So, this is surprising. Apparently there’s more cause, effect, and sequencing in diffusion models than what I expected, which would be roughly ‘none’. Google here uses SD 1.4, as the core of the diffusion model, which is a nice reminder that open models are useful to even giant cloud monopolies.
The two main things of note I took away from the summary were: 1) they got infinite training data using agents playing doom (makes sense), and 2) they added Gaussian noise to source frames and rewarded the agent for ‘correcting’ sequential frames back, and said this was critical to get long range stable ‘rendering’ out of the model.
That last is intriguing — they explain the intuition as teaching the model to do error correction / guide it to be stable.
Finally, I wonder if this model would be easy to fine tune for ‘photo realistic’ / ray traced restyling — I’d be super curious to see how hard it would be to get a ‘nicer’ rendering out of this model, treating it as a doom foundation model of sorts.
Just want to clarify a couple possible misconceptions:
The diffusion model doesn’t maintain any state itself, though its weights may encode some notion of cause/effect. It just renders one frame at a time (after all it’s a text to image model, not text to video). Instead of text, the previous states and frames are provided as inputs to the model to predict the next frame.
Noise is added to the previous frames before being passed into the SD model, so the RL agents were not involved with “correcting” it.
De-noising objectives are widespread in ML, intuitively it forces a predictive model to leverage context, ie surrounding frames/words/etc.
In this case it helps prevent auto-regressive drift due to the accumulation of small errors from the randomness inherent in generative diffusion models. Figure 4 shows such drift happening when a player is standing still.
The concept is that if you train a Diffusion model by feeding all the possible frames seen in the game.
The training was over almost 1 billion frames, 20 days of full-time play-time, taking a screenshot of every single inch of the map.
Now you show him N frames as input, and ask it "give me frame N+1", then it gives you the frame n. N+1 back based on how it was originally seen during training.
But it is not frame N+1 from a mysterious intelligence, it's simply frame N+1 given back from past database.
The drift you mentioned is actually a clear (but sad) proof that the model does not work at inventing new frames, and can only spit out an answer from the past dataset.
It's a bit like if you train stable diffusion on Simpsons episodes, and that it outputs the next frame of an existing episode that was in the training set, but few frames later goes wild and buggy.
I don't think you've understood the project completely. The model accepts player input, so frame 601 could be quite different if the player decided to turn left rather than right, or chose that moment to fire at an exploding barrel.
> 1 billion frames in memory... With such dataset, you have seen practically all realistic possibilities in the short-term.
I mean... no? Not even close? Multiply the number of game states with the number of inputs at any given frame gives you a number vastly bigger than 1 billion, not even comparable. Even with 20 days of play time to train no, it's entirely likely that at no point did someone stop at a certain location and look to the left from that angle. They might have done from similar angles, but the model then has to reconstruct some sense of the geometry of the level to synthesize the frame. They might also not have arrived there from the same direction, which again the model needs some smarts to understand.
I get your point, it's very overtrained on these particular levels of Doom, which means you might as well just play Doom. But this is not a hash table lookup we're talking about, it's pretty impressive work.
The map 1 has 2'518 walkable map units.
There are 65536 angles.
2'518*65'536=165'019'648
If you capture 165M frames, you already cover all the possibilities in terms of camera / player view, but probably the diffusion models don't even need to have all the frames (the same way that LLMs don't).
Like many people in case of LLMs, you're just demonstrating unawareness of - or disbelief in - the fact that the model doesn't record training data vetbatim, but smears it out in high-dimensional space, from which it then samples. The model then doesn't recall past inputs (which are effectively under extreme lossy compression), but samples from that high-dimensional space to produce output. The high-dimensional representation by necessity captures semantic understanding of the training data.
Generating "infinite Doom" is exactly what this model is doing, as it does not capture the larger map layout well enough to stay consistent with it.
But it's not a game. It's a memory of a game video, predicting the next frame based on the few previous frames, like "I can imagine what happened next".
I would call it the world's least efficient video compression.
What I would like to see is the actual predictive strength, aka imagination, which I did not notice mentioned in the abstract. The model is trained on a set of classic maps. What would it do, given a few frames of gameplay on an unfamiliar map as input? How well could it imagine what happens next?
But it's trained on the actual screen pixel data, AFAICT. It's literally a visual imagination model, not gameplay / geometry imagination model. They had to make special provisions to the pixel data on the HUD which by its nature different than the pictures of a 3D world.
> But it's not a game. It's a memory of a game video, predicting the next frame based on the few previous frames, like "I can imagine what happened next".
It's not super clear from the landing page, but I think it's an engine? Like, its input is both previous images and input for the next frame.
So as a player, if you press "shoot", the diffusion engine need to output an image where the monster in front of you takes damage/dies.
We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality.
They could down convert the entire model to only utilize the subset of matrix components from stable diffusion. This approach may be able to improve internet bandwidth efficiency assuming consumers in the future have powerful enough computers.
No, it’s predicting the next frame conditioned on past frames AND player actions! This is clear from the article. Mere video generation would be nothing new.
It's a memory of a video looped to controls, so frame 1 is "I wonder how would it look if the player pressed D instead of W", then the frame 2 is based on frame 1, etc. and couple frames in, it's already not remembering, but imagining the gameplay on the fly. It's not prerecorded, it responds to inputs during generation. That's what makes it a game engine.
> Apparently there’s more cause, effect, and sequencing in diffusion models than what I expected
To temper this a bit, you may want to pay close attention to the demo videos. The player rarely backtracks, and for good reason - the few times the character does turn around and look back at something a second time, it has changed significantly (the most noticeable I think is the room with the grey wall and triangle sign).
This falls in line with how we'd expect a diffusion model to behave - it's trained on many billions of frames of gameplay, so it's very good at generating a plausible -next- frame of gameplay based on some previous frames. But it doesn't deeply understand logical gameplay constraints, like remembering level geometry.
Even purely going forward, specks on wall textures morph into opponents and so on. All the diffusion-generated videos I’ve seen so far have this kind of unsettling feature.
What if you combine this with an engine in parallel that provides all geometry including characters and objects with their respective behavior, recording changes made through interactions the other model generates, talking back to it?
A dialogue between two parties with different functionality so to speak.
“Everything” would mean all objects and the elements they’re made of, their rules on how they interact and decay.
A modularized ecosystem i guess, comprised of “sub-systems” of sorts.
The other model, that provides all interaction (cause for effect) could either be run artificially or be used interactively by a human - opening up the possibility for being a tree : )
This all would need an interfacing agent that in principle would be an engine simulating the second law of thermodynamics and at the same time recording every state that has changed and diverged off the driving actor’s vector in time.
Basically the “effects” model keeping track of everyones history.
In the end a system with an “everything” model (that can grow overtime), a “cause” model messing with it, brought together and documented by the “effect” model.
(Again … non technical person, just fantasizing) : )
For instance, for a generated real world RPG, one process could create the planet, one could create the city where the player starts, one could create the NPCs, one could then model the relationships of the npcs with each other. Each one building off of the other so that the whole thing feels nuanced and more real.
Repeat for quest lines, new cities, etc, with the npcs having real time dialogue and interactions that happen entirely off screen, no guarantee of there being a massive quest objective, and some sort of recorder of events that keeps a running tally of everything that goes on so that as the PCs interact with it they are never repeating the same dreary thing.
If this were a MMORPG it would require so much processing and architecting, but it would have the potential to be the greatest game in human history.
It always blew my mind how well it worked on a 33 Mhz 486. I'm fairly sure it ran at 30 fps in 320x200. That gives it just over 17 clock cycles per pixel, and that doesn't even include time for game logic.
My memory could be wrong, though, but even if it required a 66 Mhz to reach 30 fps, that's still only 34 clocks per pixel on an architecture that required multiple clocks for a simple integer add instruction.
Great observation. And not entirely unlike normal human visual perception which is notoriously vulnerable to missing highly salient information; I'm reminded of the "gorillas in our midst" work by Dan Simons and Christopher Chabris [0].
If a generic human glances at an unfamiliar screen/wall/room, can they accurately, pixel-perfectly reconstruct every single element of it? Can they do it for every single screen they have seen in their entire lives?
I never said pixel perfect, but I would be surprised if whole objects , like flaming lanterns suddenly appeared.
What this demo demonstrates to me is how incredible willing we are to accept what seems familiar to us as accurate.
I bet if you look closely and objectively you will see even more anomalies. But at first watch, I didn’t see most errors because I think accepting something is more efficient for the brain.
You'd likely be surprised by a flaming lantern unless you were in Flaming Lanterns 'R Us, but if you were watching a video of a card trick and the two participants changed clothes while the camera wasn't focused on them, you may well miss that and the other five changes that came with that.
Not exactly, but our representation of what's behind us is a lot more sparse than we would assume. That is, I might not be surprised by what I see when I turn around, but it could have changed pretty radically since I last looked, and I might not notice. In fact, an observer might be quite surprised that I missed the change.
Objectively, Simons and Chabris (and many others) have a lot of data to support these ideas. Subjectively, I can say that these types of tasks (inattentional blindness, change blindness, etc.) are humbling.
Even having a clue why I'm linking this, I virtually guarantee you won't catch everything.
And even if you do catch everything... the real thing to notice is that you had to look. Your brain does not flag these things naturally. Dreams are notorious for this sort of thing, but even in the waking world your model of the world is much less rich than you think. Magic tricks like to hide in this space, for instance.
The opposite - if you turn around and there's something that wasn't there the last time - you'll likely not notice if it's not out of place. You'll just assume it was there and you weren't paying attention.
We don't memorize things that the environment remembers for us if they aren't relevant for other reasons.
Not noticing to a gorilla that ‘shouldn’t’ be there is not the same thing as object permanence. Even quite young babies are surprised by objects that go missing.
That's absolutely true. It's also well-established by Simons et al. and others that healthy normal adults maintain only a very sparse visual representation of their surroundings, anchored but not perfectly predicted by attention, and this drives the unattended gorilla phenomenon (along with many others). I don't work in this domain, but I would suggest that object permanence probably starts with attending and perceiving an object, whereas the inattentional or change blindness phenomena mostly (but not exclusively) occur when an object is not attended (or only briefly attended) or attention is divided by some competing task.
The people were told to focus very deeply on a certain aspect of the scene. Maintaining that focus means explicitly blocking things not related to that focus. Also, there is social pressure at the end to have peformed well at the task; evaluating them on a task which is intentionally completely different than the one explicitly given is going to bias people away from reporting gorillas.
And also, "notice anything unusual" is a pretty vague prompt. No-one in the video thought the gorillas were unusual, so if the PEOPLE IN THE SCENE thought gorillas were normal, why would I think they were strange? Look at any TV show, they are all full of things which are pretty crazy unusual in normal life, yet not unusual in terms of the plot.
I understand what you mean. I believe that the authors would contend that what you're describing is a typical attentional state for an awake/aware human: focused mostly on one thing, and with surprisingly little awareness of most other things (until/unless they are in turn attended).
Furthermore, even what we attend to isn't always represented with all that much detail. Simons has a whole series of cool demonstration experiments where they show that they can swap out someone you're speaking with (an unfamiliar conversational partner like a store clerk or someone asking for directions), and you may not even notice [0]. It's rather eerie.
Does that work on autistic people? Having no filters or fewer filters, should allow them to be more efficient "on guard duty" looking for unexpected things.
tangentially related but Grand Theft Auto speedrunners often point the camera behind them while driving so cars don't spawn "behind" them (aka in front of the car)
I saw a longer video of this that Ethan Mollick posted and in that one, the sequences are longer and they do appear to demonstrate a fair amount of consistency. The clips don't backtrack in the summary video on the paper's home page because they're showing a number of district environments but you only get a few seconds of each.
If I studied the longer one more closely, I'm sure inconsistencies would be seen but it seemed able to recall presence/absence of destroyed items, dead monsters etc on subsequent loops around a central obstruction that completely obscured them for quite a while. This did seem pretty odd to me, as I expected it to match how you'd described it.
Yes it definitely is very good for simulating gameplay footage, don't get me wrong. Its input for predicting the next frame is not just the previous frame, it has access to a whole sequence of prior frames.
But to say the model is simulating actual gameplay (i.e. that a person could actually play Doom in this) is far fetched. It's definitely great that the model was able to remember that the gray wall was still there after we turned around, but it's untenable for actual gameplay that the wall completely changed location and orientation.
> it's untenable for actual gameplay that the wall completely changed location and orientation.
It would in an SCP-themed game. Or dreamscape/Inception themed one.
Hell, "you're trapped in Doom-like dreamscape, escape before you lose your mind" is a very interesting pitch for a game. Basically take this Doom thing and make walking though a specific, unique-looking doorway from the original game to be the victory condition - the player's job would be to coerce the model to generate it, while also not dying in the Doom fever dream game itself. I'd play the hell out of this.
(Implementation-wise, just loop in a simple recognition model to continously evaluate victory condiiton from last few frames, and some OCR to detect when player's hit points indicator on the HUD drops to zero.)
(I'll happily pay $100 this year to the first project that gets this to work. I bet I'm not the only one. Doesn't have to be Doom specifically, just has to be interesting.)
To be honest, I agree! That would be an interesting gameplay concept for sure.
Mainly just wanted to temper expectations I'm seeing throughout this thread that the model is actually simulating Doom. I don't know what will be required to get from here to there, but we're definitely not there yet.
Fun variant: give it hidden state by doing the offscreen scratch pixel buffer thing, but not grading its content in training. Train the model as before, grading on the "onscreen" output, and let it keep the side channel to do what it wants with. It'd be interesting to see what way it would use it, what data it would store, and how it would be encoded.
MyHouse is excellent, but it mostly breaks our perception of what the Doom engine is capable of by not really using the Doom engine. It leans heavily on engine features which were embellishments by the GZDoom project, and never existed in the original Doom codebase.
Small objects like powerups appear and disappear as the player moves (even without backtracking), the ammo count is constantly varying, getting shot doesn't deplete health or armor, etc.
So for the next iteration, they should add a minimap overlay (perhaps on a side channel) - it should help the model give more consistent output in any given location. Right now, the game is very much like a lucid dream - the universe makes sense from moment to moment, but without outside reference, everything that falls out of short-term memory (few frames here) gets reimagined.
I don't see this as something that would be hard to overcome. Sora for instance has already shown the ability for a diffusion model to maintain object permanence. Flux recently too has shown the ability to render the same person in many different poses or images.
I don't know of an example (not to say it doesn't exist) but the problem is fundamentally the same as things moving out of sight/out of frame and coming back again.
> the problem is fundamentally the same as things moving out of sight/out of frame and coming back again
Maybe it is, but doing that with the entire scene instead of just a small part of it makes the problem massively harder, as the model needs to grow exponentially to remember more things. It isn't something that we will manage anytime soon, maybe 10-20 years with current architecture and same compute progress.
Then you make that even harder by remembering a whole game level? No, ain't gonna happen in our lifetimes without massive changes to the architecture. They would need to make a different model keep track of level state etc, not just an image to image model.
In this sora video the dragon covers half the scene, and its basically identical when it is revealed again ~5 seconds later, or about 150 frames later. The is lots of evidence (and some studies) that these models are in fact building internal world models.
Buckle in, the train is moving way faster. I don't think there would be much surprise if this is solved in the next few generations of video generators. The first generation is already doing very well.
Did you watch the video, it is completely different after the dragon goes past? Its still a flag there, but everything else changed. Even the stores in the background changed, the mass of people is completely different with no hint of anyone moving there etc.
You always get this from AI enthusiast, they come and post "proof" that disproves their own point.
I'm not GP, but running over that video I'm actually having a hard time finding any detail present before the dragon obscures them not either exit frame right when the camera pans left slightly near the end or not re-appear with reasonably crisp detail after the dragon gets out of the way.
Most of the mob of people are indistinct, but there is a woman in a lime green coat who is visible, and then obstructed by the dragon twice (beard and ribbon) and reappears fine. Unfortunately when dragon fully moves past she has been lost to frame right.
There is another person in black holding a red satchel which is visible both before and after the dragon has passed.
Nothing about the storefronts appear to change. The complex sign full of Chinese text (which might be gibberish text: it's highly stylized and I don't know Chinese) appears to survive the dragon passing without even any changes to the individual ideograms.
There is also a red box shaped like a Chinese paper lantern with a single gold ideogram on it at the store entrance which spends most of the video obscured by the dragon and is still in the same location after it passes (though video artifacting makes it more challenging to verify that that ideogram is unchanged it certainly does not appear substantially different)
What detail are you seeing that is different before and after the obstruction?
> What detail are you seeing that is different before and after the obstruction?
First frame, guy in blue hat next to a flag. That flag and the guy is then gone afterwards.
The two flags near the wall are gone, there is something triangular there but there was two flags before the dragon went past.
Then not to mention that the crowd is 6 people deep after the dragon went past, while just 4 people deep before, it is way more crowded.
Instead of the flag that was there before the dragon, it put in 2 more flags afterwards far more to the left.
Third second a guy was out of frame for a few frames, and suddenly gained a blue scarf. AFter dragon went by he turned into a woman. Next to that person was a guy with a blue cap, he completely disappears.
> Most of the mob of people are indistinct
No they aren't, they are mostly distinct and basically all of them changes. If you ignore that the entire mob totally changes both in number and appearance and where it is, sure it is pretty good, except it forgot the flags, but how can you ignore the mob when we talk about the model remembering details? The wall is much less information dense than the mob, so that is much easier to remember for the model, the difficulty is in the mob.
> but there is a woman in a lime green coat who is visible,
She was just out of frame for a fraction of a second, not the big bit where the dragon moves past. The guy in blue jacket and blue cap behind her disappears though, or merges with another person and becomes a woman with a muffler after the dragon moved past.
So, in the end some big strokes were kept, and that was a very tiny part of the image that was both there before and after the dragon moved past so it was far from a whole image with full details. Almost all details are wrong.
Maybe he meant that the house looked mostly the same, I agree the upper parts does, but I looked at the windows and they were completely different, it is full of people heads after the dragon moved past while before it was just clean walls.
We are looking at first generation tech and pretty much every human would recognize the "before dragon" scene as being the same place as the "after dragon" scene. The prominent features are present. The model clearly shows the ability to go beyond "image-to-image" rendering.
If you want to be right because you can find any difference. Sure. You win. But also completely missed the point.
> pretty much every human would recognize the "before dragon" scene as being the same place as the "after dragon" scene
Not in a game and those were enemies, it completely changed what and how many they are, people would notice such a massive change instantly if they looked away and suddenly there were 50% more enemies.
> The model clearly shows the ability to go beyond "image-to-image" rendering.
I never argued against that. Adding a third dimension (time) makes generating a video the same kind of problem as generating an image, it is not harder to draw a straight pencil with something covering it than to draw the scene with something covering it for a while.
But still, even though it is that simple, these models are really bad at it, because it requires very large models and much compute. So I just extrapolated based on their current abilities that we know, as you demonstrated there, to say roughly how long until we can even have consistent short videos.
Note that videos wont have the same progression as images, as the early image models were very small and we quickly scaled up there, while now for video we start at really scaled up models and we have to wait until compute gets cheaper/faster the slow way.
> But also completely missed the point.
You completely missed my point or you changed your point afterwards. My point was that current models can only remember little bits under such circumstances, and to remember a whole scene they need to be massively larger. Almost all details in the scene you showed were missed, the large strokes are there but to keep the details around you need an exponentially larger model.
Nicely summarised. Another important thing that clearly standsout (not to undermine the efforts and work gone into this) is the fact that more and more we are now seeing larger and more complex building blocks emerging (first it was embedding models then encoder decoder layers and now whole models are being duck-taped for even powerful pipelines). AI/DL ecosystem is growing on a nice trajectory.
Though I wonder if 10 years down the line folks wouldn't even care about underlying model details (no more than a current day web-developer needs to know about network packets).
PS: Not great examples, but I hope you get the idea ;)
> Google here uses SD 1.4, as the core of the diffusion model, which is a nice reminder that open models are useful to even giant cloud monopolies.
A mistake people make all the time is that massive companies will put all their resources toward every project. This paper was written by four co-authors. They probably got a good amount of resources, but they still had to share in the pool allocated to their research department.
Even Google only has one Gemini (in a few versions).
The quest to run doom on everything continues.
Technically speaking, isn't this the greatest possible anti-Doom, the Doom with the highest possible hardware requirement?
I just find it funny that on a linear scale of hardware specification, Doom now finds itself on both ends.
>Technically speaking, isn't this the greatest possible anti-Doom
When I read this part I thought you were going to say because you're technically not running Doom at all. That is, instead of running Doom without Doom's original hardware/software environment (by porting it), you're running Doom without Doom itself.
- you could build a non-real-time version of the game engine and use the neural net as a real-time approximation
- you could edit videos shot in real life to have huds or whatever and train the neural net to simulate reality rather than doom. (this paper used 900 million frames which i think is about a year of video if it's 30fps, but maybe algorithmic improvements can cut the training requirements down) and a year of video isn't actually all that much—like, maybe you could recruit 500 people to play paintball while wearing gopro cameras with accelerometers and gyros on their heads and paintball guns, so that you could get a year of video in a weekend?
That feels like the endgame of video game generation.
You select an art style, a video and the type of game you'd like to play.
The game is then generated in real-time responding to each action with respect to the existing rule engine.
I imagine a game like that could get so convincing in its details and immersiveness that one could forget they're playing a game.
Well, 2001 is actually a happy ending, as Dave is reborn as a cosmic being. Solaris, at least in the book, is an attempt by the sentient ocean to communicate with researchers through mimics.
Have you ever played a video game? This is unbelievably depressing. This is a future where games like Slay the Spire, with a unique art style and innovative gameplay simply are not being made.
Not to mention this childish nonsense about "forget they're playing a game," as if every game needs to be lifelike VR and there's no room for stylization or imagination. I am worried for the future that people think they want these things.
Its a good thing. When the printing press was invented there were probably monks and scribes who thought that this new mechanical monster that took all the individual flourish out of reading was the end of literature. Instead it became a tool to make literature better and just removed a lot of drudgery. Games with individual style and design made by people will of course still exist. They'll just be easier to make.
The problem is quite the opposite, that AI will be able to generate games so many game with so many play styles that it will totally dilute the value of all games.
Compare it to music gen algo's that can now produce music that is 100% indiscernible from generic crappy music. Which is insane given that 5 years ago it could maybe create the sound of something that maybe someone would describe as "sort of guitar-like". At this rate of progress it's probably not going to be long before AI is making better music than humans. And it's infinitely available too.
Oh, because we can link this in with biometric responses - heartrate, temperature, eye tracking etc.
We could build a 'game' which would learn and adapt to precisely the chemistry that makes someone tick and then provide them a map to find the state in which their brain releases their desired state.
Then if the game has a directive - it should be pointed to work as a training tool to allow the user to determine how to release these chemicals themselves at will. Resulting in a player-base which no longer requires anything external for accessing their own desired states.
Even if you spend 40 hours a week in video conferences, you'll have to work for over four years to get one years' worth of footage. Of course, by then the models will be even better and so you might actually have a chance of going surfing.
I guess I should start hoarding video of myself now.
the neural net doesn't need a year of video to train to simulate your face; it can do that from a single photo. the year of video is to learn how to play the game, and in most cases lots of people are playing the same game, so you can dump all their video in the same training set
If you train it on multiple games then you could produce new games that have never existed before, in the same way image generation models can produce new images that have never existed before.
It's unlikely that such a procedurally generated mashup would be perfectly coherent, stable and most importantly fun right out of the gate, so you would need some way to reach into the guts of the generated game and refine it. If properties as simple as "how much health this enemy type has" are scattered across an enormous inscrutable neural network, and may not even have a single consistent definition in all contexts, that's going to be quite a challenge. Nevermind if the game just catastrophically implodes and you have to "debug" the model.
Well, yeah. Image diffusion models only work because you can provide large amounts of training data. For Doom it is even simpler, since you don't need to deal with compositing.
Maybe, in future, techniques of Scientific Machine Learning which can encode physics and other known laws into a model would form a base model. And then other models on top could just fine tune aspects to customise a game.
There is going to be a flood of these dreamlike "games" in the next few years. This feels likes a bit of a breakthrough in the engineering of these systems.
It's insane that that this works, and that it works fast enough to render at 20 fps. It seems like they almost made a cross between a diffusion model and an RNN, since they had to encode the previous frames and actions and feed it into the model at each step.
Abstractly, it's like the model is dreaming of a game that it played a lot of, and real time inputs just change the state of the dream. It makes me wonder if humans are just next moment prediction machines, with just a little bit more memory built in.
Also recursion and nested virtualization. We can dream about dreaming and imagine different scenarios, some completely fictional or simply possible future scenarios all while doing day to day stuff.
It makes good sense for humans to have this ability. If we flip the argument, and see the next frame as a hypothesis for what is expected as the outcome of the current frame, then comparing this "hypothesis" with what is sensed makes it easier to process the differences, rather than the totality of the sensory input.
As Richard Dawkins recently put it in a podcast[1], our genes are great prediction machines, as their continued survival rests on it. Being able to generate a visual prediction fits perfectly with the amount of resources we dedicate to sight.
If that is the case, what does aphantasia tell us?
What’s the aphantasia link? I’ve got aphantasia. I’m convinced though that the bit of my brain that should be making images is used for letting me ‘see’ how things are connected together very easily in my head. Also I still love games like Pictionary and can somehow draw things onto paper than I don’t really know what they look like in my head. It’s often a surprise when pen meets paper.
I agree, it is my own experience as well. Craig Venter In one of his books also credit this way of representing knowledge as abstractions as his strength in inventing new concepts.
The link may be that we actually see differences between “frames”, rather than the frames directly. That in itself would imply that a from of sub-visual representation is being processed by our brain. For aphantasia, it could be that we work directly on this representation instead of recalling imagery through the visual system.
Many people with aphantasia reports being able to visualize in their dreams, meaning that they don't lack the ability to generate visuals. So it may be that the brain has an affinity to rely on the abstract representation when "thinking", while dreaming still uses the "stable diffusion mode".
I’m no where near qualified to speak of this with certainty, but it seems plausible to me.
Worth noting that aphantasia doesn't necessarily extend to dreams. Anecdotally - I have pretty severe aphantasia (I can conjure milisecond glimpses of barely tangible imagery that I can't quite perceive before it's gone - but only since learning that visualisation wasn't a linguistic metaphor). I can't really simulate object rotation. I can't really 'picture' how things will look before they're drawn / built etc. However I often have highly vivid dream imagery. I also have excellent recognition of faces and places (e.g.: can't get lost in a new city). So there clearly is a lot of preconscious visualisation and image matching going on in some aphantasia cases, even where the explicit visual screen is all but absent.
Pretty much the same for me. My aphantasia is total (no images at all) but still ludicrously vivid dreams and not too bad at recognising people and places.
> Many people with aphantasia reports being able to visualize in their dreams, meaning that they don't lack the ability to generate visuals. So it may be that the [aphantasia] brain has an affinity to rely on the abstract representation when "thinking", while dreaming still uses the "stable diffusion mode".
(I obviously don't know what I'm talking about, just a fellow aphant)
Obviously we're all introspecting here - but my guess is that there's some kind of cross talk in aphantasic brains between the conscious narrating semantic brain and the visual module. Such that default mode visualisation is impaired. It's specifically the loss of reflexive consciousness that allows visuals to emerge. Not sure if this is related, but I have pretty severe chronic insomnia, and I often wonder if this in part relates to the inability to drift off into imagery.
Image is 2D. Video is 3D. The mathematical extension is obvious. In this case, low resolution 2D (pixels), and the third dimension is just frame rate (discrete steps). So rather simple.
This is not "just" video, however. It's interactive in real time. Sure, you can say that playing is simply video with some extra parameters thrown in to encode player input, but still.
I think you're mistaken. The abstract says it's interactive, "We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction"
Further - "a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions." specifically "and actions"
User input is being fed into this system and subsequent frames take that into account. The user is "actually" firing a gun.
I guess you are being sarcastic, except this is precisely what it is doing. And it's not hard: player movement is low information and probably not the hardest part of the model.
It's interactive but can it go beyond what it learned from the videos. As in, can the camera break free and roam around the map from different angles? I don't think it will be able to do that at all. There are still a few hallucinations in this rendering, it doesn't look it understands 3d.
You might be surprised. Generating views from novel angles based on a single image is not novel, and if anything, this model has more than a single frame as input. I’d wager that it’s quite able to extrapolate DOOM-like corridors and rooms even if it hasn’t seen the exact place during training. And sure, it’s imperfect but on the other hand it works in real time on a single TPU.
Then why do monsters become blurry smudgy messes when shot? That looks like a video compression artifact of a neural network attempting to replicate low-structure image (source material contains guts exploding, very un-structured visual).
Uh, maybe because monster death animations make up a small part of the training material (ie. gameplay) so the model has not learned to reproduce them very well?
There cannot be "video compression artifacts" because it hasn’t even seen any compressed video during training, as far as I can see.
Seriously, how is this even a discussion? The article is clear that the novel thing is that this is real-time frame generation conditioned on the previous frame(s) AND player actions. Just generating video would be nothing new.
In a sense, poorly reproducing rare content is a form of compression artifact. Ie, since this content occurs rarely in the training set, it will have less impact on the gradients and thus less impact on the final form of the model. Roughly speaking, the model is allocating fewer bits to this content, by storing less information about this content in its parameters, compared to content which it sees more often during training. I think this isn't too different from certain aspects of images, videos, music, etc., being distorted in different ways based on how a particular codec allocates its available bits.
I simply cannot take seriously anyone who exclaims that monster death animations are a minor part of Doom. It's literally a game about slaying demons. Gameplay consists almost entirely of explosions and gore, killing monsters IS THE GAME, if you can't even get that correct then what nonsense are we even looking at.
Okay, I think you're right. My mistake. I read through the paper more closely and I found the abstract to be a bit misleading compared to the contents. Sorry.
Academic authors are consistently better at editing away unclear and ambiguous statements which make their work seem less impressive compared to ones which make their work seem more impressive. Maybe it's just a coincidence, lol.
Uff, I guess you’re right. Mea culpa. I misread their diagram to represent inference when it was about training instead. The latter is conditioned on actions, but… how do they generate the actual output frames then? What’s the input? Is it just image-to-image based on the previous frame? The paper doesn’t seem to explain the inference part at all well :(
Video is also higher resolution, as the pixels flip for the high resolution world by moving through it. Swivelling your head without glasses, even the blurry world contains more information in the curve of pixelchange.
Correct, for the sprites. However, the walls in Doom are texture mapped, and so have the same issue as videos. Interesting, though, because I assume the antialiasing is something approximate, given the extreme demands on CPUs of the era.
So are gravity and friction. I don't know how well tested or accepted it is, but being just a theory doesn't tell you much about how true it is without more info
We are. At least that's what Lisa Feldman Barrett [1] thinks. It is worth listening to this Lex Fridman podcast: Counterintuitive Ideas About How the Brain Works [2], where she explains among other ideas how constant prediction is the most efficient way of running a brain as opposed to reaction. I never get tired of listening to her, she's such a great science communicator.
Interesting talk about the brain, but the stuff she says about free will is not a very good argument. Basically it is sort of the argument that the ancient greeks made which brings the discussion into a point where you can take both directions.
Penrose (Nobel prize in physics) stipulates that quantum effects in the brain may allow a certain amount of time travel and back propagation to accomplish this.
It's unclear how that compares to a high-end consumer GPU like a 3090, but they seem to have similar INT8 TFLOPS. The TPU has less memory (16 vs. 24), and I'm unsure of the other specs.
Something doesn't add up, in my opinion, though. SD usually takes (at minimum) seconds to produce a high-quality result on a 3090, so I can't comprehend how they are like 2 orders of magnitudes faster—indicating that the TPU vastly outperforms a GPU for this task. They seem to be producing low-res (320x240) images, but it still seems too fast.
There's been a lot of work in optimising inference speed of SD - SD Turbo, latent consistency models, Hyper-SD, etc. It is very possible to hit these frame rates now.
There's been a ton of work to generate assets for games using AI: 3d models, textures, code, etc. None of that may even be necessary with a generative game engine like this! If you could scale this up, train on all games in existence, etc. I bet some interesting things would happen
But can you grab what this Ai has learned and generate the 3d models, maps and code to turn it into an actual game that can run on a user's PC? That would be amazing.
What would be the point? This model has been trained on an existing game, so turning it back into assets, maps, and code would just give you a copy of the original game you started with. I suppose you could create variations of it then... but:
You don't even need to do all of that - this trained model already is the game, i.e., it's interactive, you can play the game.
There is no text conditioning provided to the SD model because they removed it, but one can imagine a near future where text prompts are enough to create a fun new game!
Yes they had to use RL to learn what DOOM looks like and how it works, but this doesn’t necessarily pose a chicken vs egg problem. In the same way that LLMs can write a novel story, despite only being trained on existing text.
IMO one of the biggest challenges with this approach will be open world games with essentially an infinite number of possible states. The paper mentions that they had trouble getting RL agents to completely explore every nook and corner of DOOM. Factorio or Dwarf Fortress probably won’t be simulated anytime soon…I think.
With enough computation, your neural net weights would converge to some very compressed latent representation of the source code of DOOM. Maybe smaller even than the source code itself? Someone in the field could probably correct me on that.
At which point, you effectively would be interpolating in latent space through the source code to actually "render" the game. You'd have an entire latent space computer, with an engine, assets, textures, a software renderer.
With a sufficiently powerful computer, one could imagine what interpolating in this latent space between, say Factorio and TF2 (2 of my favorites). And tweaking this latent space to your liking by conditioning it on any number of gameplay aspects.
This future comes very quickly for subsets of the pipeline, like the very end stage of rendering -- DLSS is already in production, for example. Maybe Nvidia's revenue wraps back to gaming once again, as we all become bolted into a neural metaverse.
Obviously assets would get encoded too, in some form. Not necessarily corresponding to the original bitmaps, if the game does some consistent post-processing, the encoded thing would more likely be (equivalent to) the post-processed state.
That’s just an artifact of the language we use to describe an implementation detail, in the sense GP means it, the data payload bits are not essentially distinct from the executable instruction bits
The Holographic Principle is the idea that our universe is a projection of a higher dimensional space, which sounds an awful lot like the total simulation of an interactive environment, encoded in the parameter space of a neural network.
The first thing I thought when I saw this was: couldn't my immediate experience be exactly the same thing? Including the illusion of a separate main character to whom events are occurring?
> With enough computation, your neural net weights would converge to some very compressed latent representation of the source code of DOOM. Maybe smaller even than the source code itself? Someone in the field could probably correct me on that.
Neural nets are not guaranteed to converge to anything even remotely optimal, so no that isn't how it works. Also even though neural nets can approximate any function they usually can't do it in a time or space efficient manner, resulting in much larger programs than the human written code.
Could is certainly a better word, yes. There is no guarantee that it will happen, only that it could. The existence of LLMs is proof of that; imagine how large and inefficient a handwritten computer program to generate the next token would be. On the flipside, human beings very effectively predicting the next token, and much more, on 5 watts is proof that LLM in their current form certainly are not the most efficient method for generating next token.
I don't really know why everyone is piling on me here. Sorry for a bit of fun speculating! This model is on the continuum. There is a latent representation of Doom in weights. some weights, not these weights. Therefore some representation of doom in a neural net could become more efficient over time. That's really the point I'm trying to make.
The raw capacity of the network doesn't tell you how complex the weights actually are. The capacity is only an upper bound on the complexity.
It's easy to see this by noting that you can often prune networks quite a bit without any loss in performance. I.e. the effective dimension of the manifold the weights live on can be much, much smaller than the total capacity allows for. In fact, good regularization is exactly that which encourages the model itself to be compressible.
I think your confusing capacity with the training dynamics.
Capacity is autological. The amount of information it can express.
Training dynamics are the way the model learns, the optimization process, etc. So this is where things like regularization come into play.
There's also architecture which affects the training dynamics as well as model capacity. Which makes no guarantee that you get the most information dense representation.
Sorry I wasn't more clear! I'm referring to the Kolmogorov complexity of the network. The OP said:
> With enough computation, your neural net weights would converge to some very compressed latent representation of the source code of DOOM. Maybe smaller even than the source code itself? Someone in the field could probably correct me on that.
And they're not wrong! An ideally trained network could, in principle, learn the data-generating program, if that program is within its class of representable functions. I might have a NN that naively looks like it takes up GBs of space, but it might actually be parameterizing a much simpler function (hence our ability to prune/compress the weights without performance loss - most of the capacity wasn't being used for any interesting computation).
You're right that there's no guarantee that the model finds the most "dense" representation. The goal of regularization is to encourage that, though!
All over the place in ML there are bounds like:
test loss <= train loss + model complexity
Hence minimizing model complexity improves generalization performance. This is a kind of Occam's Razor: the simplest model generalizes best. So the OP is on the right track - we definitely want networks to learn the "underlying" process that explains the data, which in this case would be a latent representation of the source code (well, except that doesn't really make sense since you'd need the whole rest of the compute stack that code runs on - the neural net has no external resources/embodied complexity it calls, unlike the source code which gets to rely on drivers, hardware, operating systems, etc.)
> An ideally trained network could, in principle, learn the data-generating program
No disagreement
> I might have a NN that naively looks like it takes up GBs of space, but it might actually be parameterizing a much simpler function (hence our ability to prune/compress the weights without performance loss - most of the capacity wasn't being used for any interesting computation).
Also no disagreement.
I suggested that this probably isn't the case here since they tried distillation and saw no effect. While this isn't proof that this particular model can't be compressed more it does suggest that it is non-trivial. This is especially true given the huge difference in size. I mean we're talking about 700x...
Where I think our disagreement is in that I read the OP as saying __this__ network. If we're talking about a theoretical network, well... nothing I said anywhere is in any disagreement with that. I even said in the post I linked to that the difference shows that there's still a long way to go but that this is still cool. Why did I assume OP was talking about __this__ network? Well because we're in a thread talking about a paper and well... yes, we're talking about compression machines so theoretically (well not actually supported by any math theory) this is true for so many things and that is a bit elementary. So makes more sense (imo) that we're talking about this network. And I wanted to make it clear that this network is nowhere near compression. Can further research later result in something that is better than the source code? Who knows? For all the reasons we've both mentioned. We know they are universal approximators (which are not universal mimicers and have limits) but we have no guarantee of global convergence (let alone proof such a thing exists in many problems).
And I'm not sure why you're trying to explain the basic concepts to me. I mentioned I was an ML researcher. I see you're a PhD at Oxford. I'm sure you would be annoyed if I was doing the same to you. We can talk at a different level.
Totally fair points all. Sorry if it came across as condescending!
I agree with you that this network probably has not found the source code or something like a minimal description in its weights.
Honestly, I'm writing a paper on model compression/complexity right now, so I may have co-opted the discussion to practice talking about these things...! Just a bit over-eager (,,>﹏<,,)
Have you given much thought to how we can encourage models to be more compressible? I'd love to be able to explicitly penalize the filesize during training, but in some usefully learnable way. Proxies like weight norm penalties have problems in the limit.
Haha totally fair and it happens to me too, but trying to work on it.
I actually have some stuff I'm working on in that area that is having some success. I do need to extend it to diffusion but I see nothing stopping me.
Personally I think a major slowdown for our community is it's avoidance of math. Like you don't need to have tons of math in the papers, but many of the lessons you learn in the higher level topics do translate to usable techniques in ML. Though I would also like to see a stronger push on theory because empirical results can be deceiving (Von Neumann's elephant and all)
Similarly, you could run a very very simple game engine, that outputs little more than a low resolution wireframe, and upscale it. Put all of the effort into game mechanics and none into visual quality.
I would expect something in this realm to be a little better at not being visually inconsistent when you look away and look back. A red monster turning into a blue friendly etc.
Most games are conditioned on text, it's just that we call it "source code" :).
(Jk of course I know what you mean, but you can seriously see text prompts as compressed forms of programming that leverage the model's prior knowledge)
If it can be trained on (many) existing games, then it might work similarly to how you don't need to describe every possible detail of a generated image in order to get something that looks like what you're asking for (and looks like a plausible image for the underspecified parts).
Also: define "fun" and "new" in a "simple text prompt". Current image generators suck at properly reflecting what you want exactly, because they regurgitate existing things and styles.
Video games are gonna be wild in the near future. You could have one person talking to a model producing something that's on par with a AAA title from today. Imagine the 2d sidescroller boom on Steam but with immersive photorealistic 3d games with hyper-realistic physics (water flow, fire that spreads, tornados) and full deformability and buildability because the model is pretrained with real world videos. Your game is just a "style" that tweaks some priors on look, settings, and story.
Sorry, no offence, but you sound like those EA execs wearing expensive suits and never played a single video game in their entire life. There’s a great documentary on how Half Life was made. Gabe Newell was interviewed by someone asking “why you did that and this, it’s not realistic”, where he answered “because it’s more fun this way, you want realism — just go outside”.
From the article: We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality.
They obviously can within seconds, so it wouldn't be a result. Being able to generate gameplay that looks right even if it doesn't play right is one step.
Take a bunch of videos of the real world and calculate the differential camera motion with optical flow or feature tracking. Call this the video’s control input. Now we can play SORA.
Watch closely in the videos and you'll see that enemies often respawn when offscreen and sometimes when onscreen. Destroyed barrels come back, ammo count and health fluctuates weirdly, etc. It's still impressive, but its not perfect in that regard.
It doesn't. You need to put the world state in the input (the "prompt", even it doesn't look like prompt in this case). Whatever not in the prompt is lost.
It doesn't even remember the state of the game you look at. Doors spawning right in front of you, particle effects turning into enemies mid flight etc, so just regular gen AI issues.
Edit: Can see this in the first 10 seconds of the first video under "Full Gameplay Videos", stairs turning to corridor turning to closed door for no reason without looking away.
There's also the case in the video (0:59) where the player jumps into the poison but doesn't take damage for a few seconds then takes two doses back-to-back - they should've taken a hit of damage every ~500-1000ms(?)
Guessing the model hasn't been taught enough about that, because most people don't jump into hazards.
Ah finally we are starting to see something gaming related. I'm curious as to why we haven't seen more of neural networks applied to games even in a completely experimental fashion; we used to have a lot of little experimental indie games such as Façade (2005) and I'm surprised we don't have something similar years after the advent of LLMs.
We could have mods for old games that generate voices for the characters for example. Maybe it's unfeasible from a computing perspective? There are people running local LLMs, no?
What most programmers don't understand, that in the very near future, the entire application will be delivered by an AI model, no source, no text, just connect to the app over RDP. The whole app will be created by example, the app developer will train the app like a dog trainer trains a dog.
that might work for some applications, especially recreational things, I think we're a while away from it doing away with all things, especially where deterministic behavior, efficiency, or reliability are important.
I think it's possible AI models will generate dynamic UI for each client and stream the UI to clients (maybe eventually client devices will generate their UI on the fly) similar to Google Stadia. Maybe some offset of video that allows the remote to control it. Maybe Wasm based - just stream wasm bytecode around? The guy behind VLC is building a library for ulta low latency: https://www.kyber.video/techology.
There has definitely been research for simulating physics based on observation, especially in fluid dynamics but also for rigid body motion and collision. It's important for robotics applications actually. You can bet people will be applying this technique in those contexts.
I think for real world application one challenge is going to be the "action" signal which is a necessary component of the conditioning signal that makes the simulation reactive. In video games you can just record the buttons, but for real world scenarios you need difficult and intrusive sensor setups for recording force signals.
(Again for robotics though maybe it's enough to record the motor commands, just that you can't easily record the "motor commands" for humans, for example)
It's called predictive coding. By trying to predict sensory stimuli, the brain creates a simplified model of the world, including common sense physics. Yann LeCun says that this is a major key to AGI. Another one is effective planning.
But while current predictive models (autoregressive LLMs) work well on text, they don't work well on video data, because of the large outcome space. In an LLM, text prediction boils down to a probability distribution over a few thousand possible next tokens, while there are several orders of magnitude more possible "next frames" in a video. Diffusion models work better on video data, but they are not inherently predictive like causal LLMs. Apparently this new Doom model made some progress on that front though.
Howver, this is due how we actually digitize video. From a human point a view, looking in my room reduces the load to the _objects_ in the room and everyhing else is just noise ( like the color of the wall could be just a single item to remember, while otherwise in the digital world, it needs to remember all the pixels )
So, any given sequence of inputs is rebuilt into a corresponding image, twenty times per second. I wonder how separate the game logic and the generated graphics are in the fully trained model.
Given a sufficient enough separation between these two, couldn't you basically boil the game/input logic down to an abstract game template? Meaning, you could just output a hash that corresponds to a specific combination of inputs, and then treat the resulting mapping as a representation of a specific game's inner workings.
To make it less abstract, you could save some small enough snapshot of the game engine's state for all given input sequences. This could make it much less dependent to what's recorded off of the agents' screens. And you could map the objects that appear in the saved states to graphics, in a separate step.
I imagine this whole system would work especially well for games that only update when player input is given: Games like Myst, Sokoban, etc.
I see enemies and decreasing health on hit. But even if it lacked those, it seems like a pretty irrelevant nitpick that is completely underplaying what we're seeing here. The fact that this is even possible at all feels like science fiction.
Key: "predicts next frame, recreates classic Doom". A game that was analyzed and documented to death. And the training included uncountable runs of Doom.
A game engine lets you create a new game, not predict the next frame of an existing and copiously documented one.
429 comments
[ 2.5 ms ] story [ 295 ms ] threadThe two main things of note I took away from the summary were: 1) they got infinite training data using agents playing doom (makes sense), and 2) they added Gaussian noise to source frames and rewarded the agent for ‘correcting’ sequential frames back, and said this was critical to get long range stable ‘rendering’ out of the model.
That last is intriguing — they explain the intuition as teaching the model to do error correction / guide it to be stable.
Finally, I wonder if this model would be easy to fine tune for ‘photo realistic’ / ray traced restyling — I’d be super curious to see how hard it would be to get a ‘nicer’ rendering out of this model, treating it as a doom foundation model of sorts.
Anyway, a fun idea that worked! Love those.
The diffusion model doesn’t maintain any state itself, though its weights may encode some notion of cause/effect. It just renders one frame at a time (after all it’s a text to image model, not text to video). Instead of text, the previous states and frames are provided as inputs to the model to predict the next frame.
Noise is added to the previous frames before being passed into the SD model, so the RL agents were not involved with “correcting” it.
De-noising objectives are widespread in ML, intuitively it forces a predictive model to leverage context, ie surrounding frames/words/etc.
In this case it helps prevent auto-regressive drift due to the accumulation of small errors from the randomness inherent in generative diffusion models. Figure 4 shows such drift happening when a player is standing still.
The training was over almost 1 billion frames, 20 days of full-time play-time, taking a screenshot of every single inch of the map.
Now you show him N frames as input, and ask it "give me frame N+1", then it gives you the frame n. N+1 back based on how it was originally seen during training.
But it is not frame N+1 from a mysterious intelligence, it's simply frame N+1 given back from past database.
The drift you mentioned is actually a clear (but sad) proof that the model does not work at inventing new frames, and can only spit out an answer from the past dataset.
It's a bit like if you train stable diffusion on Simpsons episodes, and that it outputs the next frame of an existing episode that was in the training set, but few frames later goes wild and buggy.
If it would be able to invent action and maps and let the user play "infinite doom", then it would be very different (and impressive!).
I mean... no? Not even close? Multiply the number of game states with the number of inputs at any given frame gives you a number vastly bigger than 1 billion, not even comparable. Even with 20 days of play time to train no, it's entirely likely that at no point did someone stop at a certain location and look to the left from that angle. They might have done from similar angles, but the model then has to reconstruct some sense of the geometry of the level to synthesize the frame. They might also not have arrived there from the same direction, which again the model needs some smarts to understand.
I get your point, it's very overtrained on these particular levels of Doom, which means you might as well just play Doom. But this is not a hash table lookup we're talking about, it's pretty impressive work.
The map 1 has 2'518 walkable map units. There are 65536 angles.
2'518*65'536=165'019'648
If you capture 165M frames, you already cover all the possibilities in terms of camera / player view, but probably the diffusion models don't even need to have all the frames (the same way that LLMs don't).
Correct. You are certainly not moving between the tiles as discrete units in doom.
And Doom movement isn't tile based. The map may be, but you can be in many many places on a tile.
Generating "infinite Doom" is exactly what this model is doing, as it does not capture the larger map layout well enough to stay consistent with it.
They succeeded in the research, gained knowledge, and might be able to do something awesome with it.
It’s a success even if they don’t sell anything.
I would call it the world's least efficient video compression.
What I would like to see is the actual predictive strength, aka imagination, which I did not notice mentioned in the abstract. The model is trained on a set of classic maps. What would it do, given a few frames of gameplay on an unfamiliar map as input? How well could it imagine what happens next?
It's not super clear from the landing page, but I think it's an engine? Like, its input is both previous images and input for the next frame.
So as a player, if you press "shoot", the diffusion engine need to output an image where the monster in front of you takes damage/dies.
We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality.
To temper this a bit, you may want to pay close attention to the demo videos. The player rarely backtracks, and for good reason - the few times the character does turn around and look back at something a second time, it has changed significantly (the most noticeable I think is the room with the grey wall and triangle sign).
This falls in line with how we'd expect a diffusion model to behave - it's trained on many billions of frames of gameplay, so it's very good at generating a plausible -next- frame of gameplay based on some previous frames. But it doesn't deeply understand logical gameplay constraints, like remembering level geometry.
If on the backend you could record the level layouts in memory you could have exploration teams that try to find new areas to explore.
What if you combine this with an engine in parallel that provides all geometry including characters and objects with their respective behavior, recording changes made through interactions the other model generates, talking back to it?
A dialogue between two parties with different functionality so to speak.
(Non technical person here - just fantasizing)
“Everything” would mean all objects and the elements they’re made of, their rules on how they interact and decay.
A modularized ecosystem i guess, comprised of “sub-systems” of sorts.
The other model, that provides all interaction (cause for effect) could either be run artificially or be used interactively by a human - opening up the possibility for being a tree : )
This all would need an interfacing agent that in principle would be an engine simulating the second law of thermodynamics and at the same time recording every state that has changed and diverged off the driving actor’s vector in time.
Basically the “effects” model keeping track of everyones history.
In the end a system with an “everything” model (that can grow overtime), a “cause” model messing with it, brought together and documented by the “effect” model.
(Again … non technical person, just fantasizing) : )
Repeat for quest lines, new cities, etc, with the npcs having real time dialogue and interactions that happen entirely off screen, no guarantee of there being a massive quest objective, and some sort of recorder of events that keeps a running tally of everything that goes on so that as the PCs interact with it they are never repeating the same dreary thing.
If this were a MMORPG it would require so much processing and architecting, but it would have the potential to be the greatest game in human history.
It always blew my mind how well it worked on a 33 Mhz 486. I'm fairly sure it ran at 30 fps in 320x200. That gives it just over 17 clock cycles per pixel, and that doesn't even include time for game logic.
My memory could be wrong, though, but even if it required a 66 Mhz to reach 30 fps, that's still only 34 clocks per pixel on an architecture that required multiple clocks for a simple integer add instruction.
[0]: https://en.wikipedia.org/wiki/Inattentional_blindness#Invisi...
What this demo demonstrates to me is how incredible willing we are to accept what seems familiar to us as accurate.
I bet if you look closely and objectively you will see even more anomalies. But at first watch, I didn’t see most errors because I think accepting something is more efficient for the brain.
Objectively, Simons and Chabris (and many others) have a lot of data to support these ideas. Subjectively, I can say that these types of tasks (inattentional blindness, change blindness, etc.) are humbling.
Even having a clue why I'm linking this, I virtually guarantee you won't catch everything.
And even if you do catch everything... the real thing to notice is that you had to look. Your brain does not flag these things naturally. Dreams are notorious for this sort of thing, but even in the waking world your model of the world is much less rich than you think. Magic tricks like to hide in this space, for instance.
[0]: https://www.youtube.com/watch?v=wBoMjORwA-4
We don't memorize things that the environment remembers for us if they aren't relevant for other reasons.
edit: someone should train it on MyHouse.wad
The people were told to focus very deeply on a certain aspect of the scene. Maintaining that focus means explicitly blocking things not related to that focus. Also, there is social pressure at the end to have peformed well at the task; evaluating them on a task which is intentionally completely different than the one explicitly given is going to bias people away from reporting gorillas.
And also, "notice anything unusual" is a pretty vague prompt. No-one in the video thought the gorillas were unusual, so if the PEOPLE IN THE SCENE thought gorillas were normal, why would I think they were strange? Look at any TV show, they are all full of things which are pretty crazy unusual in normal life, yet not unusual in terms of the plot.
Why would you think the gorillas were unusual?
Furthermore, even what we attend to isn't always represented with all that much detail. Simons has a whole series of cool demonstration experiments where they show that they can swap out someone you're speaking with (an unfamiliar conversational partner like a store clerk or someone asking for directions), and you may not even notice [0]. It's rather eerie.
[0]: https://www.youtube.com/watch?v=FWSxSQsspiQ&t=5s
or do we believe it's an inherent limitation in the approach?
tangentially related but Grand Theft Auto speedrunners often point the camera behind them while driving so cars don't spawn "behind" them (aka in front of the car)
If I studied the longer one more closely, I'm sure inconsistencies would be seen but it seemed able to recall presence/absence of destroyed items, dead monsters etc on subsequent loops around a central obstruction that completely obscured them for quite a while. This did seem pretty odd to me, as I expected it to match how you'd described it.
But to say the model is simulating actual gameplay (i.e. that a person could actually play Doom in this) is far fetched. It's definitely great that the model was able to remember that the gray wall was still there after we turned around, but it's untenable for actual gameplay that the wall completely changed location and orientation.
It would in an SCP-themed game. Or dreamscape/Inception themed one.
Hell, "you're trapped in Doom-like dreamscape, escape before you lose your mind" is a very interesting pitch for a game. Basically take this Doom thing and make walking though a specific, unique-looking doorway from the original game to be the victory condition - the player's job would be to coerce the model to generate it, while also not dying in the Doom fever dream game itself. I'd play the hell out of this.
(Implementation-wise, just loop in a simple recognition model to continously evaluate victory condiiton from last few frames, and some OCR to detect when player's hit points indicator on the HUD drops to zero.)
(I'll happily pay $100 this year to the first project that gets this to work. I bet I'm not the only one. Doesn't have to be Doom specifically, just has to be interesting.)
Mainly just wanted to temper expectations I'm seeing throughout this thread that the model is actually simulating Doom. I don't know what will be required to get from here to there, but we're definitely not there yet.
Wouldn't be as pure though.
Fun variant: give it hidden state by doing the offscreen scratch pixel buffer thing, but not grading its content in training. Train the model as before, grading on the "onscreen" output, and let it keep the side channel to do what it wants with. It'd be interesting to see what way it would use it, what data it would store, and how it would be encoded.
https://en.wikipedia.org/wiki/MyHouse.wad
Maybe it is, but doing that with the entire scene instead of just a small part of it makes the problem massively harder, as the model needs to grow exponentially to remember more things. It isn't something that we will manage anytime soon, maybe 10-20 years with current architecture and same compute progress.
Then you make that even harder by remembering a whole game level? No, ain't gonna happen in our lifetimes without massive changes to the architecture. They would need to make a different model keep track of level state etc, not just an image to image model.
In this sora video the dragon covers half the scene, and its basically identical when it is revealed again ~5 seconds later, or about 150 frames later. The is lots of evidence (and some studies) that these models are in fact building internal world models.
https://www.youtube.com/watch?v=LXJ-yLiktDU
Buckle in, the train is moving way faster. I don't think there would be much surprise if this is solved in the next few generations of video generators. The first generation is already doing very well.
You always get this from AI enthusiast, they come and post "proof" that disproves their own point.
Most of the mob of people are indistinct, but there is a woman in a lime green coat who is visible, and then obstructed by the dragon twice (beard and ribbon) and reappears fine. Unfortunately when dragon fully moves past she has been lost to frame right.
There is another person in black holding a red satchel which is visible both before and after the dragon has passed.
Nothing about the storefronts appear to change. The complex sign full of Chinese text (which might be gibberish text: it's highly stylized and I don't know Chinese) appears to survive the dragon passing without even any changes to the individual ideograms.
There is also a red box shaped like a Chinese paper lantern with a single gold ideogram on it at the store entrance which spends most of the video obscured by the dragon and is still in the same location after it passes (though video artifacting makes it more challenging to verify that that ideogram is unchanged it certainly does not appear substantially different)
What detail are you seeing that is different before and after the obstruction?
First frame, guy in blue hat next to a flag. That flag and the guy is then gone afterwards.
The two flags near the wall are gone, there is something triangular there but there was two flags before the dragon went past.
Then not to mention that the crowd is 6 people deep after the dragon went past, while just 4 people deep before, it is way more crowded.
Instead of the flag that was there before the dragon, it put in 2 more flags afterwards far more to the left.
Third second a guy was out of frame for a few frames, and suddenly gained a blue scarf. AFter dragon went by he turned into a woman. Next to that person was a guy with a blue cap, he completely disappears.
> Most of the mob of people are indistinct
No they aren't, they are mostly distinct and basically all of them changes. If you ignore that the entire mob totally changes both in number and appearance and where it is, sure it is pretty good, except it forgot the flags, but how can you ignore the mob when we talk about the model remembering details? The wall is much less information dense than the mob, so that is much easier to remember for the model, the difficulty is in the mob.
> but there is a woman in a lime green coat who is visible,
She was just out of frame for a fraction of a second, not the big bit where the dragon moves past. The guy in blue jacket and blue cap behind her disappears though, or merges with another person and becomes a woman with a muffler after the dragon moved past.
So, in the end some big strokes were kept, and that was a very tiny part of the image that was both there before and after the dragon moved past so it was far from a whole image with full details. Almost all details are wrong.
Maybe he meant that the house looked mostly the same, I agree the upper parts does, but I looked at the windows and they were completely different, it is full of people heads after the dragon moved past while before it was just clean walls.
If you want to be right because you can find any difference. Sure. You win. But also completely missed the point.
Not in a game and those were enemies, it completely changed what and how many they are, people would notice such a massive change instantly if they looked away and suddenly there were 50% more enemies.
> The model clearly shows the ability to go beyond "image-to-image" rendering.
I never argued against that. Adding a third dimension (time) makes generating a video the same kind of problem as generating an image, it is not harder to draw a straight pencil with something covering it than to draw the scene with something covering it for a while.
But still, even though it is that simple, these models are really bad at it, because it requires very large models and much compute. So I just extrapolated based on their current abilities that we know, as you demonstrated there, to say roughly how long until we can even have consistent short videos.
Note that videos wont have the same progression as images, as the early image models were very small and we quickly scaled up there, while now for video we start at really scaled up models and we have to wait until compute gets cheaper/faster the slow way.
> But also completely missed the point.
You completely missed my point or you changed your point afterwards. My point was that current models can only remember little bits under such circumstances, and to remember a whole scene they need to be massively larger. Almost all details in the scene you showed were missed, the large strokes are there but to keep the details around you need an exponentially larger model.
Though I wonder if 10 years down the line folks wouldn't even care about underlying model details (no more than a current day web-developer needs to know about network packets).
PS: Not great examples, but I hope you get the idea ;)
A mistake people make all the time is that massive companies will put all their resources toward every project. This paper was written by four co-authors. They probably got a good amount of resources, but they still had to share in the pool allocated to their research department.
Even Google only has one Gemini (in a few versions).
You didn't say open _what_ models. Was that intentional?
When I read this part I thought you were going to say because you're technically not running Doom at all. That is, instead of running Doom without Doom's original hardware/software environment (by porting it), you're running Doom without Doom itself.
Not really? The greatest anti-Doom would be an infinite nest of these types of models predicting models predicting Doom at the very end of the chain.
The next step of anti-Doom would be a model generating the model, generating the Doom output.
Isn't that possible by setting arbitrarily high goals for ray-cast rendering?
- you could build a non-real-time version of the game engine and use the neural net as a real-time approximation
- you could edit videos shot in real life to have huds or whatever and train the neural net to simulate reality rather than doom. (this paper used 900 million frames which i think is about a year of video if it's 30fps, but maybe algorithmic improvements can cut the training requirements down) and a year of video isn't actually all that much—like, maybe you could recruit 500 people to play paintball while wearing gopro cameras with accelerometers and gyros on their heads and paintball guns, so that you could get a year of video in a weekend?
I imagine a game like that could get so convincing in its details and immersiveness that one could forget they're playing a game.
Not to mention this childish nonsense about "forget they're playing a game," as if every game needs to be lifelike VR and there's no room for stylization or imagination. I am worried for the future that people think they want these things.
Compare it to music gen algo's that can now produce music that is 100% indiscernible from generic crappy music. Which is insane given that 5 years ago it could maybe create the sound of something that maybe someone would describe as "sort of guitar-like". At this rate of progress it's probably not going to be long before AI is making better music than humans. And it's infinitely available too.
What makes you think a mechanical "predict next frame based on existing games" will be any good?
We could build a 'game' which would learn and adapt to precisely the chemistry that makes someone tick and then provide them a map to find the state in which their brain releases their desired state.
Then if the game has a directive - it should be pointed to work as a training tool to allow the user to determine how to release these chemicals themselves at will. Resulting in a player-base which no longer requires anything external for accessing their own desired states.
I guess I should start hoarding video of myself now.
Abstractly, it's like the model is dreaming of a game that it played a lot of, and real time inputs just change the state of the dream. It makes me wonder if humans are just next moment prediction machines, with just a little bit more memory built in.
As Richard Dawkins recently put it in a podcast[1], our genes are great prediction machines, as their continued survival rests on it. Being able to generate a visual prediction fits perfectly with the amount of resources we dedicate to sight.
If that is the case, what does aphantasia tell us?
[1] https://podcasts.apple.com/dk/podcast/into-the-impossible-wi...
The link may be that we actually see differences between “frames”, rather than the frames directly. That in itself would imply that a from of sub-visual representation is being processed by our brain. For aphantasia, it could be that we work directly on this representation instead of recalling imagery through the visual system.
Many people with aphantasia reports being able to visualize in their dreams, meaning that they don't lack the ability to generate visuals. So it may be that the brain has an affinity to rely on the abstract representation when "thinking", while dreaming still uses the "stable diffusion mode".
I’m no where near qualified to speak of this with certainty, but it seems plausible to me.
> Many people with aphantasia reports being able to visualize in their dreams, meaning that they don't lack the ability to generate visuals. So it may be that the [aphantasia] brain has an affinity to rely on the abstract representation when "thinking", while dreaming still uses the "stable diffusion mode".
(I obviously don't know what I'm talking about, just a fellow aphant)
Heck, it is far simpler than video, because the point of view and frame is fixed.
I highly suggest you to read the paper briefly before commenting on the topic. The whole point is that it's not just generating a video.
Further - "a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions." specifically "and actions"
User input is being fed into this system and subsequent frames take that into account. The user is "actually" firing a gun.
There cannot be "video compression artifacts" because it hasn’t even seen any compressed video during training, as far as I can see.
Seriously, how is this even a discussion? The article is clear that the novel thing is that this is real-time frame generation conditioned on the previous frame(s) AND player actions. Just generating video would be nothing new.
Yup, see https://en.wikipedia.org/wiki/Predictive_coding
[1] https://en.wikipedia.org/wiki/Lisa_Feldman_Barrett
[2] https://www.youtube.com/watch?v=NbdRIVCBqNI&t=1443s
This is an incredibly complex hypothesis that doesn't really seem justified by the evidence
This, to me, seems extremely reductionist. Like you start with AI and work backwards until you frame all cognition as next something predictors.
It’s just the stochastic parrot argument again.
It is running on an entire v5 TPU (https://cloud.google.com/blog/products/ai-machine-learning/i...)
It's unclear how that compares to a high-end consumer GPU like a 3090, but they seem to have similar INT8 TFLOPS. The TPU has less memory (16 vs. 24), and I'm unsure of the other specs.
Something doesn't add up, in my opinion, though. SD usually takes (at minimum) seconds to produce a high-quality result on a 3090, so I can't comprehend how they are like 2 orders of magnitudes faster—indicating that the TPU vastly outperforms a GPU for this task. They seem to be producing low-res (320x240) images, but it still seems too fast.
You don't even need to do all of that - this trained model already is the game, i.e., it's interactive, you can play the game.
Yes they had to use RL to learn what DOOM looks like and how it works, but this doesn’t necessarily pose a chicken vs egg problem. In the same way that LLMs can write a novel story, despite only being trained on existing text.
IMO one of the biggest challenges with this approach will be open world games with essentially an infinite number of possible states. The paper mentions that they had trouble getting RL agents to completely explore every nook and corner of DOOM. Factorio or Dwarf Fortress probably won’t be simulated anytime soon…I think.
At which point, you effectively would be interpolating in latent space through the source code to actually "render" the game. You'd have an entire latent space computer, with an engine, assets, textures, a software renderer.
With a sufficiently powerful computer, one could imagine what interpolating in this latent space between, say Factorio and TF2 (2 of my favorites). And tweaking this latent space to your liking by conditioning it on any number of gameplay aspects.
This future comes very quickly for subsets of the pipeline, like the very end stage of rendering -- DLSS is already in production, for example. Maybe Nvidia's revenue wraps back to gaming once again, as we all become bolted into a neural metaverse.
God I love that they chose DOOM.
The first thing I thought when I saw this was: couldn't my immediate experience be exactly the same thing? Including the illusion of a separate main character to whom events are occurring?
Neural nets are not guaranteed to converge to anything even remotely optimal, so no that isn't how it works. Also even though neural nets can approximate any function they usually can't do it in a time or space efficient manner, resulting in much larger programs than the human written code.
I don't really know why everyone is piling on me here. Sorry for a bit of fun speculating! This model is on the continuum. There is a latent representation of Doom in weights. some weights, not these weights. Therefore some representation of doom in a neural net could become more efficient over time. That's really the point I'm trying to make.
https://news.ycombinator.com/item?id=41377398
^__^It's easy to see this by noting that you can often prune networks quite a bit without any loss in performance. I.e. the effective dimension of the manifold the weights live on can be much, much smaller than the total capacity allows for. In fact, good regularization is exactly that which encourages the model itself to be compressible.
Capacity is autological. The amount of information it can express.
Training dynamics are the way the model learns, the optimization process, etc. So this is where things like regularization come into play.
There's also architecture which affects the training dynamics as well as model capacity. Which makes no guarantee that you get the most information dense representation.
Fwiw, the authors did also try distillation.
> With enough computation, your neural net weights would converge to some very compressed latent representation of the source code of DOOM. Maybe smaller even than the source code itself? Someone in the field could probably correct me on that.
And they're not wrong! An ideally trained network could, in principle, learn the data-generating program, if that program is within its class of representable functions. I might have a NN that naively looks like it takes up GBs of space, but it might actually be parameterizing a much simpler function (hence our ability to prune/compress the weights without performance loss - most of the capacity wasn't being used for any interesting computation).
You're right that there's no guarantee that the model finds the most "dense" representation. The goal of regularization is to encourage that, though!
All over the place in ML there are bounds like:
test loss <= train loss + model complexity
Hence minimizing model complexity improves generalization performance. This is a kind of Occam's Razor: the simplest model generalizes best. So the OP is on the right track - we definitely want networks to learn the "underlying" process that explains the data, which in this case would be a latent representation of the source code (well, except that doesn't really make sense since you'd need the whole rest of the compute stack that code runs on - the neural net has no external resources/embodied complexity it calls, unlike the source code which gets to rely on drivers, hardware, operating systems, etc.)
I suggested that this probably isn't the case here since they tried distillation and saw no effect. While this isn't proof that this particular model can't be compressed more it does suggest that it is non-trivial. This is especially true given the huge difference in size. I mean we're talking about 700x...
Where I think our disagreement is in that I read the OP as saying __this__ network. If we're talking about a theoretical network, well... nothing I said anywhere is in any disagreement with that. I even said in the post I linked to that the difference shows that there's still a long way to go but that this is still cool. Why did I assume OP was talking about __this__ network? Well because we're in a thread talking about a paper and well... yes, we're talking about compression machines so theoretically (well not actually supported by any math theory) this is true for so many things and that is a bit elementary. So makes more sense (imo) that we're talking about this network. And I wanted to make it clear that this network is nowhere near compression. Can further research later result in something that is better than the source code? Who knows? For all the reasons we've both mentioned. We know they are universal approximators (which are not universal mimicers and have limits) but we have no guarantee of global convergence (let alone proof such a thing exists in many problems).
And I'm not sure why you're trying to explain the basic concepts to me. I mentioned I was an ML researcher. I see you're a PhD at Oxford. I'm sure you would be annoyed if I was doing the same to you. We can talk at a different level.
I agree with you that this network probably has not found the source code or something like a minimal description in its weights.
Honestly, I'm writing a paper on model compression/complexity right now, so I may have co-opted the discussion to practice talking about these things...! Just a bit over-eager (,,>﹏<,,)
Have you given much thought to how we can encourage models to be more compressible? I'd love to be able to explicitly penalize the filesize during training, but in some usefully learnable way. Proxies like weight norm penalties have problems in the limit.
I actually have some stuff I'm working on in that area that is having some success. I do need to extend it to diffusion but I see nothing stopping me.
Personally I think a major slowdown for our community is it's avoidance of math. Like you don't need to have tons of math in the papers, but many of the lessons you learn in the higher level topics do translate to usable techniques in ML. Though I would also like to see a stronger push on theory because empirical results can be deceiving (Von Neumann's elephant and all)
I would expect something in this realm to be a little better at not being visually inconsistent when you look away and look back. A red monster turning into a blue friendly etc.
Not really. This is a reproduction of the first level of Doom. Nothing original is being created.
(Jk of course I know what you mean, but you can seriously see text prompts as compressed forms of programming that leverage the model's prior knowledge)
Sit down and write down a text prompt for a "fun new game". You can start with something relatively simple like a Mario-like platformer.
By page 300, when you're about halfway through describing what you mean, you might understand why this is wishful thinking
Also: https://news.ycombinator.com/item?id=41376722
Also: define "fun" and "new" in a "simple text prompt". Current image generators suck at properly reflecting what you want exactly, because they regurgitate existing things and styles.
I'm convinced this is the code that gives Data (ST TNG) his dreaming capabilities.
The demo is actual gameplay at ~20 FPS.
Like if I kill an enemy in some room and walk all the way across the map and come back, would the body still be there?
Edit: Can see this in the first 10 seconds of the first video under "Full Gameplay Videos", stairs turning to corridor turning to closed door for no reason without looking away.
Guessing the model hasn't been taught enough about that, because most people don't jump into hazards.
We could have mods for old games that generate voices for the characters for example. Maybe it's unfeasible from a computing perspective? There are people running local LLMs, no?
You mean in real time? Or just in general?
There are a lot of mods that use AI-generated voices. I'll say it's the norm of modding community now.
I was playing around with the idea in this: https://github.com/StreamUI/StreamUI. Thinking is take the ideas of Elixir LiveView to the extreme.
I too have been thinking about how to push dynamic wasm to the client for super low latency UIs.
LiveView is just the beginning. Your readme is dreamy. I'll dive into your project at the end of Sept when I get back into deep tech.
I'm wondering when people will apply this to other areas like the real world. Would it learn the game engine of the universe (ie physics)?
I think for real world application one challenge is going to be the "action" signal which is a necessary component of the conditioning signal that makes the simulation reactive. In video games you can just record the buttons, but for real world scenarios you need difficult and intrusive sensor setups for recording force signals.
(Again for robotics though maybe it's enough to record the motor commands, just that you can't easily record the "motor commands" for humans, for example)
https://slatestarcodex.com/2017/09/05/book-review-surfing-un...
It's called predictive coding. By trying to predict sensory stimuli, the brain creates a simplified model of the world, including common sense physics. Yann LeCun says that this is a major key to AGI. Another one is effective planning.
But while current predictive models (autoregressive LLMs) work well on text, they don't work well on video data, because of the large outcome space. In an LLM, text prediction boils down to a probability distribution over a few thousand possible next tokens, while there are several orders of magnitude more possible "next frames" in a video. Diffusion models work better on video data, but they are not inherently predictive like causal LLMs. Apparently this new Doom model made some progress on that front though.
Given a sufficient enough separation between these two, couldn't you basically boil the game/input logic down to an abstract game template? Meaning, you could just output a hash that corresponds to a specific combination of inputs, and then treat the resulting mapping as a representation of a specific game's inner workings.
To make it less abstract, you could save some small enough snapshot of the game engine's state for all given input sequences. This could make it much less dependent to what's recorded off of the agents' screens. And you could map the objects that appear in the saved states to graphics, in a separate step.
I imagine this whole system would work especially well for games that only update when player input is given: Games like Myst, Sokoban, etc.
A game engine lets you create a new game, not predict the next frame of an existing and copiously documented one.
This is not a game engine.
Creating a new good game? Good luck with that.