By Yunzhong, from凹非寺
量子位 | 公众号QbitAI
A real-time world model that grew out of a domestic virtual film studio has won an award on NVIDIA's stage.
On October 15, the 2026 NVIDIA Startup Showcase Half-way Presentation was held at Suzhou Jinji Lake. Dawa Technology from Chongqing won the "Half-way Showcase Outstanding Enterprise Award" for its real-time world model productScenes AItook home「Half-way Showcase Outstanding Enterprise Award」。
This year's competition held 11 regional events nationwide, with more than 580 tech startups registering, and ultimately 30 members of the NVIDIA Inception program entered the half-way showcase.
Unlike most video models that compete on "how alike the next frame looks," Scenes AI takes a different path:
instead of predicting the next frame, itmaintains a continuously running world state— the scene, characters, camera, lighting, objects and physics each form a set of variables that can be read and written independently.
Whenever something changes in the world, what the model must do is not redraw the entire sequence of footage, but update only the part of the state that changed, then write the result back into the world, letting the next moment continue to evolve on that basis.
So the question it must answer has changed too: it is not "what does the next frame look like," butif you change one variable in the world, what will happen to the world next?
What's more crucial, its input is not a game engine or agent footage, buta real space captured by an ordinary camera。
In Dawa's own words:A video is an output, while a world is a process. A video is an output. A world is a process.
First, look at the results: you can change an entire world, yet modify only one wall
Let's start with the latest version of the demo.
The top-left corner is the real room captured by the camera at this moment: white walls, a green screen, an iron chair, a teddy bear in the corner, and a person standing in the middle adjusting the brim of a hat.
The large picture is the model's output at the same moment: the room has become a hand-drawn-style summer grassland, and the person has become an animated girl wearing a yellow baseball cap, with movements in sync.
Looking closely, you'll find that only the appearance has changed. The chair is still behind the girl, the little bear still sits in its original "corner," the position and perspective of the dark cabinet on the right have not moved, and the grassland even faintly retains the fold line where two walls meet.
Test point one: global re-rendering, the structure must not fall apart.
In this frame, the only variable that has been rewritten is the "art style.".
Character movements, furniture positions, occlusion relationships and spatial perspective are all locked in as preservation conditions. The model must first recognize what the person, chair, doll and cabinet each are, then keep them pinned in place while restyling the whole room — and it must hold up in every single frame.
Test point two: local rewriting, and it has to remember.
In the same scene, a child writes "DW Future" on the wall, then picks up a plush toy and runs out of the frame; the young man stands up, trees grow in the room, and a campfire burns on the floor; next he raises his hand to control the weather, and the room first sees snow falling, then rain.
Through this chain of events, people come and go, new things grow in the scene, and the weather changes several times, yet those four characters on the wall remain the whole time, with their position and strokes unchanged.
What this tests is whether the world state is correctly written back and persistently maintained.
For frame-by-frame generative models, the more changes in the picture, the more easily previously left details get "washed away," and every re-render may quietly alter them; but in Scenes AI, "there is writing on the wall" becomes part of the world state the moment it is written down. No matter what happens afterwards, as long as no event modifies it, it stays there.
The September (9月) debut demo followed the same logic: after the snow monster burst out of the ground and left, the pit in the ground stayed in place; the wrought-iron chair was rewritten twice in a row into an iron throne with a wolf head, and the streamer could still sit back in the original spot; when the bottom of the curtain caught fire, the adjacent plush toy and flower pot were untouched.
The entire generation runs locally on a single RTX graphics card, with the frame rate displayed in real time in the upper right corner of the picture.
A different route: from "repainting the next frame" to "updating the world state"
Behind the demo is Dawawa's basic judgment about world models:
Video generation is already powerful enough, but the world has not yet begun to run.
The deliverable of mainstream video generation is 'a single result': you give a prompt, it generates a nice-looking video, and then it's over.
Even when made into frame-by-frame continuous synthesis, the focus is still on 'what the next frame should look like' — errors in identity, geometry, lighting, and physics accumulate little by little over time.
Dawa defines a world model differently:after changing one state variable, why the world changes and how it will evolve.
Take the simplest example, inputting the phrase 'turn the lighting into sunset':
- Traditional video model: regenerates the entire video, with the scene, characters, camera, and objects all potentially drifting along;
- Scenes AI: the scene, characters, camera, and objects are frozen as held conditions; only the single variable 'lighting' changes, and a new frame is generated from the new world state.
Dawa condenses this approach into one sentence: modify one variable while preserving the entire world.
In computational terms, this means shifting from 'repainting the next frame' to"updating the state delta ΔS"— only updating the parts that changed, while the rest of the world state remains stable, queryable, and predictable.
Under this definition, a continuously running world is broken down into five parts:
A persistent world equals world memory plus neural physics plus sparse diffusion plus neural rendering plus AI lighting.
namely world memory, neural physics, sparse diffusion, neural rendering, and AI lighting.
In Dawa's view, moving from generating a video to running an interactive world requires passing five gates simultaneously:
cross-frame memory, local controllability, physics and causality, real-time computational efficiency, and data that can explain 'why the world changes this way'.
In their words,the difficulty is not building a bigger model, but making state, control, physics, and compute all hold true at the same time.
The precise control and real-time rendering to be discussed later are exactly this route's answers to the 'control' and 'compute' gates.
Technical architecture: make the control layer heavy and the generation layer light
Turning this route into an engineering system, Dawa's approach can be summed up in one sentence:make the control layer heavy and the generation layer light.
Scenes AI does not hand all problems to one large model to repaint frame by frame; instead, between 'creative intent' and 'image generation', it adds a control layer centered on world state, which Dawa calls thecontrol layer for video generation。
- control layer: it decomposes continuous video into queryable, updatable, constrainable state factors — scene and space, characters and actions, camera and lens, lighting and materials, assets and wardrobe/props, physics and time; it is responsible for planning actions, maintaining state, and simulating outcomes.
- rendering layer: it receives the conditions issued by the control layer, generates only the parts that need to change, and composites them back into the frame.
A typical execution pipeline is:
world state → action or contact event → state update → local generation and compositing → result written back.
Again, take the earlier 'DW Future' demo as an example:
- when the child writes on the wall, the event layer recognizes the writing action and the contact location, i.e., the wall;
- The control layer updates the world state accordingly, recording “there is writing on the wall” onto the wall object;
- Afterwards, trees grow in the scene, a campfire is lit, and it snows and rains; the control layer only updates the corresponding scene elements and weather variables, the rendering layer only generates the trees, flames, snowflakes and raindrops along with the lighting and shadow changes they bring, while the wall and the writing directly reuse their existing representations;
- Weather processes like rain and snow come and go, making them transient changes; the writing on the wall, by contrast, persists as durable state in the world, to be continuously read by subsequent shots and interactions.
There is a key distinction here:which changes are “transient” and which are “persistent”.。
The former can be handled and discarded by the rendering layer, while the latter must enter the world state. In the Demo, “the pit is still there” and “the writing is still there” are both outward manifestations of this mechanism.
Viewed from an industry perspective, “separating state maintenance from visual rendering” is gradually becoming a consensus for real-time world models; the disagreement lies inwhere the state comes from.。
One approach has the upstream engine or program explicitly write out the rules and state, with the model rendering through reference frames as the interface, suitable for games and prototypes with well-defined rules;
Scenes AI targets real filmed and livestreamed spaces, where the world state must be parsed by the model from camera footage, then continuously maintained, inferred upon, and written back.
Precise control: change one variable while everything else stays fixed
Precise control resides in the control layer, whose core module isLatent Stage。
First, variable decoupling.
Latent Stage decomposes scene, characters, camera, lighting, clothing, props, and physics into independent "stages" in latent space, which can evolve individually or be jointly constrained.
Every edit has a clear "change set" and a "keep set": when changing lighting, the scene, characters, and camera position are locked; when changing clothing, character identity and body motion continue to constrain generation.
The screenshot showing "art style fully changed, furniture untouched" is the result of separating the change set and keep set cleanly enough.
Second, the non-linear video matrix.
Dawa's proposed non-linear view matrix fusion method turns video from "linear time" into an "editable matrix": characters, environment, camera, lighting, objects, and motion are broken into visual elements that can be recombined, making it easier for the control layer to understand camera movement and blend real with generated content.
The related demonstration paper, "NVD: Non-Linear View Matrix Fusion for 2D Video Using Multimodal Foundation Models," has been accepted by ACM UbiComp / ISWC 2026 and was presented on site on October 13 at the Shanghai International Convention Center. The research was jointly completed by the Dawa AIGC Research Institute together with The Hong Kong Polytechnic University, the University of Illinois Urbana-Champaign, and Newcastle University.
Third, world memory.
World Memory persistently caches fragments of historical states, providing queryable and updatable context for subsequent generation.
When the camera leaves and comes back, the model reads the state record of that region instead of re-imagining a wall.
Dawa regards long-horizon memory as the most critical challenge today: identity, space, physics, and story state must all remain consistent across time.
Without causal memory, no amount of GPU accumulation will solve the problem.
Fourth, physical constraints.
Neural Physics covers rigid bodies, cloth, fluids, and particles, organizing generation conditions around collision and gravity, determining where fire should burn up to and how cloth should be pushed apart; Neural Renderer then converts geometry, materials, and lighting into images.
Physical state thereby becomes the common reference for generation, training, and evaluation, rather than treating the video itself as the answer.
In continuous editing tests, Scenes AI achieved atarget edit success rate of 98.2%，and a character identity preservation rate of 93.3%。
Real-time rendering: dedicating compute to the world that is actually changing
The rendering layer is handled by the real-time generative modelLiveStage. The problem it solves is very concrete:
keeping the imagery continuously and low-latently tracking a real person within the compute and memory budget of a single consumer-grade graphics card.
First, sparse diffusion autoregression.
The autoregressive component handles continuous evolution over time, while sparse computation limits the scope of each update—meaning the model predicts state deltas rather than an entirely new frame.
Based on analysis of data from more than 300 real commercial shoots, Dawa found that in typical scenes, the regions that actually change per frame account for only a fraction of the image.10%—20%。
As a result, stable regions such as walls and floors are directly reused via a latent cache (Latent Cache), concentrating compute on characters, faces, hands, cloth, smoke, light and shadow, and interaction regions.
Second, causal sparse micro-diffusion.
The rendering layer receives a set of conditions issued by the control layer—including event type, position, velocity, intensity, object identity, materials, and local history—and performs only a few sampling steps on the changing regions, then composites the local results back into the persistent world image.
"Causalization" means each step depends only on history that has already occurred, making it suitable for streaming output;
"micro-diffusion" means each step processes only a small local region rather than the entire image.
Third, sparse activation.
The model has 34B total parameters, with only about 4B activated at each step, an activation ratio of about 11.8%.
Spatially local updates reduce "how much needs to be computed", while the model's internal sparse activation controls "how many parameters are used in each computation"; together these constitute the characteristic of "the capacity of a large model, with a real-time computation path".
The result of combining all three: compared with traditional video diffusion approaches,GPU consumption is reduced by 40%—60%, and inference speed is improved 3—5 times.。
When the frame requires an overall change, such as the entire room in the screenshot turning into a grassland, where the area of change far exceeds 20%, the system dynamically allocates computation according to the extent of change, and local editing and full-scene transformation share the same pipeline.
Ultimately, on a single NVIDIA GeForce RTX consumer-grade graphics card, Scenes AI achieves real-time output at 720P and 25 frames.
Underlying compute: natively built on the NVIDIA technology stack.
Dawa's technology partnership with NVIDIA began in 2018, the year Dawa purchased its first NVIDIA DGX-1.
Today, Scenes AI uses CUDA for training and inference acceleration, TensorRT for inference optimization, and NVENC to support real-time video encoding output.
Sparse updating reduces wasted computation, the Latent Cache facilitates memory reuse, and the layered architecture is naturally suited to layered deployment—the mechanism itself is designed around the GPU's parallel computing and computation reuse.
Core moat: not a Video Dataset, but a State Dataset
With the route clarified, the real difficulty lies in actually training it. Dawa's accumulated expertise can be broken down into three layers.
Layer one: state data
To learn "how the world changes," video alone is not enough.
Most videos on the internet are edited final products—how the camera moved, how the lights were set, when props were picked up—this process information has almost all been edited out, and the pixels themselves cannot explain why the world changes the way it does.
Dawa's formulation is:Not a Video Dataset, but a State Dataset.
Dawa operates an 8.4万平方米 full-process digital film studio — correction: an 84,000-square-meter full-process digital technology film studio in Yongchuan, Chongqing.
Over the past 5 years, the team has continuously collected, annotated, and curated data from more than 300 real commercial projects, accumulating approximately180TB of structured world state data, covering six core dimensions: camera trajectories, motion trajectories, object states, lighting states, physical states, and narrative states.
A single scene receives structured annotation across more than 300 dimensions, aligning shots, optics, space, performance, and motion onto the same timeline.
Film sets also naturally concentrate the long-tail samples that are hardest for real-time generation: wide sleeves and skirts deforming with movement, hands repeatedly making contact with props, multiple characters occluding one another, the camera moving continuously during performance, and lighting changes acting simultaneously on people and environment.
Beyond real data, Dawa uses NVIDIA Omniverse and PhysX to build simulation scenes, batch-generating annotated data with precise physical states, and uses NVIDIA Cosmos on H200 to generate "process-type" data such as melting ice cream, growing trees, and melting snow.
The three types of data have clear divisions of labor:Real shooting provides the distribution, simulation provides precise states, and generated data fills coverage gaps.
Layer two: training route
Dawa is training not a "video generator" but a continuously evolving world system, divided into five overall stages:
- World cognition pre-training: self-supervised learning on 180TB of film and television data to establish a unified latent space;
- Spatiotemporal prediction training: future prediction, video completion, and state reconstruction, learning temporal evolution rather than continuous video stitching;
- Specialized capability training: modules such as local sparse micro-diffusion, nonlinear video matrices, neural rendering, and neural physics each specialized;
- Joint optimization: multi-task joint training and curriculum learning, end-to-end optimization in the shared latent space;
- Agent-driven reinforcement learning: multi-agent collaboration among directors, cinematography, lighting, and animation, continuously optimizing shot quality and scene consistency.
Layer three: the production loop
Dawa itself is already engaged in real film and television production and short drama production. Clients' revisions, feedback from interactions, and cases of model errors flow back as entry points for the next round of data screening, supplementary collection, and training — in other words, the model is tested on set, and the set in turn feeds the model.
"We shoot real films and short dramas every day—every lighting setup and every camera movement teaches the model how this world works," said Chen Liang, co-founder of Dawa Technology. "The next battle for world models belongs to teams that possess real-world processes."
From "generating a result" to "running a world"
World models should not remain an abstract concept; when applied to concrete use cases, this approach corresponds to several clear categories of scenarios:
- Film, Television and Short Drama: Directors can adjust and transform more creative ideas in controllable settings, instantly converting each spark of inspiration into parameterized weather, lighting, wardrobe, and environment plans;
- Live Streaming and E-commerce: turning individual "grid livestream rooms" into true "scenario livestream rooms," where the host's surrounding space switches in real time with products and topics—selling snow boots in the snow, hiking shoes on a high mountain;
- culture & tourism, gaming, and interactive display: content is no longer an object of one-way viewing; users can directly enter real spaces to interact with, rewrite, and extend them.
They share the same set of requirements: changes must be controllable, output must be timely, and costs must be estimable.
It runs on a single consumer-grade graphics card and a local machine, with no reliance on cloud computing power—this is what moving from laboratory demos toward a real environment closer to actual deployment looks like.
"Video generation is already powerful enough, but the world has not yet started to run. A video can end; a world must keep running." At the mid-stage showcase roadshow, Li Zongxing, Chief Product Officer of Dawa Technology, summarized the starting point of Scenes AI this way.
"NVIDIA provides a solid computing foundation for real-time world models," he said. "We hope to work with the NVIDIA ecosystem to bring world models from the laboratory into real production, so that every creator can use a single graphics card to build, rewrite, and step into a continuously running world in real time."
△ Photo of Dawa Technology CPO Li Zongxing at the mid-stage showcase roadshow
Looking back at Dawa's path over the past few years, the key puzzle pieces have already fallen into place.
The data is there: an 8.4 (in 10,000s of square meters) film studio base—84,000 square meters—over 300 real projects, and 180TB of world-state data, still continuously accumulating;
The model works: a 34B world model running in real time on a single consumer-grade graphics card, with precise control and real-time rendering achieved simultaneously;
Recognition has come too: an Outstanding Enterprise award at the Shenzhen station in August, a UbiComp paper presented in October, and in the same month the NVIDIA Mid-Stage Showcase Outstanding Enterprise Award.
As the industry's attention begins to shift from "generating a video" to "running a world," this route that grew out of the film studio is being seen by more and more people.
Videos end; worlds do not. For Scenes AI, this world has only just begun to run.
*This article is republished by QbitAI with authorization; the views belong solely to the original author.