement that says that I declare that I own the content, which then gives Meta plausible deniability when it comes to training with copyrighted material?

(apologies for the run-on sentence - it is early still)

onlyrealcuzzo2• 3 years ago

Is training with user-generated content a way to launder copyrighted images?

Doubt it. If you upload child porn to Instagram and they distribute it - it's still an Instagram problem, AFAIK.

dragonwriter3• 3 years ago

Child porn is not a copyright issue, so the DMCA safe harbor for UGC doesn't apply, and its criminal, so the Section 230 safe harbor doesn't apply, so its very much not an applicable example as to whether use of UGC in other contexts is a way of leveraging safe harbor protections for content, whether for copyright or more generally.

glimshe3• 3 years ago

I think Meta is already assuming that there will be no liability for training with copyrighted material. I find it very unlikely that image owners will win the AI training battle.

lxgr3• 3 years ago

I'd be extremely surprised if the "Mickey Mouse standing on the moon" example image was a legitimate way to "launder copyright".

The interesting question is just who will be liable for the copyright violation: The party that hosts the AI service? The party that trained it on copyrighted images? The user entering a prompt? The (possibly different) user publishing the resulting image?

codingdave• 3 years ago

It is in big bold letters right in instagram's terms of service: "We do not claim ownership of your content, but you grant us a license to use it."

This isn't about copyright, it is about the fact that most people don't realize that by posting photos, they are licensing those photos.

__loam4• 3 years ago

Ultra shitty corporate interests win again...

sp332• 3 years ago

They don't own the copyright, but they do have a "non-exclusive, royalty-free, transferable, sub-licensable, worldwide license to host, use, distribute, modify, run, copy, publicly perform or display, translate, and create derivative works". https://www.facebook.com/help/instagram/478745558852511

bee_rider• 3 years ago

They user might upload something that they don’t have rights to.

Technically the user is the one misbehaving, but we, Facebook, and any reasonable court know that users are doing that.

Bjartr2• 3 years ago

If they didn't have that (or something similar) they couldn't serve the image to other users. Well, they could, but without something like that someone will sue them for showing a picture they uploaded to someone they didn't want to see it (or any number of other gotchas).

They store the image or video (host/copy), distribute it over their network and to users (use/run), they resize it and change the image format (modify/translate), their site then shows it to the user (display/derivative work), and they can't control the setting in which a user might choose to pull up an image they have access to (the "publically" caveat)

It sounds like a lot, but AFAIK that's what that clause covers and why it's necessary for any site like them.

costcofries• 3 years ago

This is why you don't also download the music from stories when you download stories, no such agreement with Spotify.

ezoe2• 3 years ago

You're just stating an agreement between Meta and cowboyscott. The copyright holder of Ironman image never agreed it.

The problem here is cowboyscott doesn't own a copyright of Ironman image. But his uploading of image may match the condition of fair use of US or similar copyright exemption rule in their country's copyright law. It effectively works as copyright laundering.

laylower• 3 years ago

You forgot the "in perpetuity" /s

sosodev• 3 years ago

It seems like this is still very much a legal gray area. If it's concretely decided in court that generative AI cannot produce copyrighted work then I assume it makes no difference what the source of the copyrighted training material was.

KaiserPro4• 3 years ago

When an image us uploaded is it re-licensed:

  > When you share, post, or upload content that is covered by intellectual property rights (like photos or videos) on or in connection with our Service, you hereby grant to us a non-exclusive, royalty-free, transferable, sub-licensable, worldwide license to host, use, distribute, modify, run, copy, publicly perform or display, translate, and create derivative works of your content (consistent with your privacy and application settings). This license will end when your content is deleted from our systems. You can delete content individually or all at once by deleting your account.

ROFISH• 3 years ago

So if you delete your image the entire trained data set is invalid because they no longer have license to the copyright?

carstenhag2• 3 years ago

Yeah, derative works in this case afaik was always be meant as "we can generate thumbnails etc" and not "we will train our AI with it". I am pretty sure this is illegal in many countries...

raincole• 3 years ago

At this point all big players assume it's okay to train on copyrighted materials.

If you can[0］crawl materials from other sites, why can't you crawl from your own site?

[0]: "can" in quotes

carstenhag2• 3 years ago

Because your users have agreed to terms of service that don't mention analyzing the images to train an AI model.

PeterisP2• 3 years ago

It's not a legal way to "launder" copyrighted images, because for things where copyright law grants exclusive rights to the authors, they need the author's permission, and having permission from someone and plausible deniability is not a defense against copyright violation - the only thing that it can change is when damages are assessed, then successfully arguing that it's not intentional can ensure that they have to pay ordinary damages, not punitive triple amount.

However, as others note, all the actions of the major IT companies indicate that their legal departments feel safe in assuming that training a ML model is not a derivative work of the training data, they are willing to defend that stance in court, and expect to win.

Like, if their lawyers wouldn't be sure, they'd definitely advise the management not to do it (explicitly, in writing, to cover their arses), and if executives want to take on large risks despite such legal warning, they'd do that only after getting confirmation from board and shareholders (explicitly, in writing, to avoid major personal liability), and for publicly traded companies the shareholders equals the public, so they'd all be writing about these legal risks in all caps in every public company report to shareholders.

rpdillon• 3 years ago

However, as others note, all the actions of the major IT companies indicate that their legal departments feel safe in assuming that training a ML model is not a derivative work of the training data, they are willing to defend that stance in court, and expect to win.

I think the move will be to argue fair use, declaring the derivative work to be transformative, and possibly to point out that only a small amount (1%-3%) of the original data is retained.

caesil• 3 years ago

Training on copyrighted content isn't a copyright violation. Sarah Silverman is currently learning that the hard way.

ezoe2• 3 years ago

Another method of copyright laundering is doing ML learning in a country where it doesn't protected under copyright law.

Personally, I'm on a side of using copyrighted data for machine learning input source doesn't violate copyright. Statistically, learned model for generative Ai doesn't retain even 1 bit of input. It's hard to say NN model data infringe any copyright of the input source. The copyright is applied to the expression, not the process. If the generative AI produces an image that's clearly a copy of a specific Ironman image which existed before the image generation, that's copyright infringement.

WanderPanda• 3 years ago

I agree with you but I think the argument is flawed. If you think about it like this h265 also just steals 10% (or whatever the compression ratio is) of an artifact

abrookewood• 3 years ago

"learned model for generative Ai doesn't retain even 1 bit of input". If that was true, it shouldn't be possible to trick the models into regurgitating their source material, but cleary that is possible [0].

[0] https://stackdiary.com/chatgpts-training-data-can-be-exposed...

AlotOfReading• 3 years ago

Copyright doesn't require a single bit of input to be shared. You can't avoid copyright by using a paintbrush for example, you're simply creating a derived work. You might still be in violation even if you create an entirely new context around the copied elements or substitute for the original in the market, as was the case with Warhol vs. Goldsmith.

Obviously not every generative output is a copyright violation, but it seems equally clear that there are outputs that would be if they were produced by humans.

lawlessone• 3 years ago

learned model for generative Ai doesn't retain even 1 bit of input.

It does. The data is just obfuscated.

zeruch3• 3 years ago

"Is training with user-generated content a way to launder copyrighted images?" Pretty much.

incrudible3• 3 years ago

You are very very unlikely to stumble upon something resembling a training image closely enough for copyright to take effect, and in any event this is not the purpose of these systems. You may be running into trademarked content, but in that case you can not speak of laundry, because you can not use a trademark even if the image is AI generated.

FpUser2• 3 years ago

It is not any different than actual live artist learning from works of others

SirMaster• 3 years ago

What about all the photos of people at Disney taking pictures of themselves standing next to Mickey Mouse etc.

I don’t think there’s a question that people are allowed to upload photos like that.

JohnFen8• 3 years ago

I don’t think there’s a question that people are allowed to upload photos like that.

Technically, that's a copyright violation. Disney just opts not to enforce their rights for that sort of use.

Similarly, you technically can't take and post pictures of statues, paintings, some buildings, etc., and some rightsholders do enforce their copyright when people do those things.

lumost• 3 years ago

This is almost certainly going to be used to generate actual pictures of real people in the nude etc.

0cf8612b2e1e• 3 years ago

Fake celebrity nudes pre-date the internet.

rchaud2• 3 years ago

Barriers to entry were a lot higher, and distribution capacity was a lot lower. Surely you can see how the change in that combination could make for a significantly different reality now.

btbuildem• 3 years ago

I really really doubt that. If anything, it'll be nerfed into complete uselessness.

delecti• 3 years ago

Doesn't seem to be possible. I tried a variety of real people (Tom Hanks, George Bush, George Washington) and each time got the error "This image can't be generated. Please try something else." It did work with some fictional characters though, namely Santa and Mickey Mouse. I'd rather not try asking for nudes while at work, so I can't attest to that part either way. Though "Sherlock Holmes dancing" looked pretty clearly like Benedict Cumberbatch (though the face was pretty mangled looking).

wongarsu2• 3 years ago

That has been a thing since 2019's DeepNude, and the world hasn't ended. If anything it has been relegated to obscurity.

KaiserPro4• 3 years ago

Its not obscure. There are a bunch of paid apps that allow you to "virtually undress" any image you upload.

Which is already causing pain for a bunch of people.

acdha3• 3 years ago

Is it obscure or just not in the news you follow? There have been many reports about significant impacts on school students:

https://www.technologyreview.com/2023/12/01/1084164/deepfake...

https://www.cbsnews.com/news/deepfake-nude-images-teen-girls...

https://www.theverge.com/2023/6/8/23753605/ai-deepfake-sexto...

It’s also showing up in elections:

https://www.telegraph.co.uk/news/2023/05/14/turkey-deepfake-...

https://www.wired.com/story/deepfakes-cheapfakes-and-twitter...

GeoAtreides• 3 years ago

If anything it has been relegated to obscurity.

oh man, if /b/ could read this they would be very upset right now

KaiserPro4• 3 years ago

For that to work, you need to have a dataset of nudes to start with.

Given that instagram is pretty anti nudity (well women's nipples at least) I'd be surprised if there is enough data to work properly.

Its not impossible, but I'd be surprised.

soultrees• 3 years ago

At this point who cares honestly. The more ‘fake’ generated nudes out there, means it’s just not going to be a novelty. And if everyone has the ability to generate an image of everyone naked, the value for ‘real’ nudes will go high but it will also be good cover for people who get their nudes leaked.

merrywhether2• 3 years ago

I’ve wondered about a similar thing. If there was something automatically constantly generating nudes of everyone, surely the noise would desensitize people to the signal.

dopa42365• 3 years ago

How's that any different from the gazillions of more or less good "how would you look like older/younger", "how would your kids look like", "how would you look like as barbie" and what not tools? One click to generate a thousand waifus. It's not real, who cares.

nextworddev• 3 years ago

I tried this and was floored how good this was

nothrowaways• 3 years ago

The title is misleading. It uses publicly available photos, which means it uses the same image as other AI models like GPT, midjiurney ...

holoduke2• 3 years ago

Who is gonna use these heavily moderated generators?. You cant even generate a nipple or a famous person. There is almost no control or finetuning. There are zillion of checkpoints, loras, controlnets and ipadapters out there to get almost anything with sd. No filters. You can literally generate whatever you like.

miguelazo• 3 years ago

Wow, another reason to delete my accounts.

leptons• 3 years ago

If nothing else they've done so far hasn't convinced you to delete your accounts, then why would this? They've done worse before.

WendyTheWillow• 3 years ago

Because it’s trained on “real” people, will it be easier to generate ugly people? I have a hard time convincing DALL-E to give me ugly DnD character portraits.

wobbly_bush• 3 years ago

Aren't Insta images heavily edited?

rchaud2• 3 years ago

Yes, with filters supplied by Instagram, so they would still have the original camera images.

doctorpangloss• 3 years ago

Because it’s trained on “real” people, will it be easier to generate ugly people?

In the literature, testing concepts in image generation is asking human graders "which image do you prefer more for this caption?," so the answer is probably no. You could speculate on all the approaches that would help this system learn the concept "ugly," and they would probably work, but it would be hard to measure.

PUSH_AX2• 3 years ago

In order for a model to understand what ugly is, someone or something has to tag training data as “ugly”, I find this to be a complete can of worms

Jerrrry• 3 years ago

In order for a model to understand what ugly is, someone or something has to tag training data as “ugly”,

that is a very dated (2008) concept.

the model "understands" that 50% of people are below/above median.

consequently, those that are not "OMG girl ur BEAUTIFUL"-tagged are horse-faced.

It understands that the girl with the profile picture with 200 likes and 2k friends is better looking than the girl with 4 likes and 500 friends.

Guillaume86• 3 years ago

Put beautiful/pretty in the negative prompt, should get a similar result without the need for tagging ugly in the training set.

hbossy• 3 years ago

Try asking for asymmetry. The more images of faces you average, the better they look.

brucethemoose22• 3 years ago

It can handle complex prompts better than Stable Diffusion XL, but perhaps not as well as DALL-E 3

This is a interesting statement, as Stable Diffusion XL implementations vary from "worse than SD 1.5" to "Competitive with DALL-E 3."

sjfjsjdjwvwvc• 3 years ago

It depends what you want to gen and what prompting style you prefer. I have found SD 1.5/6 to be far more flexible than SDXL. SDXL seems more „neutered“ and biased towards a specific style (like dalle/midj); but this may change as people train more diverse checkpoints and loras for SDXL.

brucethemoose22• 3 years ago

See, this is totally my opposite experience. SDXL handles styles incredibly well... With style prompting.

Hence my point. SDXL implementations vary wildly. For reference I am using Fooocus.

al_be_back• 3 years ago

to me these innovations seem akin to Concept Cars in the Motor industry; there's some utility, until some executive takes it center-stage, and pisses-off most of the core users.

the biggest value in these networks is real User-generated content, you can't beat billions of real users capturing real content and sharing habitually.

even if wording in the Terms permit certain research/usage, you've got market and political climates to consider.

jafitc• 3 years ago

All I can say is it’s really fast

junto• 3 years ago

Before anyone tries it out from the EU, be warned that it will push to make a Meta account and merge any Facebook/ Instagram profiles together and once you’ve finally bitten that bullet, it will tell you that it isn’t available in your region.

kevincox• 3 years ago

Same in Canada

mrits• 3 years ago

It took me about 10 minutes to do what should have been a single click. They even wanted me to generate a new password even though I login with facebook. In US though and had some fun. It isn't as good as ChatGPT but very impressive.

CGamesPlay• 3 years ago

Strangely, it appears to decide this based on IP geolocation. My account is listed as US-based but the site does not work when using non-US VPN.

phatskat• 3 years ago

I wouldn’t be surprised if they have several layers of location checks and if any of them fail they bail. Typically with geolocation on projects I’ve done we will rely on the best available info - location permission, IP geolocation, or the user telling us where they are via form.

seydor• 3 years ago

So it's just faces?

dmazzoni• 3 years ago

If you ask it to generate an image of Taylor Swift, it refuses. But if you ask it to generate an image of a popular celebrity singer performing the song "Blank Space", it generates an image that looks exactly like Taylor Swift some fraction of the time.

a_wild_dandan• 3 years ago

I wonder if celebrity doppelgangers can't find modeling work. Like, without EVER referencing your celebrity twin, how closely can your work implicitly approach Swifthood before your free expression gets violated? To dramatize for effect:

Can you act in films? Or model a company's products like a guitar/microphone? Or genuinely start a band? Can your credits/band name reference you, if your given name is coincidentally also "Taylor Swift"? Can Facebook AIs train on your Facebook images, and produce a "celebrity female singer" images (with/without a "Blank Space" reference)? What if your LLM's purpose is strictly "parody, caricature, and images whose likeness is purely coincidental"? Can generative AIs have intention? Let alone intention to break copyright?

The consequences are endless in both kind/degree when pretending that "likeness" is some unique fingerprint. Ditto for thought-policing what (artificial or human) neural networks can learn from without paying royalties or whatever. It's all absurd.

What's more, our society must face these issues. We can't dismiss them as all hyperbolic catastrophizing about slippery slopes. Our system is already subjective, inadequate, and incapable of sorting itself out. The situation becomes more dire each day. Given our trend of sacrificing public interest for private greed (e.g. Disney's hatchet job on copyright), I'm worried about our future.

RegW• 3 years ago

I wonder what other purposes FB have used those 1.1B+ publicly visible photos to train models for?

mr_toad4• 3 years ago

Anything that classifies and/or recommends images will likely be a deep learning model these days.

Havoc• 3 years ago

Meta is asking me to log in with my facebook account. Then after authenticating with my FB account meta says I don't have a meta account.

Is this all some sort of scam to get me to click accept on whatever godforsaken ToS comes with a meta account? If the FB account is good enough to freakin AUTHENTICATE me then just use that ffs.

__loam4• 3 years ago

This is extremely shitty to a lot of users.

andrewstuart• 3 years ago

And weirdly, every image it generates is sort of a combination of your grandma and an influencer on a beach on a tropical island.

astrange2• 3 years ago

Doesn't do "hard prompts" better than other systems I've tried. Looks pretty similar to them too.

eg: "horse riding an astronaut", "upside-down mini cooper", "kanji alphabet soup".

ClumsyPilot2• 3 years ago

Ooh, so that's what they are called! I tried to do "bicycle in a jar" and they could do it, but when I did "car in a jar" or "toy car in a jar" all of them failed.

astrange2• 3 years ago

I don't know if that's the right kind of term, it's just a list of prompts I've noticed usually don't work in image generation models - specifically they ignore what you said and just make an image with some of the words in it.

This looks like a SD1.5-like latent diffusion model though. The giveaway is that it can't spell.

prawn• 3 years ago

Sometimes you need to massage the prompt a bit to avoid it getting distracted. e.g., it took me a few tries to get a family having dinner inside a home aquarium. I'd specified it being a large feature wall aquarium, and it got hung up on showing the dining table in front of the feature wall, rather than literally inside the aquarium.

tiffanyh• 3 years ago

1.1B is tiny.

Given that FB & IG combined have ~0.5B photos uploaded daily, this effectively translates to training data from just a few days of user generated content.

https://www.brandwatch.com/blog/facebook-statistics/#:~:text....

https://www.zippia.com/advice/instagram-statistics/#:~:text=....

next_xibalba• 3 years ago

But is it tiny with respect to the volume of data required to create a good model and the compute costs associated with the training and operation of that model?

acchow• 3 years ago

They are using only the publicly available photos. Not the ones you share only with friends.

ssss11• 3 years ago

And once they notice no one gives a sh* they’ll throw 100b at it, then more…..

FpUser2• 3 years ago

Canada. It asked me to create Meta account only to tell me that it is "not available in your region".

Fuck you Meta and fuck you Zukerberg.

neilv2• 3 years ago

Interesting. Unlike some other popular image generation training, is there a chance that Meta technically got copyright permission for many/most of the images that were posted to its properties?

I'm thinking: When the user who uploaded the image was also the copyright holder, that might've been covered by an agreement that technically permitted this use by Meta.

(Copyright isn't the only legal issue, though. For example, a person in a photo that someone else uploaded doesn't necessarily lose right to their likeness being used for every purpose to which a generative AI service might be put.)

deegles• 3 years ago

They probably added a clause to their terms of service retroactively granting them permission to use your images for this purpose.

neilv2• 3 years ago

I was thinking Meta might be on more solid copyright ground here than most large generative AI models have.

(Not that legal ground is going to stop anyone, in a generative AI race worth trillions of dollars.)

neom• 3 years ago

Really struggles with fingers, probably worse than any AI image generator I've seen so far. Maybe there aren't a lot of finger-showing images on IG and FB!

andsoitis• 3 years ago

The images of ourselves have now been absorbed into an AI.

An intelligence that knows a shit ton about a very very large number of people.

squigglydonut• 3 years ago

Artists...leave.

miked85• 3 years ago

Yet another reason to steer far clear of anything Meta.

zoklet-enjoyer• 3 years ago

I'm not sure if they cut me off for generating too many images or because of the content of my images. Everything is now giving the response "This image can't be generated. Please try something else."

This only started after I put in the prompt manbearpig did 9/11. It was ok with some really weird stuff though

Meta's new AI image generator was trained on 1.1B Instagram and FB photos

Related Stories