Hacker News Reader

Comment by tomatowurst | Hacker News Reader

This DALL-E 2 is absolutely the wildest thing I've seen by far. If we can get it to render 2D images, imagine 2d films, fully textured 3d mesh & environments is almost a certainty in the future.

The implications of this is tremendous. We could be looking at crowd-directed AI generated consumable content.

Imagine that you could watch infinite variations of the ending in Sopranos season 6. Imagine that you could generate a production quality 3D animatable human mesh that can interact with the AI generated 3D environment.

Pair this with ultrasound based non-invasive haptic feedback, taste, smell with some type of 3D holographic projector. The possibilities are literally endless.

Holy moly!!!! We are on the cusp of a completely new way to generate and consume content. Copyright laws will have to be rewritten. Once this pandora box is open it can't be closed.

View on HN

> We are on the cusp of a completely new way to generate and consume content. Copyright laws will have to be rewritten. Once this pandora box is open it can't be closed.

The question is not if but why

We don't need more content, we need more curated and higher quality content.

Random by definition has a very low probability of producing quality.

Case in point: The walking dead.

It's already an infinite variation of the same topic but the good bits ended after season 2...

The holy grail is to generate content for you!

The future will combine these tool with data about you—aesthetic preferences, story arches you respond to well, your current mood—-and generate content on the fly. I assume there will still be people working on the skeleton or basic structure of such content, but the rest will be filled in by these tools.

I don’t buy it. There was a discussion on a post about Netflix about how their “long tail” of niche content is getting totally blown out of the water by the likes of Ted Lasso, Severance, and pretty much any of the Mouse’s tentpole properties. Hell Netflix’s most successful shows right now are all big tentpole productions by Shonda Rimes.

The problem with art and aesthetics (and many other things) (1) is that people often don’t know what they want until they see or hear it, and (2) there is a social element to it - people derive value from being part of a shared experience.

> The future will combine these tool with data about you

I'm not sure anyone has let you know yet but this future you're imagining was cancelled a long time ago.

We're going to start seeing more global food shortages this year, it won't impact you yet (except small price increases). But in a decade expect your quality of life to be noticeably worse than it is now.

The world you're imagining living in is a world powered by tremendous amounts of fossil fuel energy. But either a.) we run out of fuels, or b.) we don't and cook the planet.

The mass delusions we're living through right now when people start talking about the "future" is just a bit too much for me to let go any more. We need to stop pretending for all of our mental health.

> The future

But probably not in my lifetime.

> aesthetic preferences, story arches you respond to well, your current mood

that's not how content consumption works.

> will still be people working on the skeleton or basic structure of such content

That's not a real issue, we are still writing stories based on very few skeletons and very few archetypes, invented millennia ago (and maybe they are even older of the ones we discovered).

How does content consumption work then?

Would you not say there are clear characteristics, like your preferred genres, that matter quite a bit whether you will like a movie or not?

> How does content consumption work then?

simplified, there are two main branches:

- those who consume everything, kinda like maximizing the investment or "it makes me company"

- those who consume things that need to pass some kind filter

generated content would (maybe) reward the first group, but without some very good hit they will probably go back to downloading stuff, but not the second.

At least not yet and not in the foreseeable future.

There's a difference between computer aided content and computer generated.

Jurassic park wouldn't exist as it is without technology advancements, but the novel was already good.

Another example by Crichton is WestWorld. The original was written and directed by Crichton himself in 1973, the original movie is great, the second chapter is good, the tv-show is visually astonishing, but after 2 seasons, I stopped watching it. The material taken from the first two movies had ended and nothing really interesting have come out after from the writers (IMO).

Voyage au centre de la Terre (Journey to the Center of the Earth) is a great book from 1864 and I honestly prefer the movie from 1959 than the its latest 3D incarnation from 2008.

Not because older is better, but because the 1959 version is more centered around the story and it aged better.

Probably because Indiana Jones hadn't come out yet and the director didn't try to imitate it.

I’m pretty sure it will continue to be more profitable to just cater to the lowest(-ish) common denominator.

Technologies like this can lower the cost of production, opening the door to more competition essentially due to a larger talent pool. Quality can also go up as a result, not just quantity.

> Technologies like this can lower the cost of production

Seriously, is writing "infinite variations of the ending in Sopranos season 6" the limiting factor here?

The cost of production is already low, what's costly is distribution, for the distributor.

Make your movie and put it on YouTube it's already a thing.

> The cost of production is already low

The Sopranos cost 2 million dollars per episode. HBO will not be afford to produce an infinite number of those using their existing methods any time soon.

> The Sopranos cost 2 million dollars per episode

well, actors like to be paid...

you can make an infinite amount of breaking bad episodes without Brian Cranston, Aaron Paul, Bob Odenkirk, Giancarlo Esposito, and Jonathan Banks, but I guess people won't be so keen to pay to watch them.

But even if you use digital avatars, you have to pay them the same amount of money to use their image, because that's what people want to see, there's no way around it.

Besides, the Sopranos is a great show because they spend 2 million on each episode.

I'm sure you can recreate the same quality with a couple hundred dollars, but I'm also quite sure most of us regular humans couldn't.

Sure, but how is this different from all the blogs, vlogs, social media, and the Internet in general (with an assist from smartphones, digital cameras, game engines, camera drones, and real-time ray tracing GPUs) allowing humanity to create an unprecedented global flood of content now? Or the problem of finding something satisfying to watch out literally tens of thousands of shows on Netflix, Amazon Prime, Hulu, Disney+, HBO Max, AppleTV+, etc?

That’s why we have critics, reviews, review aggregators, search engines, recommendation engines, community discussion boards…

Clearly, the answer to an AI generated problem (pretty much anyone being able to content previously requiring massive budgets) is more AI! We can evolve content critics to help people navigate a world of practically unlimited content to find stuff they enjoy.

Also, consider all the potentially great films that never got made because even the best filmmakers can only get a fraction of their ideas actually produced.

> Also, consider all the potentially great films that never got made because even the best filmmakers can only get a fraction of their ideas actually produced.

and that's a good thing.

Consider how many bad movies were not made thanks to this!

IMO, content creation should be harder, not easier.

If it becomes too easy, the quality goes toward hobbyist level, which is not what I want to pay for.

Also, bad content was a thing on its own, take for example Troma films, they are entertainingly bad, not just repetitive and dull bad.

> Or the problem of finding something satisfying to watch out literally tens of thousands of shows on Netflix, Amazon Prime, Hulu, Disney+, HBO Max, AppleTV+, etc?

the problem is exactly the mass amount of content created to lure people in.

But do I really want to watch another iteration of the superhero trope?

No, I don't.

I wanna watch something that I've never seen before.

Imagine a book written to mimic the style of my favourite writers on my favourite topic: Sci-fi.

On paper it sounds like a dream, but in reality it would be a terrible, confusing, incoherent and ultimately a trying-too-hard-to-match-my-taste product.

Like when Data reads his poems on TNG and when asked why people seemed distracted, Laforge answers "Well, your poems were clever, Data, and your Haiku was clever, and your sonnet was clever. But did it evoke an emotional response? To be honest, no"

Of course Data being Data, he knows he's clever and he knows he's an amateur poet and wants to hear the truth about his works.

But do we want to pay for content generated by the machines so that they can learn to imitate us better? [1]

I sincerely don't want to.

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

EDIT: with [1] I mean that I don't want to pay streaming content publishers to improve their algorithm so that they can mass produce more dull content for people to consume (and pay)

If they really want to improve things, they should pay people to make less shows and/or movies but better.

Shows like "Another life" (6% on Rotten Tomatoes) would not exist if they were shot on real sets that take time, money, permits and a lot people to recreate.

Pretty sure you're wrong. Look at what's happening in music right now. It used to be that if you wanted music that didn't sound like total garbage it was going to be a mainstream production, because all the tools around professional music production were expensive. Now hobbyists have tools comparable to what the pros use, indie music tends to be well produced, and we have a much greater proliferation of fringe music than in the past. Considering how shit mainstream music is from an artistic perspective today, and I have to say I'm glad.

By lowering the cost to create professional art, it reduces the barriers to producing unique fringe art. Without a need to recoup a massive investment, people can take experiments, and express themselves.

Uh, you’re not actually required to watch something just because it exists. Some subset of people not liking content isn’t, in and of itself, reason to stifle peoples ability to create said content.

The job of a literature professor is not to protect Literature by ensuring that inadequate students are never allowed to write.

If some combination of AI and community driven filtering can prevent sub-par works from entering your queue, you are no worse off.

On the other hand, some scripts like Severance, which languished for years on the best-scripts-never-made list, would likely never have been (finally) made but for the vast expansion of content creation due to the rise of streaming. In fact, if only a dozen movies were allowed to be made each year, they would probably all be Marvel movies or something with similar mass appeal. Most of the truly innovative work appears at the margins, hence expanding the margins is more likely to yield groundbreaking art (along with a ton of crap).

> Uh, you’re not actually required to watch something just because it exists

believe it or not, if the money goes all in one direction, you're at one point between a rock and a hard place when it's about choice.

> The job of a literature professor is not to protect Literature by ensuring that inadequate students are never allowed to write.

But the job of a film production company is not to produce as much content as possible, but taking the financial risk to produce something that will make an impression.

> IMO, content creation should be harder, not easier.

> If it becomes too easy, the quality goes toward hobbyist level, which is not what I want to pay for.

IMO coding should be harder, not easier.

If it becomes too easy, the quality goes toward hobbyist level, which is not what I want to pay for.

code is already hobbyist level.

it rarely escaped the mid-low quality level since it became a mass produced product.

By the same logic, we can say the same things for no-code tools, IDE, spellchecker, etc.

ML models used in this way will struggle with scene to scene consistency...

For example, when you generate someone riding a horse, and later want the same character to be at home drinking a coffee. It'll be hard to get the model to render the same face, mannerisms and style for the same person in both cases.

I think that alone will take years to properly solve.

Consistency in stills seems to be solvable with the current generation models. You can now get an AI to draw Brad Pitt on a horse wearing a Stetson and Brad Pitt in Central Perk drinking a coffee with the right source data, or move a subject into a different setting by tweaking text descriptions enough, and it seems they often get details like shadows and background borders right.

I think it'll struggle enormously with subtle details of mannerism in film and coherently fleshing out stuff where it hasn't got a lot of data in its database (getting Brad Pitt or New York from multiple angles is a lot easier than getting a consistent thisactordoesnotexist or novel fantasy setting). But I think the bigger barrier to adoption outside CGI studios (for producing content above meme-level quality) is that actually describing enough details to get everything to cohere even in static images will be beyond the average person.

That’s because Brad Pitt is a named celebrity with lots of photos from many angles that has been included by name in the training data. It can get Brad Pitt like content by name. If I say “5 foot 6 inch tall older Italian woman <insert rest of scene description here>” I’m going to get an absolutely randomised person as the focus each time, the mode has no memory and this has been the BIGGEST limitation in all these transformer based architectures. There’s no memory of each output, they are all discrete and as much as you can kinda fudge it with GPT-3 by including part of the previous paragraph or sentences in your prompt to try and keep it on track, it can, does, and inevitably will drift horribly off track unless you manually coach it over and over micromanaging it’s output to keep it useful and sensible.

With GPT it’s each text prompt, with DALL-E it’s each text or image prompt, and with a hypothetical video version it would be the same limit, each prompt has no memory of the previous and so it can’t implicitly reuse anything created by a previous prompt. You have to hack that in sound the core transformer architecture to make a facsimile of memory by things like pasting in extra data or teaching DALL-E 2 to edit a picture you upload and then uploading the picture you just generated using it, video would be no different and consequently this approach will not scale efficiency beyond individual items of content like an arbitrary sized block of text, a single photo or small group of photo edits, or a single “shot” of video content containing a scene or two worth of contiguous narrative flow.

To create specific people and spatiotemporal consistent content, we would use conditional image generation where we condition on both the text and some seed vector, and proceed autoregressively. If we wanted a random person, we would use a random seed vector. If we wanted a specific person, we would use 3D GAN Inversion to find the seed vector which would give us the person in a provided query image. We would also likely condition the latent space so we could retarget facial expressions from one person to another by manipulating the differences between latent codes.

My point isn't that the procedurally generated woman is as easy for the AI to conceive in many contexts as Brad, but that there are ways of explicitly reusing parts of images and combining an actor, a hat and a location in a novel but acceptable way is a tough challenge text prompt driven neural networks appear to now perform well at. Manually tagging items in each frame and requesting that they're in the next frame with text prompts doesn't scale, but adding a degree of automation to classifying parts of a frame and updating the data corpus and text prompt for the next frame doesn't seem an insurmountable problem. Many existing CGI techniques don't scale well (even human CGI artists working on Hollywood movies expect instructions on frame composition far more explicit than a director's comments and storyboard; I had a flatmate who was that translation layer). I don't see us getting coherent cinematography from feeding it Hollywood scripts soon if ever and agree there will be significant uncanny valley issues with human mannerisms, but we can do a lot of convincing and selective creative work with text prompts that we couldn't until recently. Especially in combination with other tools, so if DALL-E's concept of geometry is nowhere near capable of generating the side profile of procedurally generated 5'6" Italian woman, we create the woman in a specialised tool for 3D procedural face generation, and feed DALL-E as many images with different angles and lighting as it possesses of Brad. I'd agree we're a long way from doing without human curation and letting the neural networks do all the heavy lifting, but it looks like its now a curation and classification problem rather than a "neural networks can't discern that" problem.

You are reminding me of the Many Worlds interpretation of quantum mechanics.

Not so much. Check out E3GAN from Gordon Wetzstein’s lab at Stanford: https://matthew-a-chan.github.io/EG3D/. In particular, note how they are able to apply a process called GAN Inversion to find a point in the 3D GAN’s representation space to generate a 3D model of a person’s face (one of the collaborators) from a single query image.

Seems like this will also get to a point where we would just take a photo and animate it completely as a 3D mesh. Like that picture, just on a few photos, at some point, would generate a photorealistic, ray-traced 3D asset and also a fully interactable 2D motion paired with GPT-4 or 5 where you could have indistinguishable level of human like interaction (ex. Fable studio)

After a couple dozen more papers and prototypes, we could have ultrasound haptic feedback and some 3D holographic projector where we achieve a true augmented reality.

Going a bit further with this idea, you could essentially create augmented 3D objects that would be indistinguishable from the real thing until you apply enough force enough to break through some upper limit of ultrasound generated haptic pressure (as to not hurt the end user through some legislative regulation). For example, opposing magnets that can be pushed together past a certain point where the illusion of tactility is gone.

What would a reality like that be like? Where you could not only have limitless content media generator but you could conjure up those assets in your living room and have them interact kinetically with the real world. We are talking 3D holographic assets with ultrasound haptic simulator that can apply pressure to any object from all directions.

https://www.youtube.com/watch?v=fwTcfwbrNO4

https://www.youtube.com/watch?v=jDRTghGZ7XU

We might not even need expensive robots in the future. Just conjure up a 3D human via some ultrasound/holograph projector which can apply enough force to objects.

But it's much trickier to find the point in a GAN's representation space for "the exact same fluffy pink earrings that the actress was wearing in the last scene".

Mostly because one can't write code to judge equal-ness of pink earrings. Or the fact the trash can in the background had a lid in the last scene and is now missing the lid. Or the fact the brick wall has turned into a painted wall. etc.

Having said that, I'm sure some filmmakers will go for it anyway - after all, human-made films have some consistency issues, and many viewers simply don't care.

I think the main thing is it becomes a programming problem rather than an AI limitation. From what I've seen it's now possible for the AI to generate EarringsAsset57 and KitchenWall51 and then to ask the AI to draw the actor wearing EaringsAsset57 in front of Kitchenwall51 and for stills at least, it can produce acceptable results

> But it's much trickier to find the point in a GAN's representation space for "the exact same fluffy pink earrings that the actress was wearing in the last scene".

Not so much, since you can supply an image of the desired item and get back a bunch of vectors to reuse.

You can decompose an entire scene into a collection of 3D GANs, each of which is fit by GAN Inversion to the object instance which best explains what is seen in the images.

Also, we absolutely can tell the difference between trash cans with and without lids, as well as brick walls vs painted walls. This falls under many areas of ML-based Computer Vision, such as metric, contrastive, and zero shot machine learning, also open world object detection / recognition. These frame level semantic discrepancies can be detected by discriminative DNNs, which we can use to provide a temporal consistency loss to guide the generator to produce temporally consistent videos.

For film-length consistency, I wouldn’t be surprised by it taking years to solve, despite what @dougabug says in a sibling comment.

But “years” isn’t very much time.

(I have the mental image of it being as easy to make a film as it was for the character of Mariner in that episode of Star Trek Lower Decks where she makes a filmified version of the ship, in the holodeck, in a few seconds; but the more I think about it, the more I realise I think I don’t feel the normal drives for expensive economic signalling in a way that would allow me to actually predict what people would do given the technology).

You need to loop back the generated output to provide context for the next output, just like how GPT-3 works. This at least establishes the needed data to train for consistent outputs. Given a deep well-trained network, the high-level meta-data that needs to be consistent across scenes is just a small set and the network will discover this just as GPT-3 did.

Right. In the literature this is typically referred to as autoregressive sequence generation, and you can use it generate a whole story consistent with a given prompt (of any length), or even a whole image where you give it the first N seed pixels.

> Imagine that you could watch infinite variations of the ending in Sopranos season 6

This sounds like a special version of hell…

I am really quite sceptical about these wild wild claims. There's quite a big difference between "generate an image that at first glance kinda looks like an armchair in the shape of an avocado", and "create actual meaningful content" such as you describe.

Just like several years ago you'd say "there's a big difference between these trippy deep-dream images and an actual meaningful illustration with a style and a coherent theme". But here we are.

>style and coherent theme

Can I play with it then, to verify that myself? Or can I just view cherry-picked examples where it works sort of well?

That's one of the major flaws of nn approaches, since you don't have an understanding of what's going on, just an uninterpretable set of billions of numbers, you can't know what are the failure modes and the limitations of your algorithm.

Lots of people have written big posts at this point about this topic. DALL-E2 let's you guide the image generation with text prompts, and mask part of the image and regenerate the masked parts.

This allows collaboration between the use and the algorithm to iteratively converge on what the user wants. There's also been lots of exploration to find text prompt recipes that produce interesting and more consistent results.

In the end, it's a tool! Tools are used by humans to do something that matters to the human user. Humans adapt to the tools they use to maximize the tool's utility.

Understanding that is fundamental for thinking about where ML will take us. It's not magic, it's tools.

> Just like several years ago you'd say "there's a big difference between these trippy deep-dream images and an actual meaningful illustration with a style and a coherent theme". But here we are.

Exactly, at the very least, that will massively disrupt the fields of illustration and concept art.

If this kind of thing becomes the norm, what will the DALL-E 200 train itself on? At some point it's going to be sucking its own exhaust.

You don't need humans. 'Self-distillation' and related approaches work surprisingly well. Like with GPT-3, you can get much better question-answering or French translation simply by generating a bunch of examples, ranking them with the model's own internal metrics (a simple ranking would be to accept an answer if the majority of one's samples agree on the same answer), and train further on that - with no humans or additional 'real' data involved. Remember that DALL-E 1 got a big quality boost by generating like 50 samples and using CLIP to rank them by similarity to the original text input. (You can also retrain the DALL-E 1 model to run in reverse - image to text - to score images. CogView did this.)

And if you are training on Internet data, then you are presumably training on human-curated selections. Even a minimal level of selection, like picking the best out of a dozen, by a DALL-E 2 API user, or by downstream humans selectively resharing higher-quality images, is a (slow) bootstrap.

> 'Self-distillation' and related approaches work surprisingly well.

Surprising is an understatement. That's nuts.

> it's going to be sucking its own exhaust.

Arguably that’s not too different from what humanity has always done. Inspiration mostly comes from what other humans have done before.

At some point, it will have to evolve beyond learning from human labeled (or even just collected) examples to learn by its own open ended dynamics and self-directed experience. This is analogous to the original AlphaGo bootstrapping from records of millions of human played games, to AlphaZero learning to play Go purely from self-play.

That's a beautiful analogy. At some point Alpha Go learnt all it could from the narratives of humans, and eventually it had to become the creator of its own narratives.

Applied generally enough, we see the narrative value of humans, at least in the eyes of the machine, to go down down down, as the value of its own richer and more unknown narratives go up up up.

In this way it becomes the dominant spectacle on the Earth, leaving humanity with a questionable fate.

Exactly, at that point you let the AIs play against each other and go back to watch some TV. (or in the case of DALL-E, we won't be able to understand the concept/description space of the images anymore, let alone the images generated from it).

I'm having trouble drawing that analogy, because Alpha Go knew what it was to win before it taught itself to play. The equivalent for DALL-E would be to start off knowing what an elephant looks like, what water looks like, what elephants in water look like ...

All analogies are flawed to some degree. The objective is clearly simpler in the case of Go and other strategy game. Although recreating physical objects (either animate or inanimate) is in principle straightforward: you just need a certain number of photos of a representative set of instances each object w/ enough perspective. Similar to learn to reproduce action you upgrade from a set of stills to video clips.

The harder goal of creative content creation might be driven by reward signals came from human likes, clicks, engagement / interaction, downloads, purchases, positive reviews, etc. That would probably lead to catering to pedestrian tastes, accusations of derivative authorship lacking originality. One signal for predictably might be measures of entropy, such as the capacity required to reproduce the story, or how far ahead the story arc could be predicted (“surprise factor,” internal consistency, etc).

To win awards it would need to weight expert critical opinion, eventually learning to simulate and generate meaningful and informed criticism, ultimately leading to an intrinsic judgement wrt quality and originality that holds up to scrutiny. The generators would evolve by both seeking approval from and pushing back against the collective feedback from this community of AI and human critics.

As a thought experiment, imagine if this functionality already existed but only for a small subset. Perhaps for secretive 3-letter agencies?

What would someone with that ability be able to do? How convincing could their imagery, films, etc be? Could they even shift perceptions of the world, get people to do things because of the imagery? How powerful would that ability be - is it an ability that the military etc would have looked into, and investigated possible weaponisation?

Is it possible that this has been the case for years?

It's doubtful.

The asymmetry between theory and scale has been shrinking on most fronts when it comes to the government. The government certainly has brilliant scientists on its payroll and in the past has been able to shove resources at problems that no other entity has, but these are no longer the days of the Manhattan Project and this isn't the atom bomb. You can't throw money at this problem beyond buying computing power (which the private sector now holds more of than the government does), and it requires incremental advancements in dozens of fields internationally over the course of years, along with huge collaboratively captioned datasets that simply can't be produced by the government without tipping their hand.

Being able to create fake films doesn't really buy you anything, when people aren't believing them. And for believability, you need connections with reality and trust, that you don't get from fake films alone. People can just look at other sources and see that none of this happened at the given time and place. And random videos that pop up on the Internet are very low in terms of trust anyway.

If you want to influence public opinion, there are already far more powerful tools. Just be Google or Twitter and create some strategically placed filter bubbles. Everything people will see, actually happened and can be verified. Nothing is hidden and will show up with a specific enough search query. But every less specific search query will nudge people into whatever direction you want them to move.

To counter that, imagine police bodycams with embedded AI touching up live-recording so that the police was always right, suspect always did something wrong. If you build the source device with this touch-up capability, if you are the certificate authority about these, if you are also defining how these recordings would be used in court, then a perfect fake is useful. (None of us see the actual sensor output from our mobile phones)

TL;DR In order for a fake to be useful, you have to fake AND be the only reputable source. If other recorders cannot get close enough, then your touched-up "real" video is the only material available.

> Sopranos season 6 A new episode of TNG, VOY and the Sopranos every single day!

I’m going to try “TNG in the style of DS9 seasons 4-6”. :)

Reply on news.ycombinator.com