AI video you can watch and interact with, in real-time(experience.odyssey.world) |
AI video you can watch and interact with, in real-time(experience.odyssey.world) |
We talk about this challenge in our blog post here (https://odyssey.world/introducing-interactive-video). There's specifics in there on how we improved coherence for this production model, and our work to improve this further with our next-gen model. I'm really proud of our work here!
> Compared to language, image, or video models, world models are still nascent—especially those that run in real-time. One of the biggest challenges is that world models require autoregressive modeling, predicting future state based on previous state. This means the generated outputs are fed back into the context of the model. In language, this is less of an issue due to its more bounded state space. But in world models—with a far higher-dimensional state—it can lead to instability, as the model drifts outside the support of its training distribution. This is particularly true of real-time models, which have less capacity to model complex latent dynamics. Improving this is an area of research we're deeply invested in.
In second place would absolutely be model optimization to hit real-time. That's a gnarly problem, where you're delicately balancing model intelligence, resolution, and frame-rate.
It seems like it has a high chance of leading to even more narcissism as well because we are reducing our dependence on others to such a degree that we will care about others less and less, which is something that has already started happening with increasingly advanced interactive technology like AI.
I don't think its a step toward that; I think this is literally trained using techniques to generate more immersive virtual reality that already exists and takes less compute, to produce a more computationally expensive and less accurate AI version.
At least, that's what every other demo of a real-time interactive AI world model has been, and they aren't trumpeting any clear new distinction.
I got to the graffiti world and there were some stairs right next to me. So I started going up them. It felt like I was walking forward and the stairs were pushing under me until I just got stuck. So I turned to go back down and half way around everything morphed and I ended up back down at the ground level where I originally was. I was teleported. That's why I feel like something is cheating here. If we had mode collapse I'm not sure how we should be able to completely recover our entire environment. Not unless the model is building mini worlds with boundaries. It was like the out of bond teleportation you get in some games but way more fever dream like. That's not what we want from these systems, we don't want to just build a giant poorly compressed videogame, we want continuous generation. If you have mode collapse and recover, it should recover to somewhere new, now where you've been. At least this is what makes me highly suspicious.
What makes this AI generated over just rendering a generated 3D scene?
Like it may seem impressive to have no glitches (often in AI generated works you can turn around a full rotation and you're what's in front of you isn't what was there originally) but here it just acts as a fully modelled 3D scene rendering at low resolution? I can't even walk outside of certain bounds which doesn't make sense if this really is generated on the fly.
This needs a lot of skepticism and i'm surprised you're the first commenting on the lack of actual generation here. It's a series of static scenes rendered at low fidelity with limited bounds.
I think what's happening is this is AI generated but it is very very overfitted to real world 3D scenes. The AI is almost rendering exactly a real world scene and not much more. They can't travel out of bounds or the model stops working since it's so overfitted to these scenes. The overfitting solves hallucinations but it also makes it almost indistinguishable from pre modelled 3D scenes.
To clarify: this is a diffusion model trained on lots of video, that's learning realistic pixels and actions. This model takes in the prior video frame and a user action (e.g. move forward), with the model then generating a new video frame that resembles the intended action. This loop happens every ~40ms, so real-time.
The reason you're seeing similar worlds with this production model is that one of the greatest challenges of world models is maintaining coherence of video over long time periods, especially with diverse pixels (i.e. not a single game). So, to increase reliability for this research preview—meaning multiple minutes of coherent video—we post-trained this model on video from a smaller set of places with dense coverage. With this, we lose generality, but increase coherence.
We share a lot more about this in our blog post here (https://odyssey.world/introducing-interactive-video), and share outputs from a more generalized model.
> One of the biggest challenges is that world models require autoregressive modeling, predicting future state based on previous state. This means the generated outputs are fed back into the context of the model. In language, this is less of an issue due to its more bounded state space. But in world models—with a far higher-dimensional state—it can lead to instability, as the model drifts outside the support of its training distribution. This is particularly true of real-time models, which have less capacity to model complex latent dynamics.
> To improve autoregressive stability for this research preview, what we’re sharing today can be considered a narrow distribution model: it's pre-trained on video of the world, and post-trained on video from a smaller set of places with dense coverage. The tradeoff of this post-training is that we lose some generality, but gain more stable, long-running autoregressive generation.
> To broaden generalization, we’re already making fast progress on our next-generation world model. That model—shown in raw outputs below—is already demonstrating a richer range of pixels, dynamics, and actions, with noticeably stronger generalization.
Let me know any questions. Happy to go deeper!
Additionally, curious about what exactly the difference between the new mode of storytelling you’re describing and something like a crpg or visual novel is - is your hope that you can just bake absolutely everything into the world model instead of having to implement systems for dialogue/camera controls/rendering/everything else that’s difficult about working with a 3D engine?
Or is the next frame a function of just the previous frame and the user input? Like (previous frame, input) -> next frame
I'm asking because, if some world has two distinct locations that look exactly the same, will the AI distinguish them, or will they get coalesced into one location?
To be honest most of the appeal to me of this type of thing is the fact that it gets incoherent and morph-y and rotating 360 degrees can completely change the scenery. It's a trippy dreamlike experience whereas this kind of felt like a worse version of existing stuff.
(Some of this will be for benefit of other HN non-researcher readers)
I'm hoping you can provide some more. Are these training on single video moving through these environments, where the camera is not turning? What I am trying to understand is what is being generated vs what is being recalled.
It may be a more contentious view, but I do not think we're remotely ready to call these systems "world models" if they are primarily performing recall. Maybe this is the bias from an education in physics (I have a degree), but world modeling is not just about creating consistent imagery, but actually being capable of recovering the underlying physics of the videospace (as opposed to reality which the videos come from). I've yet to see a demonstration of a model that comes anywhere near this or convinces me we're on the path towards this.
The key difference here is are we building Doom which has a system requirements of 100MB disk and 8MB RAM with minimal computation or are we building a extremely decompressed version that requires 4GB of disk and a powerful GPU to run only the first level and can't even get critical game dynamics right like shooting the right enemy (GameNGen).
The problem is not the ability to predict future states based on previous ones, the problem is the ability to recover /causal structures/ from observation.
Critically, a p̶h̶y̶s̶i̶c̶s̶ world model is able to process a counterfactual.
Our video game is able to make predictions, even counterfactual predictions, with its engine. Of course, this isn't generated by observation and environment interaction, it is generated through directed programming and testing (where the testing includes observing and probing the environment). If the goal was just that, then our diffusion models would comparatively be a poor contender. It's the wrong metric. The coherence is a consequence of the world modeling (i.e. game engine) but it coherence can also be developed from recall. Recall alone will be unable to make a counterfactual.
Certainly we're in research phase and need to make tons of improvements, but we can't make these improvements if we're blindly letting our models cheat the physics and are only capable of picking up "user clicks fire" correlating with "monster usually dies when user shoots". LLMs have tons of similar problems with making such shortcuts and the physics will tell you that you are not going to be able to pick up such causal associations without some very specific signals to observe. Unfortunately, causality can not be determined from observation alone (a well known physics result![0]). You end up with many models that generate accurate predictions, and these become non-differentiable without careful factorization, probing, and often requiring careful integration of various other such models. It is this much harder and more nuanced task that is required of a world model rather than memory.
Essentially, do we have "world models" or "cargo cult world models" (recall or something else).
That's the context of my data question. To help us differentiate the two. Certainly the work is impressive and tbh I do believe there is quite a bit of utility in the cargo cult setting, but we should also be clear about what is being claimed and what isn't.
I'm also interested in how you're trying to address the causal modeling problem.
[0] There is much discussion on the Duhem-Quine thesis, which is a much stronger claim than I stated. There's the famous Michelson-Morley experiment, which actually did not rule out an aether, but rather only showed that it had no directionality. Or we could even use the classic Heisenberg Uncertainty Principle which revolutionized quantum mechanics showing that there are things that are unobservable, leading to Schrodinger's Cat (some weird hypotheses of multiverses). And we even have String Theory, which the main gripe remains that it is non-differentiable from other TOES due to differences in predictions being non-observable.
Right. I was never able to get very far from the starting point, and kept getting thrown back to the start. It looks like they generated a little spherical image, and they're able to extrapolate a bit from that. Try to go through a door or reach a distant building, and you don't get there.
> I mean... https://news.ycombinator.com/item?id=44121671 informed you of exactly why this happens a whole hour before you posted this comment and the creator is chatting with people in the comments.
> and the creator is chatting with people in the comments.
> is effectively a stereotypical "who needs dropbox" levels of shallow dismissal.
https://news.ycombinator.com/item?id=44147777
> and i only say this because it's the top comment on this post
I call BS.
"To improve autoregressive stability for this research preview, what we’re sharing today can be considered a narrow distribution model: it's pre-trained on video of the world, and post-trained on video from a smaller set of places with dense coverage. The tradeoff of this post-training is that we lose some generality, but gain more stable, long-running autoregressive generation."
I wonder if it'd break our brains more if the environment changes as the viewpoint changes, but doesn't change back (e.g. if there's a horse, you pan left, pan back right, and the horse is now a tiger).
In a way, that almost makes it more dreamlike, in that you have what feels like high local coherence (just enough not to immediately tip you off that it’s a dream) that de-coheres over time as you move through it.
Fascinatingly strange demo.
[0] Minecraft with object impermanence (229 points, 146 comments) https://news.ycombinator.com/item?id=42762426
What AI can dream up in milliseconds could take hundreds of human hours to encode using traditional tech (meshes, shaders, ray tracing, animation, logic scripts, etc.), and it still wouldn't look as natural and smooth as AI renderings — I refer to the latest developments in video synthesis like Google's Veo 3. Imagine it as a game engine running in real time.
Eventually managed to leave the first room, but then got teleported somewhere else.
I LOVE dreamy AI content. That stuff where everything turned into dogs for example.
As AI is maturing, we are slowly losing that im favor of boring realism and coherence.
We think it's a glimpse of a totally new medium of entertainment, where models imagine compelling experiences in real-time and stream them to any screen.
Once you've taken the research preview for a whirl, you can learn a lot more about our technical work behind this here (https://odyssey.world/introducing-interactive-video).
ie. as opposed to first generating a 3d env then doing some sorts of img2img on top of it?
This would explain:
1. How collisions / teleportation work and why they're so rigid (the WM is mimicking hand-implemented scene-bounds logic)
2. Why the scenes are static and, in the case of should-be-dynamic elements like water/people/candles, blurred (the WM is mimicking artifacts from the 3D representation)
3. Why they are confident that "There's no map or explicit 3D representation in the outputs. This is a diffusion model, and video in/out" https://x.com/olivercameron/status/1927852361579647398 (the final product is indeed a diffusion WM trained on videos, they just have a complicated pipeline for getting those training videos)
- Open Source Diamond WM that you can run on consumer hardware [1]
- Google's Genie 2 (way better than this) [2]
- Oasis [3]
[1] https://diamond-wm.github.io/
[2] https://deepmind.google/discover/blog/genie-2-a-large-scale-...
[3] https://oasis.decart.ai/welcome
There are a lot of papers and demos in this space. They have the same artifacts.
From our perspective, what separates our work is two things:
1. Our model is able to be experienced by anyone today, and in real-time at 30 FPS.
2. Our data domain is real-world, meaning learning life-like pixels and actions. This is, from our perspective, more complex than learning from a video game.
> Why are you going all in on world models instead of basing everything on top of a 3D engine that could be manipulated / rendered with separate models?
I absolutely think there's going to be super cool startups that accelerate film and game dev as it is today, inside existing 3D engines. Those workflows could be made much faster with generative models.
That said, our belief is that model-imagined experiences are going to become a totally new form of storytelling, and that these experiences might not be free to be as weird and whacky as they could because of heuristics or limitations in existing 3D engines. This is our focus, and why the model is video-in and video-out.
Plus, you've got the very large challenge of learning a rich, high-quality 3D representation from a very small pool of 3D data. The volume of 3D data is just so small, compared to the volumes generative models really need to begin to shine.
> Additionally, curious about what exactly the difference between the new mode of storytelling you’re describing and something like a crpg or visual novel
To be clear, we don't yet know what shape these new experiences will take. I'm hoping we can avoid an awkward initial phase where these experiences resemble traditional game mechanics too much (although we have much to learn from them), and just fast-forward to enabling totally new experiences that just aren't feasible with existing technologies and budgets. Let's see!
> is your hope that you can just bake absolutely everything into the world model instead of having to implement systems for dialogue/camera controls/rendering/everything else that’s difficult about working with a 3D engine?
Yes, exactly. The model just learns better this way (instead of breaking it down into discrete components) and I think the end experience will be weirder and more wonderful for it.
Isn’t the entire aim of world models (at least, in this particular case) to learn a very high quality 3D representation from 2D video data? My point is if that you manage to train a navigable world model for a particular location, that model has managed to fit a very high quality 3D representation of that location. There’s lots of research dealing with NERFs that demonstrate how you can extract these 3D scenes as meshes once a model has managed to fit it. (NERFs are another great example of learning a high quality 3D representation from sparse 2D data.)
>That said, our belief is that model-imagined experiences are going to become a totally new form of storytelling, and that these experiences might not be free to be as weird and whacky as they could because of heuristics or limitations in existing 3D engines. This is our focus, and why the model is video-in and video-out.
There’s a lot of focus in the material on your site about the models learning physics by training on real world video - wouldn’t that imply that you’re trying to converge on a physically accurate world model? I imagine that would make weirdness and wackiness rather difficult
> To be clear, we don't yet know what shape these new experiences will take. I'm hoping we can avoid an awkward initial phase where these experiences resemble traditional game mechanics too much (although we have much to learn from them), and just fast-forward to enabling totally new experiences that just aren't feasible with existing technologies and budgets. Let's see!
I see! Do you have any ideas about the kinds of experiences that you would want to see or experience personally? For me it’s hard to imagine anything that substantially deviates from navigating and interacting with a 3D engine, especially given it seems like you want your world models to converge to be physically realistic. Maybe you could prompt it to warp to another scene?
> wouldn’t that imply that you’re trying to converge on a physically accurate world model?
Unreal engine 5 has been demoing these features for a while now, I heard about it early 2020 iirc, but the techniques like gaussian splattering predate it.
I have no experience in either of these, but I believe MegaScans and RealityCapture are two examples doing this. And the last nanite demo touched on it, too.