This isn't a matter of human-level AI or superhuman-level AI; it's just straight up impossible. If you want the information to match, it has to be provided. If it isn't there, an AI can fill in the gaps with "something" that will make the scene work, but expecting it to fill in the gaps the way you "want" even though you gave it no indication of what that is is expecting literal magic.

Long term, you'll never have a coherent movie produced by stringing together a series of textual snippets because, again, that's just impossible. Some sort of long-form "write me a horror movie staring a precocious 22-year old elf in a far-future Ganymede colony with a message about the importance of friendship" AI that generates a coherent movie of many scenes will have to be doing a lot of some sort of internal communication in an internal language to hold the result together between scenes, because what it takes to hold stuff coherent between scenes is an amount of English text not entirely dissimilar in size from the underlying representation itself. You might as well skip the English middleman and go straight to an embedding not constrained by a human language mapping.

And this applies to language / code outputs as well.

The number of times I’ve had engineers at my company type out 5 sentences and then expect a complete react webapp.

But what I’ve found in practice is using LLMs to generate the prompt with low-effort human input (eg: thumbs up/down, multiple-choice etc) is quite useful. It generates walls of text, but with metaprompting, that’s kind of the point. With this, I’ve definitely been able to get high ROI out of LLMs. I suspect the same would work for vision output.

Stick the video you want to replicate into -o1 and ask for a descriptive prompt to generate a video with the same style and content. Take that prompt and put it into Sora. Iterate with human and o1 generated critical responses.

I suspect you can get close pretty quickly, but I don't know the cost. I'm also suspicious that they might have put in "safeguards" to prevent some high profile/embarrassing rip-offs.

Why snippets? Submit a whole script the way a writer delivers a movie to a director. The (automated) director/DP/editor could maintain internal visual coherence, while the script drives the story coherence.

Yeah, and Apollo 11 would have been utterly astonishing a decade before it occurred. And, yet, if you tried to project out from it to what further frontiers manned spaceflight would reach in the following decades, you’d…probably grossly overestimate what actually occurred.

> Long term progress can be surprising.

Sure, it can be surprising for optimists as well as naysayers; as a good rule of thumb, every curve that looks exponential in an early phase ends up being at best logistic.

Ask anyone with a chronic illness about the future and they'll tell you we're about 5 years off a cure. They've been saying that for decades. Who knows where the future advancements will be.

Generative AI will never produce an experience like that. I know never is a long time, but I’m still gonna call it. You simply can’t produce such a fresh idea by gathering a bunch of data and interpolating.

Maybe someday enough AI will be good enough to create shorter or longer videos with some dialog and even a coherent story (though I doubt it), but it won‘t be fresh or creative. And we humans will at best enjoy it for its stupidity or sloppiness. Not for its cleverness or artistry.

This is the at-first-fun-but-now-frustrating infinite goal move. "AI (a stand in for literally anything) will do (anything) soon." -> "It won't do (thing), it's too complex." -> "Who said AI will do (thing)?"

So no I don’t think this will happen either. Authors may use use AI them selves as one tool in their tool box as they write their script, but we will not see entire production screen plays being written by generative AI set for theatrical release. The industry will simply not allow that to happen. At most you can have AI write a screen play for your own amusement, not for publication.

Generative AI will not be able to approach the artistry of your average actor (not even a bad actor), it won’t be able match the lighting or the score to the mood (unless you carefully craft that in your prompt). It won‘t get creative with the camera angles (again unless you specifically prompt for a specific angle) or the cuts. And it probably won’t stay consistent with any of these, or otherwise break the consistency at the right moments, like an artist could.

If you manage to prompt the generative AI to create a full feature film with excellent acting, the correct lighting given the mood, a consistent tone with editing to match, etc. you have probably spent much more time and money into crafting the prompt than would otherwise have gone into simply hiring the crew to create your movie. The AI movie will certainly contain slop and be visibly so bad it guaranteed will not be in theaters.

Now if you hired that crew to make the movie instead, that crew might use AI as a tool to enhance their artistry, but you still need your specialized artists to use that tool correctly. That movie might make it to the theaters.

In music you also have plenty of artists that have no clue how to play their instruments, or progress their songs, but the music is nonetheless amazing.

Skill is not the only quality of art. A brilliant artist works with their limitation to produce work which is better than the sum of its part. It will take AI the luck of ten billion universes before it produces anything like that.

The AI isn't creating the fresh ideas. People are.

Saying that large video models will be in theaters sounds like a completely different and much more ambitious prediction. I interpreted it as if large video models will produce whole movies on their own from a script of prompts. That there will be a single film maker with only a large video model and some prompts to make the movie. Such films will never be in the theater, unless by some grifter, and than it is certain to be a flop.

What does "EXT. NIGHT" mean in a script? Is it cloudy? Rainy? Well lit? What are camera locations? Is the scene important for the context of the movie? What are characters wearing? What are they looking at?

What do actors actually do? How do they actually behave?

Here are a few examples of script vs. screen.

Here's a well described script of Whiplash. Tell me the one hundred million things happening on screen that are not in the script: https://www.youtube.com/watch?v=kunUvYIJtHM

Or here's Joker interrogation from The Dark Night Rises. Same million different things, including actors (or the director) ignoring instructions in the script: https://www.youtube.com/watch?v=rqQdEh0hUsc

Here's A Few Good Men: https://www.youtube.com/watch?v=6hv7U7XhDdI&list=PLxtbRuSKCC...

and so on

---

Edit. Here's Annie Atkins on visual design in movies, including Grand Budapest Hotel: https://www.youtube.com/watch?v=SzGvEYSzHf4. And here's a small article summarizing some of it: https://www.itsnicethat.com/articles/annie-atkins-grand-buda...

Good luck finding any of these details in any of the scripts. See minute 14:16 where she goes through the script

Edit 2: do watch The Kerning chapter at 22:35 to see what it actually takes to create something :)

This is most Hacker News comments summarized lmao. It's kinda my favorite thing of this place: just open any thread and you immediately see so many people rushing to say ''well just do X or Y'' or ''actually it's X or Y and not Z like the experts claim''. Love it.

Of course, HN being the place that it is, the same type of comments are made about quantum entanglement and solar panel efficiency.

At the same time I am curious in the "that person has too many fingers" sense at what a system trained on tens of thousands of movies plus scripts plus subtitles plus metadata etc. would generate.

I thought about it for a bit and I would want to watch a computer generated Sharknado 7 or Hallmark Christmas movie.

Lets pick something concrete. It's a medieval script, it opens with two knights fighting. OK so later in the script we learn their characters, historic counterparts etc. So your LLM can match nefarious villain to some kind of embedding, and doubtless has trained on countless images of a knight.

But the result is not naively going to understand the level of reality the script is going for - how closely to stick to historic parallels, how much to go fantastical with the depiction. The way we light and shoot the fight and how it coheres with the themes of the scene, the way we're supposed to understand the characters in the context of the scene and the overall story, the references the scene may be making to the genre or even specific other films etc.

This is just barely scraping the surface of the beginnings of thinking about mise en scene, blocking, framing etc. You can't skip these parts - and they're just as much of a challenge as temporal coherence, or performance generation or any of the other hard 'technical issues' that these models have shown no capacity to solve. They're decisions that have to be made to make a film coherent at all - not yet good or tasteful or creative or whatever.

Put another way - you'd need AGI to comprehend a script at the level of depth required to do the job of any HOD on any film. Such a thing is doubtless possible, but it's not going to be shortcut naively the way generation an image is - because it requires understanding in context, precisely what LLMs lack.

We can already get detailed style guidance into picture generation. Declaring you want Picasso cubist, Warner brothers cartoon, or hyper realistic works today. So does lighting instructions, color palettes, on and on.

These future models will not be large language models, they will be multi-modal. Large movie models if you like. They will have tons of context about how scenes within movies cohere, just as LLMs do within documents today.

- you have to provide correct detailed instructions on lighting

- you have to provide correct detailed instructions on props

- you have to provide correct detailed instructions on clothing

- you have to provide correct detailed instructions on camera position and movement

- you have to provide correct detailed instructions on blocking

- you have to provide correct detailed instructions on editing

- you have to provide correct detailed instructions on music

- you have to provide correct detailed instructions on sound effects

- you have to provide correct detailed instructions on...

- ...

- repeat that for literally every single scene in the movie (up to 200 in extreme cases)

There's a reason I provided a few links for you to look at. I highly recommend the talk by Annie Atkins. Watch it, then open any movie script, and try to find any of the things she is talking about there (you can find actual movie scripts here: https://imsdb.com)

That's how I've been using the image generators - lots of experimentation and throwing out the stuff that doesn't work. Then once I've got enough good generated images collected out of the tons of garbage, I fine tune a model and create a workflow that more consistently gives me those styles.

Now the models and UX to do this at a cinematic quality are probably 5-10 years away for video (and the studios are probably the only ones with the data to do it), but I'm relatively bullish on AI in cinema. I don't think AI will be doing everything end to end, but it might be a shortcut for people who can write a script and figure out the UX to execute the rest of the creative process by trial and error.

Where did you find AI/ML that are good at filling in actual required and consistent details.

I beg of you to watch Annie Atkins' presentation I linked: https://www.youtube.com/watch?v=SzGvEYSzHf4 and tell me how much intervention would AI/ML need to create all that, and be consistent throughout the movie?

> once these models can generate coherent scenes, people can start using them to explore the creative space and figure out what they like.

Define "coherent scene" and "explore". A scene must be both coherent and consistent, and conform to the overall style of the movie and...

Even such a simple thing as shot/reverse shot requires about a million various details and can be shot in a million different ways. Here's an exploration of just shot/reverse shot: https://www.youtube.com/watch?v=5UE3jz_O_EM

All those are coherent scenes, but the coherence comes from a million decisions: from lighting, camera position, lens choice, wardrobe, what surrounds the characters, what's happening in the background, makeup... There's no coherence without all these choices made beforehand.

Around 4:00 mark: "Think about how well you know this woman just from her clothes, and workspace". Now watch that scene. And then read its description in the script https://imsdb.com/scripts/No-Country-for-Old-Men.html:

--- start quote ---

    Chigurh enters. Old plywood paneling, gunmetal desk, litter
          of papers. A window air-conditioner works hard.
          A fifty-year-old woman with a cast-iron hairdo sits behind
          the desk.

And right after that there's a section on the rhythm of editing. Another piece in the puzzle of coherence in a scene.

> Then once I've got enough good generated images collected out of the tons of garbage, I fine tune a model and create a workflow that more consistently gives me those styles.

So, literally what I wrote here: https://news.ycombinator.com/item?id=42375280 :)

You’re really failing to let go of the idea that you need to prescribe every little thing. Like Midjourney today, you’ll be able to give general guidance.

Now, I don’t expect we’ll get the best movies this way. But paint by numbers stuff like many movies already are? A Hallmark Channel weepy? I bet we will.

No jump.

Your original claim: "Submit a whole script the way a writer delivers a movie to a director. The (automated) director/DP/editor could maintain internal visual coherence, while the script drives the story coherence."

Two comments later it's this: "We can already get detailed style guidance into picture generation. Declaring you want Picasso cubist, Warner brothers cartoon, or hyper realistic works today. So does lighting instructions, color palettes, on and on."

I just re-wrote this with respect to movies.

> I was thinking more like ‘make it look like Bladerunner if Kurosawa directed it, with a score like Zimmer.’

Because, as we all know, every single movie by Kurosawa is the same, as is every single score by Hans Zimmer, so it's ridiculously easy to recreate any movie in that style, with that music.

> You’re really failing to let go of the idea that you need to prescribe every little thing. Like Midjourney today, you’ll be able to give general guidance.

Yes, and Midjounrey today really sucks at:

- being consistent

- creating proper consistent details

A general prompt will give you a general result that is usually very far from what you actually have in mind.

And yes, you will have to prescribe a lot of small things if you want your movie to be consistent. And for your movie to make any sense.

Again, tell me how exactly your amazing magical AI director will know which wardrobe to chose, which camera angles to setup, which typography to use, which sound effects to make just from the script you hand in?

you can start ,with a very simple scene I referenced in my original reply: two people talking at the table in Whiplash.

> But paint by numbers stuff like many movies already are? A Hallmark Channel weepy? I bet we will.

Even those movies have more details and more care than you can get out of AIs (now, or in foreseeable future)

I think you're still assuming I always want to choose those things. That's why we're talking past each other. A good movie making model would choose for me unless I give explicit directions. Today we don't see long-range coherence in the results of movie (or game engine) models, but the range is increasing, and I'm willing to bet we will see movie-length coherence in the next decade or so.

By the way, I also bet that if I pasted exactly the No Country for Old Men script scene description from up this thread into Midjourney today it would produce at least some compelling images with decent choices of wardrobe, lighting, set dressing, camera angle, exposure, etc etc. That's what these models do, because they're extrapolating and interpolating between the billion images they've seen that contained these human choices.

AFAIK Midjourney produces single images, so the relevant scope of consistency is inside the single image only. Not between images. A movie model needs coherence across ~160,000 images, which is beyond the state of the art today but I don't see why it's impossible or unreasonable in the long run.

> A general prompt will give you a general result that is usually very far from what you actually have in mind.

Which is only a problem if I have something in mind. Alternatively I can give no guidance, or loose guidance, make half a dozen variations, pick the one I like best. Maybe iterate a couple of times into that variation tree. Just like the image generators do.

https://x.com/minchoi/status/1862975323433795726

"movies will be one of the last things to be replaced by ai"

https://www.youtube.com/watch?v=ypURoMU3P3U

including this quote: "being a craftsman is knowing how to work, art is knowing when to stop"

The text encoder may not be able to know complex relationships, but the generative image/video models that are conditioned on said text embeddings absolutely can.

Flux, for example, uses the very old T5 model for text encoding, but image generations from it can (loosely) adhere to all rules and nuances in a multi-paragraph prompt: https://x.com/minimaxir/status/1820512770351411268

Flux certainly does not consistently do so across an arbitrary collection of multi-paragraph prompts, as anyone whose run more than a few long prompts past it would recongize; also, the tweet is wrong in the other direction, as well, longer language-model-preprocessed prompts for models that use CLIP (like various SD1.5 and SDXL derivatives) are, in fact, a common and useful technique. (You’d kind of think that the fact that generated prompt here is significantly longer than the 256 token window of T5 would be a clue that the 77 token limit of CLIP might not be as big of a constraint as the tweet was selling it as, too.)

How would you ever tweak or debug it in that case? It doesn't strictly have to be English, but some kind of human-readable representation of the intermediate stages will be vital.

https://dreambooth.github.io/ https://textual-inversion.github.io/

InstantID (https://replicate.com/zsxkib/instant-id) fixes that issue.

If you just want a face, InstandID/Pulid work - but it’s not going to be very varied. Doing actual training means you can get any perspective, lighting, style, expression, etc - and have the whole body be accurate.

Even the models based off danbooru and E621 still aren't the best at that. And us furries like to tag art in detail.

The best we can really do at the moment is regional prompting, perhaps they need something similar for video.

Sora performs worse than closed source Kling and Hailuo, but more importantly, it's already trumped by open source too.

Tencent is releasing a fully open source Hunyuan model [1] that is better than all of the SOTA closed source models. Lightricks has their open source LTX model and Genmo is pushing Mochi as open source. Black Forest Labs is working on video too.

Sora will fall into the same pit that Dall-E did. SaaS doesn't work for artists, and open source always trumps closed source models.

Artists want to fine tune their models, add them to ComfyUI workflows, and use ControlNets to precision control the outputs.

Images are now almost 100% Flux and Stable Diffusion, and video will soon be 100% Hunyuan and LTX.

Sora doesn't have much market apart from name recognition at this point. It's just another inflexible closed source model like Runway or Pika. Open source has caught up with state of the art and is pushing past it.

[1] https://github.com/Tencent/HunyuanVideo

The real trick is that the AI needs to be able to participate in iteration cycles, where the human can say "okay this is all mostly good, but I've circled some areas that don't look quite right and described what needs to be different about them." As far as I've played with it, current AIs aren't very good at revisiting their own work— you're basically just tweaking the original inputs and otherwise starting over from scratch each time.

https://www.shapeshift.ink/

Check out the Banodoco Discord community [2]. These are the people pioneering steerable AI video, and it's all being built on top of open source.

[1] https://github.com/Tencent/HunyuanVideo

[2] https://banodoco.ai/

Now expand that to movies and games and you can get why this whole generative-AI bubble is going to pop.

What will save it is that, no matter how picky you are as a creator, your audience will never know what exactly was that you dreamed up, so any half-decent approximation will work.

In other words, a corollary to your corollary is, "Fortunately, you don't need them to be, because no one cares about low-order bits".

Or, as we say in Poland, "What the eye doesn't see, the heart doesn't mourn."

Part of the problem is the "half decent approximations" tend towards a clichéd average, the audience won't know that the cool cyberpunk cityscape you generated isn't exactly what you had in mind, but they will know that it looks like every other AI generated cyberpunk cityscape and mentally file your creation in the slop folder.

I think the pursuit of fidelity has made the models less creative over time, they make fewer glaring mistakes like giving people six fingers but their output is ever more homogenized and interchangable.

https://www.astralcodexten.com/p/how-did-you-do-on-the-ai-ar...

In other words, someone willing to tweak the prompt and press the button enough times to say "yeah, that one, that's really good" is going to have a result which cannot in fact be reliably binned as AI-generated.

The detailed breakdown you mention? Maybe it's accurate to that artist's thought process, maybe it's more of a rationalization; either way, it's not a general rule they, or anyone, could apply to any of the other AI images. Most of those in the article don't exhibit those "telltale signs", and the one that does - the Victorian Megaship - was actually made by human artist with no AI in the mix.

EDIT:

Another image that stands out to me is Riverside Cafe. Myself, like apparently a lot of other people, going by articles' comments, assumed it's a human-made one, because we vaguely remembered Vang Gogh painted something like it. He did, it's called Café Terrace at Night - and yet, despite immediately evoking the association, Riverside Cafe was made by AI, and is actually nothing like Café Terrace at Night at any level.

(I find it fascinating how this work looks like a copy of Van Gogh at first glance, for no obvious reason, but nothing alike once you pause to look closer. It's like... they have similar low-frequency spectra or something?)

EDIT2:

Played around with the two images in https://ejectamenta.com/imaging-experiments/fourifier/. There are some similarities in the spectra, I can't put my finger on them exactly. But it's probably not the whole answer. I'll try to do some more detailed experimentation later.

--

[0] - Nor should you expect it - it would mean either a perfect calibration, or be the equivalent of flipping a coin and getting heads 30 times in a row; it's not impossible, but you shouldn't expect to see it unless you're interviewing fewer people than literally the entire population of the planet.

> The average participant scored 60%, but people who hated AI art scored 64%, professional artists scored 66%, and people who were both professional artists and hated AI art scored 68%.

> The highest score was 98% (49/50), which 5 out of 11,000 people achieved. Even with 11,000 people, getting scores this high by luck alone is near-impossible.

If 0.0005% of people who are specifically judging art as AI or not AI, in a test which presumably attracts people who would like to be able to do that thing, can do a 98% accurate job, and the average is around 60%: that isn't reliable.

If that doesn't work for you, I encourage you to take the test. Obviously since you've read the article there are some spoilers, but there's still plenty of chances to get it right or wrong. I think you'll discover that you, too, cannot do this reliably. Let us know what happens.

Imagine the space of ideas as a circle, with stuff in the middle being more easy to reach (the "cliched average"). Previously, traversing the circle was incredibly hard - we had to use tools like DeviantArt, Instragram, etc to agglomerate the diverse tastes of artists, hoping to find or create the style we're looking for. Creating the same art style is hiring the artist. As a result, on average, what you see is the result of huge amounts of human curation, effort, and branding teams.

Now reduce the effort 1000x, and all of a sudden, it's incredibly easy to reach the edge of the circle (or closer to it). Sure, we might still miss some things at the very outer edge, but it's equivalent to building roads. Motorists appear, people with no time to sit down and spend 10000 hours to learn and master a particular style can simply remix art and create things wildly beyond their manual capabilities. As a result, the amount of content in the infosphere skyrockets, the tastemaking velocity accelerates, and you end up with a more interesting infosphere than you're used to.

When you're commissioning an artist to make you some art, you're basically sampling from the entire distribution. Stuff in the middle is, as you say, easiest to reach, so that's what you'll most likely get. Generative models let more people do art, meaning there's more sampling happening, so the stuff further from the centre will be visited more often, too.

However, AI tools also make another thing easier: moving and narrowing the sampling area. Much like with a very good human artist, you can find some work that's "out there", and ask for variations of it. However, there are only so many good artists to go around. AI making this process much easier and more accessible means more exploration of the circle's edges will happen. Not just "more like this weird thing", but also combinations of 2, 3, 4, N distinct weird things. So in a way, I feel that AI tools will surface creative art disproportionally more than it'll boost the common case.

Well, except for the fly in the ointment that's the advertising industry (aka. the cancer on modern society). Unfortunately, by far most of the creative output of humanity today is done for advertising purposes, and that goal favors the common, as it maximizes the audience (and is least off-putting). Deluge of AI slop is unavoidable, because slop is how the digital world makes money, and generative AI models make it cheaper than generative protein models that did it so far. Don't blame AI research for that, blame advertising.

Tastes are almost never normally distributed along a spectrum, but multi-modal. So the more dimensions you explore in, the more you end up with “islands of taste” on the surface of a hyper sphere and nothing like the normal distribution at all. This phenomenon is deeply tied to why “design by committee” (eg, in movies) always makes financial estimates happy but flops with audiences — there is almost no customer for average anything.

I agree with your conclusion.

My experience with customer surveys indicates the opposite — that customers prefer you have an opinion.

Inside Out 2 had the largest box office of any movie in 2024. Checkout the "research and writing" section in its wikipedia article https://en.wikipedia.org/wiki/Inside_Out_2#Research_and_writ... ... psychological consultants, a feedback loop with a group of teenagers, test screenings.

Or how about "Die with a smile" - currently number 1 in the global top 50 on Spotify. 5 songwriters

Or "APT." - currently number 2 in the global top 50 on Spotify. 11 songwriters

You don't have to look very hard

Consulting with SMEs, testing with audiences, etc isn’t “design by committee”.

Similarly, “Die With a Smile” seems to have been the work of two people with developed styles with support — again, not a committee:

> The collaboration was a result of Mars inviting Gaga to his studio where he had been working on new music. He presented the track in progress to her and the duo finished writing and recording the song the same day.

Apt seems to have started with a single person goofing around, then pitched as a collaboration and the expanded team entered at that point.

(And I do blame the advertisers, but frankly anyone handing them new amplifiers, with entirely predictable consequences, is also not blameless.)

That is based on perception that it is easier than ever to create fake content, but fails to account for the fact that creating real content (for example, simply taking a video) is even much easier. So while there is more fake content, there is also lot more real content, and so manipulation of reality (for example, denying a genocide) is much harder today than ever.

Anyway, "the AI slop will win" is based on a similar misconception, that total creative output will not increase. But like with fake news, it probably will not be the case, and so the actual amount of good art will increase, too.

I think we are OK as long as normal humans prefer to create real news rather than fake news, and create innovative art rather than cliched art.

So we're not OK.

I think I need to state my assumptions/beliefs here more explicitly.

First of all, "AI slop" is just the newest iteration on human-produced slop, which we're already drowning in. Not because people prefer to create slop, but because they're paid to do it, because most content is created by marketers and advertisers to sell you shit, and they don't want it to be better than strictly necessary for purpose.

It's the same with fake news, really. Fake news isn't new. Almost all news is fake news; what we call "fake news" is a particular flavor of bullshit that got popular as it got easier for random humans to publish stories competing with established media operations.

In both cases, AI is exacerbating the problem, but it did not create it - we were already drowning in slop.

Which leads me to related point:

> Anyway, "the AI slop will win" is based on a similar misconception, that total creative output will not increase.

It will. But don't forget Sturgeon's law - "ninety percent of everything is crap"[0]. Again, for the past couple decades, we've been drowning in "creative output". It's not a new problem, it's just increasingly noticeable in the past years, because the Web makes it very easy for everyone to create more "creative output" (most of which is, again, advertising), and it finally started overwhelming our ability to filter out the crap and curate the gems.

Adding AI to the mix means more output, which per Sturgeon's law, means disproportionately more crap. That's not AI's fault, that's ours; it's still the same problem we had before.

--

[0] - https://en.wikipedia.org/wiki/Sturgeon%27s_law

That may be true of any one model (though I don’t think it really is, either, I think newer image gen models are individually capable of a much wider array of styles than earlier models), but it is pretty clearly not true of the whole range of available models, even if you look at a single model “family” like “SDXL derivatives”.

Ironically, we're long past that point with human creators, at least when it comes to movies and games.

Take sci-fi movies, compare modern ones to the ones from the tail end of the 20th century. Year by year, VFX gets more and more detailed (and expensive) - more and better lights, finer details on every material, more stuff moving and emitting lights, etc. But all that effort arguably killed immersion and believability, by making scenes incomprehensible. There's way too much visual noise in action scenes in particular - bullets and lighting bolts zip around, and all that detail just blurs together. Contrast the 20th century productions - textures weren't as refined, but you could at least tell who's shooting who and when.

Or take video games, where all that graphics works makes everything look the same. Especially games that go for realistic style, they're all homogenous these days, and it's all cheap plastic.

(Seriously, what the fuck went wrong here? All that talk, and research, and work into "physically based rendering", yet in the end, all PBR materials end up looking like painted plastic. Raytracing seems to help a bit when it comes to liquids, but it still can't seem to make metals look like metals and not Fischer-Price toys repainted to gray.)

So I guess in this way, more precision just makes the audience give up entirely.

> they will know that it looks like every other AI generated cyberpunk cityscape and mentally file your creation in the slop folder.

The answer here is the same as with human-produced slop: don't. People are good at spotting patterns, so keep adding those low-order bits until it's no longer obvious you're doing the same thing everyone else is.

EDIT: Also, obligatory reminder that generative models don't give you average of training data with some noise mixed up; they sample from learned distribution. Law of large numbers apply, but it just means that to get more creative output, you need to bias the sampling.

Western movie studios may discover the same thing soon, with the number of high-budget productions tanking lately.

[0] https://en.wikipedia.org/wiki/Fortiche

[1] https://en.wikipedia.org/wiki/Arcane_(TV_series)

If you are not beholden to a precise vision or maybe just want to create something that sells, these tools will likely be significant productivity multipliers.

So far ChatGPT is not for writing books, but is great for SEO-spam blogposts. It is already killing the content marketing industry.

So far Dall-E is not for making master paintings, but it's great for stock images. It might kill most of the clipart and stock image industry.

So far Udio and other song generators are not able to make symphonies, but it's great for quiet background music. It might kill most of the generic royalty-free-music industry.

Actual long form art like a movie works because it includes many well informed choices that work together as a whole.

There seems to be a large gap between generating a few seconds of video vaguely like one's notion, and trying to create 90 minutes that are related and meaningful.

Which doesn't mean that you can't build from this starting place build more robust tools. But if you think that this is a large, hard amount of work, it certainly could call into question optisimitic projections from people who don't even seem to notice that there is work need at all.

It's less obvious because people project personality onto the content they see, because they implicitly assume the artist cared, and had some vision in mind. Cheap shit doesn't look like cheap shit in isolation. Except when you know it's AI-generated, because this removes the artist from the equation, and with it, your assumptions that there's any personality involved.

People can generally see the lack of artistic intent when consuming entertainment.

So, while GenAI tools make it easier to create superficially decent work that lacks creative intent, the studios managed to do it just fine with human intelligence only, suggesting the problem isn't AI, but the studios and their modern management policies.

Right now AI is more the latter, but many people want it to be the former

A director letting actors "just be" knows exactly what he/she wants, and choses actors accordingly. Just as the directors that want the most minute detail.

Clint Eastwood tries to do at most one take of a scene. David Fincher is infamous for his dozens of takes.

AI is neither Fincher nor Eastwood.

People may not think they care, but obviously they do. That’s why marvel movies do better than DC ones.

People absolutely care about details in their media.

Not all details matter, some do. And, it's better to not show the details at all, than to be inconsistent in them.

Like, idk., don't identify a bomb as a specific type of existing air-fuel ordnance and then act about it as if it was a goddamn tactical nuke. Something along these lines was what made me stop watching Arrow series.

This is a key observation, unfortunately generally solving for what details matter is extremely difficult.

I don’t think video generation models help with that problem, since you have even less control of details than you do with film.

At least before post.

The movies have just had much worse audience and critical reception.

The last production I worked on averaged 16 hours per frame for the final rendering. The amount of information encoded in lighting, models, texture, maps, etc is insane.

I don’t remember how long the final rendering took but it was nearly two months and the final compute budget was 7 or 8 figures. I think we had close to 100k cores running at peak from three different render farms during crunch time, but don’t take my word for it I wasn’t producing the picture.

Weren't the rendering algos ported to CUDA yet?

The 3D you see in things like commercials is usually done on GPUs though because at their smaller scale it's much faster.

A friend recently told me about a complex scene (I think it was a Marvel or Star Wars flick) where they had so much going on in the scene with smoke, fire, and other special effects that they had to wait for a specialized server with 2TB of RAM to be assembled. They only had one such machine so by the time the rest of the movie was done rendering, that one scene still had a month to go.

https://www.disneyanimation.com/data-sets/?drawer=/resources...

https://datasets.disneyanimation.com/moanaislandscene/island...

> When everything is fully instantiated the scene contains more than 15 billion primitives.

Pixar's stuff famously takes days per frame.

Do you have a citation for this? My guess would be much closer to a couple of hours per frame.

Trust me, even if you work with human artists, you'll keep saying "it's not quite I initially invisioned, but we don't have budget/time for another revision, so it's good enough for now." all the time.

If you find a silver bullet then everything else is largely irrelevant.

Idk if you noticed but that “if” is carrying an insane amount of weight.

I have a bad feeling that you'd be a horrible manager if you ever were one.

(2021) https://arxiv.org/abs/2103.13915 : An Image is Worth 16x16 Words, What is a Video Worth?

(2024) https://arxiv.org/abs/2406.07550 : An Image is Worth 32 Tokens for Reconstruction and Generation

First paper is the most famous and prompted a lot of research to using text generation tools in the image generation ___domain : 256 "words" for an image, Second paper is 24 reference image per minutes of video, Third paper is a refinement of the first saying you only need 32 "tokens". I'll let you multiply the numbers.

In kind of the same way as a who's who game, where you can identify any human on earth with ~32bits of information.

The corollary being that contrary to what parent is telling there is no theoretical obstacle to obtaining a video from a textual description.

These papers, from my quick skim (tho I did read the first one fully years ago,) seem to show that some images and to an extent video can be generated from discrete tokens, but does not show that exact images nor that any image can be.

For instance, what combination of tokens must I put in to get _exactly_ Mona Lisa or starry night? (Tho these might be very well represented in the data set. Maybe a lesser known image would be a better example)

As I understand, OC was saying that they can’t produce what they want with any degree of precision since there’s no way to encode that information in discrete tokens.

VQ-VAE (Vector Quantised-Variational AutoEncoder), (2017) https://arxiv.org/abs/1711.00937

The whole encoding-decoding process is reversible, and you only lose some imperceptible "details", the process can be either trained with a L2Loss, or a perceptual loss depending what you value.

The point being that images which occurs naturally are not really information rich and can be compressed a lot by neural networks of a few GB that have seen billions of pictures. With that strong prior, aka common knowledge, we can indeed paint with words.

Taking an existing image and reversing the process to get the tokens that led to it then redoing that doesn’t seem the same as inserting token to get a precise novel image.

Especially since, as you said, we’d lose some details, it suggests that not all images can be perfectly described and recreated.

I suppose I’ll need to play around with some of those techniques.

Natural Image-> Sequence of token, but not all possible sequence of token will be reachable. Like plenty of letters put together form non-sensical words.

Sequence of token -> Natural Image : if the initial sequence of token is unsensical the Natural image will be garbage.

So usually you then modelize the sequence of token so that it produce sensical sequences of token, like you would with a LLM, and you use the LLM to generate more tokens. It also gives you a natural interface to control the generation of token. You can express with words what modifications to the image you should do. Which will allow you to find the golden sequence of token which correspond to the mona-lisa by dialoguing with the LLM, which has been trained to translate from english to visual-word sequence.

Alternatively instead of a LLM you can use a diffusion model, the visual words usually are continuous, but you can displace them iteratively with text using things like "controlnet" (stable diffusion).

Thats my full quote on this topic. And I think it stands. Sure, people won't describe a picture. instead, they will take an existing picture or video, and do modifications of it, using AI. That is much much simpler and more useful, if you can file a scene, and then animate it later with AI.

The prior sentence does not imply the conclusion.

A word is worth a thousand pictures. (E.g Love)

It is abstraction all the way

Being too early about this and being wrong are the same.

Especially anything involving fluid/smoke dynamics, or fast dynamic momements of humans and animals all suffer from the same weird motion artifacts. I can't describe it other than that the fluidity of the movements are completely off.

And as all genai video tools I've used are suffering from the same problem, I wonder if this is somehow inherent to the approach & somehow unsolvable with the current model architectures.

I saw a Santa dancing video today and the suspension of disbelief was almost instantly dispelled when the cuffs of his jacket moved erratically. The GenAI was trying to get them to sway with arm movements but because it didn't understand why they would sway it just generated a statistical approximation of swaying.

GenAI also definitely doesn't understand 3D structures easily demonstrated by completely incorrect morphological features. Even my dogs understand gravity, if I drop an object they're tracking (food) they know it should hit the ground. They also understand 3D space, if they stand on their back legs they can see over things or get a better perspective.

I've yet to see any GenAI that demonstrates even my dogs' level of understanding the physical world. This leaves their output in the uncanny valley.

To that end, it is actually extremely important to nit-pick this stuff. For those of us using these tools, we need to be able to talk shop about which ones are keeping up, which are work like shit in practice, and which ones work but only in certain situations, and which situations those are.

We would love to learn more about the origin of your certainty.

Fascinating! I wish I had the knowledge and wherewithal to do that and become rich instead of wasting my time on HN.

Neural networks are largely black box piles of linear algebra which are massaged to minimize a loss function.

How would you incorporate smooth kinematic motion in such an environment?

The fact that you discount the knowledge of literally every single employee at OpenAI is a big signal that you have no idea what you’re talking about.

I don’t even really like OpenAI and I can see that.

I have been working with NLP and neural networks since 2017.

They aren’t just black boxes, they are _largely_ black boxes.

When training an NN, you don’t have great control over what parts of the model does what or how.

Now instead of trying to discredit me, would you mind answering my question? Especially since, as you say, the theory is so simple.

How would you incorporate smooth kinematic motion in such an environment?

You’ve convinced me that you’re small and know very little about the subject matter.

You don’t need to reply to this. I’m done with this convo.

Do you know anyone or any companies working on that?

Maybe we're in a honeymoon period where your average user hasn't gotten annoyed by all the slop out there and they will soon, but at least for now, there is real value here. Yes, out of 20 ads maybe only 2 outperform the manually created ones, but if I can create those 20 with a couple hundred bucks in GenAI credits and maybe an hour or two of video editing that process wipes the floor with the competition, which is several thousand dollars per ad, most of which are terrible and end up thrown away, too. With the way the platforms function now, ad creative is quickly becoming a volume-driven "throw it at the wall and see what sticks" game, and AI is great for that.

It’s this. A video ad with a person morphing into a bird that takes off like a rocket with fire coming out of its ass, sure it might perform well because we aren’t saturated with that yet.

You’d probably get a similar result by giving a camera to a 5 year old.

But you also have to ask what that’s doing long term to your brand.

My guess is that the criticism of AI not being that good is correct, but many people don't realize that most humans also aren't that good, and that it's quite possible that the AI performs better than mediocre humans.

This shouldn't be much of a surprise, we've seen automation replace low skilled labor in a lot of industries. People seem uncomfortable with the possibility that there's actually a lot of low skilled labor in the creative industry that could also be easily replaced.

All the generated video startups seem to generate videos with much lower than 10% usable output, without significant human-guided edits. Given the massive amount of compute needed to generate a video relative to hyperoptimized LLMs, the quality issue will handicap gen video for the foreseeable future.

The same "prompt" they'd give the creative person they hired... Say, "I want an ad for my burgers that make it look really good, I'm thinking Christmas vibes, it should emphasize our high quality meat, make it cheerful, and remember to hint at our brand where we always have smiling cows."

Now that creative person would go make you that advert. You might check it, give a little feedback for some minor tweaks, and at some point, take what you got.

You can do the same here. The difference right now is that it'll output a lot of junk that a creative person would have never dared show you, so that initial quality filtering is missing. But on the flip side, it costs you a lot less, can generate like 100 of them quickly, and you just pick one that seems good enough.

That's what GenAI is doing, too. After all, the audience only sees the final product; they never get know what the writer had in mind.

If a 60 year old grizzled detective is introduced in page 1, a human artist will draw the same grizzled detective in page 2, 3 and so on, not procedurally generate a new grizzled detective each time.

Most models people currently use don't keep state between invocations, and whatever interpretation they make from provided context (e.g. reference image, previous frame) is surface level and doesn't translate well to output. This is akin to giving each panel in a comic to a different artist, and also telling them to sketch it out by their gut, without any deep analysis of prior work. It's a big limitation, alright, but researchers and practitioners are actively working to overcome it.

(Same applies to LLMs, too.)

https://www.comicsexperience.com/wp-content/uploads/2018/09/...

Or if you can't do this, explain why the feature you mentioned cannot do this, and what it or good for?

- An LLM to inflate descriptions in the script to very detailed prompts (equivalent to artist thinking up how characters will look, how the scene is organized);

- A step to generate a representative drawing of every character via txt2img - or more likely, multiple ones, with a multimodal LLM rating adherence to the prompt;

- A step to generate a lot of variations of every character in different poses, using e.g. ControlNet or whatever is currently the SOTA solution used by the Stable Diffuison community to create consistent variations of a character;

- A step to bake all those character variations into a LoRA;

- Finally, scenes would be generated by another call to txt2img, with prompts computed in step 1, and appropriate LoRAs active (this can be handled through prompt too).

Then iterate on that, e.g. maybe additional img2img to force comic book style (with a different SD derivative, most likely), etc.

Point being, every subproblem of the task has many different solutions already developed, with new ones appearing every month - all that's left to have an "AI artist" capable of solving your challenge is to wire the building blocks up. For that, you need just a trivial bit of Python code using existing libraries (e.g. hooking up to ComfyUI), and guess what, GPT-4 and Claude 3.5 Sonnet are quite good at Python.

EDIT: I asked Claude to generate "pseudocode" diagram of the solution from our two comments:

http://www.plantuml.com/plantuml/img/dLLDQnin4BthLmpn9JaafOR...

Each of the nodes here would be like 3-5 real ComfyUI nodes in practice.

I suspect if someone went to the trouble to implement your above solution they'd find the end result isn't as good as they'd hoped. In practice you'd probably find one or more steps don't work correctly- for example, maybe today's multimodal LLM's can't evaluate prompt adherence acceptably. If the technology was ready the evidence would be pretty clear- I'd expect to see some very good, very quickly made comic books shown off by AI enthusiast on reddit rather then the clearly limited/ not very good comic book experiments which have been demonstrated so far.

A human has to work at it too; more than few hours when doing more than few quick sketches (memory has its limits; there's a reason artists keep reference drawings around), and obviously they already put years into learning their skills than before, but fair - the human artist already knows how to do things that any given model doesn't yet[0], we kind of have to assemble the overall flow ourselves for now[1].

Then again, you only need to assemble it once, putting those hours of work up front - and if it's done, and it works, it becomes fair to say that AI can, in fact, generate self-consistent comic books.

> I suspect if someone went to the trouble to implement your above solution they'd find the end result isn't as good as they'd hoped. In practice you'd probably find one or more steps don't work correctly- for example, maybe today's multimodal LLM's can't evaluate prompt adherence acceptably.

I agree. I obviously didn't try this myself either (yet, I'm very tempted to try it, to satisfy my own curiosity). However, between my own experience with LLMs and Stable Diffusion, and occasionally browsing Stable Diffusion subreddits, I'm convinced all individual steps work well (and have multiple working alternatives), except for the one you flagged, i.e. evaluating prompt adherence using multimodal LLM - that last one I only feel should work, but I don't know for sure. However, see [1] for alternative approach :).

My point thus is, all individual steps are possible, and wiring them together seems pretty straightforward, therefore the whole thing should work if someone bothers to do it.

> If the technology was ready the evidence would be pretty clear- I'd expect to see some very good, very quickly made comic books shown off by AI enthusiast on reddit rather then the clearly limited/ not very good comic book experiments which have been demonstrated so far.

I think the biggest concentration of enthusiasm is to be found in NSWF uses of SD :). On the one hand, you're right; we probably should've seen it done already. On the other hand, my impression is that most people doing advanced SD magic are perfectly satisfied with partially manual workflows. And it kind of makes sense - manual steps allow for flexibility and experimentation, and some things are much simpler to wire by hand or patch up with some tactical photoshopping, than to try and automate them fully. In particular, things judging the quality of output is both easy for humans and hard to automate.

Still, I've recently seen ads of various AI apps claiming to do complex work (such as animating characters in photos) end-to-end automatically - exactly the kind of work that's typically done in partially manual process. So I suspect fully-automated solutions are being built on a case-by-case basis, driven by businesses making apps for the general population; a process that lags some months behind what image gen communities figure out in the open.

--

[0] - Though arguably, LLMs contain the procedural knowledge of how a task should be done; just ask it to ELI5 or explain in WikiHow style.

[1] - In fact, I just asked Claude to solve this problem in detail, without giving it my own solution to look at (but hinting at the required complexity level); see this: https://cloud.typingmind.com/share/db36fc29-6229-4127-8336-b... (and excuse the weird errors; Claude is overloaded at the moment, so some responses had to be regenerated; also styling on the shared conversation sucks, so be sure to use the "pop out" button on diagrams to see them in detail).

At very high level, it's the same as mine, but one level below, it uses different tools and approaches, some of which I never knew about - like keeping memory in embedding space instead of text space, and using various other models I didn't know exist.

EDIT: I did some quick web search for some of the ideas Claude proposed, and discovered even more techniques and models I never heard of. Even my own awareness of the image generation space is only scratching the surface of what people are doing.

In comparison I've messed around with prompting image generator models quite a bit and it's not possible to get remotely close to the quality level of even rough paid concept work by a professional, and the credits to run these models aren't particularly cheap.

I think people are getting way ahead of their skis here. Even in 2D I can’t for example generate inventory images for weapons and items for a game yet. Which is an orders of magnitude simpler test case than video. They all are slightly different styles. If I don’t care that they all look different in strange ways then it’s useful - but any consumer will think it looks like crap

AI or not, no one but you cares about the lower order bits of your idea.

GenAI is great at filling in those lower order bits but until stuff like ControlNet gets much better precision and UX, I think genAI will be stuck in the uncanny valley because they’re inconsistent between scenes, frames, etc.

Not disagreeing, just noting: this is not how [most?] people's minds work {I don't think you're holding to that opinion particularly, I'm just reflecting on this point}. We have vague ideas until an implementation is shown, then we examine it and latch on to a detail and decide if it matches our idea or not. For me, if I'm imagining "a superhero planting vegetables in his garden" I've no idea what they're actually wearing, but when an artist or genAI shows me it's a brown coat then I'll say "no something more marvel". Then when ultimately they show me something that matches the idea I had _and_ matches my current conception of the idea I had... then I'll point out the fingernails are too long, when in the idea I hadn't even perceived the person had fingers, never mind too-long fingernails!

I'd warrant any actualised artistic work has some delta with the artists current perception of the work; and a larger delta with their initial perception of it.

These are situations with relatively simple logical constraints, but an infinite number of valid solutions.

Keep in mind that we are not requiring any particular configuration of circuit diagram, just any diagram that makes sense. There are an infinite number of valid ones.

That doesn't mean it can't work with AI - it's that you may need to add something extra to the generative pipeline, something that can do circuit diagrams, and make the diffusion model supply style and extra noise (er, beautifying elements).

> Keep in mind that we are not requiring any particular configuration of circuit diagram, just any diagram that makes sense. There are an infinite number of valid ones.

On that note. I'm the kind of person that loves to freeze-frame movies to look at markings, labels, and computer screens, and one thing I learned is that humans fail at this task too. Most of the time the problems are big and obvious, ruining my suspension of disbelief, and importantly, they could be trivially solved if the producers grabbed a random STEM-interested intern and asked for advice. Alas, it seems they don't care.

This is just a specific instance of the general problem of "whatever you work with or are interested in, you'll see movies keep getting it wrong". Most of the time, it's somewhat defensible - e.g. most movies get guns wrong, but in way people are used to, and makes the scenes more streamlined and entertaining. But with labels, markings and computer screens, doing it right isn't any more expensive, nor would it make the movie any less entertaining. It seems that the people responsible don't know better or care.

Let's keep that in mind when comparing AI output to the "real deal", as to not set an impossible standards that human productions don't match, and never did.

In particular, internal consistency is one of the important constraints which viewers will immediately notice. If you’re just using sora for 5 second unrelated videos it may be less of an issue but if you want to do anything interesting you’ll need the clips to tie together which requires internal consistency.

Sora is obviously not Photoshop, but given that you can write basically anything you can think of I reckon it's going to take a long time to get good at expressing your vision in words that a model like Sora will understand.

FWIW I too have been quite frustrated iterating with AI to produce a vision that is clear in my head. Past changing the broad strokes, once you start “asking” for specifics, it all goes to shit.

Still, it’s good enough at those broad strokes. If you want your vision to become reality, you either need to learn how to paint (or whatever the medium), or hire a professional, both being tough-but-fair IMO.

Things like rearranging things in the scene with drag'n'drop sound implementable (although incredibly GPU heavy)

Also, AI sucks at understanding detail expressed in symbolic communication, because it doesn't understand symbols the way linguistic communication expects the receiver to understand them.

My own experience is that all the AI tools are great for shortcutting the first 70-80% or so. But the last 20% goes up an exponential curve of required detail which is easier and easier to express directly using tooling and my human brain.

Consider the analogy to a contract worker building or painting something for you. If all you have is a vague description, they'll make a good guess and you'll just have to live with that. But the more time you spend with them communicating (through description, mood boards rough sketches etc) the more accurate to your detailed version it will get. But you only REALLY get exactly what you want if you do it yourself, or sit beside them as they work and direct almost every step. And that last option is almost impossible if they can't understand symbolic meaning in language.

(Talking about visual generative AI in general)

I find generative AI frustrating because I know what I want. To this point I have been trying but then ultimately sitting it out — waiting for the one that really works the way I want.

If you want higher quality/precision, you’ll likely want to ask a professional, and I don’t expect that to change in the near future.

People pay for stock images.

But that’s probably something you don’t pay for directly, instead paying for e.g. a phone that has those features.

What happens is a description becomes a longer specification or script that's still good and hangs together in itself and then further iterations involving professionals who can't do "exactly what the director wants" but rather do something further that's good and close enough to what the director wants.

Eventually, it starts to muck with the earlier work that it did good on, when I'm just asking it to add onto it.

I was still happy with what I got in the end, but it took trial and error and then a lot of piecemeal coaxing with verification that it didn't do more than I asked along the way.

I can imagine the same for video or images. You have to examine each step post prompt to verify it didn't go back and muck with the already good parts.

With ChatGPT, you can iteratively improve text (e.g., "make it shorter," "mention xyz"). However, for pictures (and video), this functionality is not yet available. If you could prompt iteratively (e.g., "generate a red car in the sunset," "make it a muscle car," "place it on a hill," "show it from the side so the sun shines through the windshield"), the tools would become exponentially more useful.

https://www.youtube.com/watch?v=uNv7XBngmLY&t=25

Sure, if you then do the same in reverse.