Dawa Technology's Scenes AI real-time world model runs a 34B-parameter model on a single RTX GPU

QbitAI | WeChat Official Account QbitAI

A real-time world model that grew out of a domestic virtual production stage has won an award on NVIDIA's stage.

On October 15, the 2026 NVIDIA Startup Showcase Midterm Presentation was held at Jinji Lake in Suzhou, where Chongqing-based Dawa Technology won the "Midterm Showcase Excellence Enterprise Award" with its real-time world model product Scenes AI.

This year's competition held 11 regional events across the country, with more than 580 tech startups registering, and ultimately 30 NVIDIA Inception program members advanced to the midterm showcase.

Unlike most video models that compete over whether "the next frame looks right," Scenes AI takes a different path:

It does not predict the next frame of imagery, but rather maintains a continuously running world state — scenes, characters, cameras, lighting, objects, and physics, each being a set of variables that can be read and written independently.

Whenever something changes in the world, what the model needs to do is not redraw the entire segment of imagery, but only update the part of the state that has changed, then write the result back into the world, letting the next moment continue to evolve on that basis.

So the question it has to answer also changes: not "what does the next frame look like," but if you change one variable in the world, what will happen to the world next?

More critically, its input is not a game engine or proxy imagery, but the real space captured by an ordinary camera.

In Dawa's own words: A video is an output. A world is a process.

Let's first look at the latest Demo.

In the upper 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 holding the brim of a hat.

The large image is the model 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 synchronized movements.

Looking closely, you'll find that only the appearance has changed. The chair is still behind the girl, the little bear is still sitting in the original "corner," and the position and perspective of the dark cabinet on the right have not moved; on the grass, the polyline where the two walls meet is even faintly retained.

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 as preservation conditions. The model must first recognize what the person, chair, doll, and cabinet each are, then pin them in place while changing the style of the entire room, and this must hold for every frame.

In the same scene, a child writes "DW Future" on the wall, then picks up a doll and runs out of the frame; a young man gets up, trees grow in the room, and a bonfire ignites on the ground; then he raises his hand to control the weather, and first snow falls indoors, then rain.

Through this series 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 are always still there, with their position and strokes unchanged.

This tests whether the world state has been correctly written back and continuously maintained.

For frame-by-frame generation models, the more changes in the image, the easier it is for previously retained details to be "washed away," and each redraw may quietly alter them; in Scenes AI, "there is writing on the wall" becomes part of the world state at the moment it is written. After that, no matter what happens, as long as no event changes it, it stays there.

The same logic applies in the September debut demo: after the snow monster breaks through the ground and leaves, the pit on the ground remains in place; the iron chair is rewritten twice in succession into a wolf-headed iron throne, and the streamer can still sit back in the original position; the hem of the curtain catches fire, and the doll and flowerpot right next to it are not affected.

The entire generation runs on a single local RTX graphics card, with the frame rate displayed in real time in the upper right corner of the image.

Behind the Demo is Dawa'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 segment of result": give a prompt, generate a good-looking video, and then it ends.

Even when made into continuous frame-by-frame synthesis, the focus is still on "what the next frame looks like," and errors in identity, geometry, lighting, and physics will gradually accumulate over time.

Dawa's definition of a world model is different: after changing a state variable, why the world changes and how it will evolve.

Take the simplest example: input the sentence "turn the lighting into sunset":

Dawa condenses this line of thinking into one sentence: modify one variable, preserve the entire world.

In computational terms, this means shifting from "redrawing the next frame" to "updating the state increment ΔS" — updating only the local part that has 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:

In Dawa's view, going from generating a segment of 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 larger model, but that state, control, physics, and compute power must all hold at the same time.

The precise control and real-time rendering to be discussed later are precisely this route's answers at the two gates of "control" and "compute power."

To turn this route into an engineering system, Dawa's approach can be summed up in one sentence: make the control layer heavy, make the generation layer light.

Scenes AI does not hand every problem to a single large model to redraw frame by frame. Instead, between "creative intent" and "image generation," it adds a control system centered on world state, which Dawa calls the control layer for video generation.

A typical operating chain is:

World state → action or contact event → state update → local generation and compositing → result written back.

Let's still use the earlier "DW Future" demo as an example:

There is a key distinction here: which changes are "transient" and which are "persistent."

The former can be handled by the rendering layer and that's it; the latter must enter the world state. In the demo, "the pit is still there" and "the text is still there" are both external manifestations of this mechanism.

Looking at the industry, "separating state maintenance from visual rendering" is gradually becoming the consensus for real-time world models; the disagreement lies in where state comes from.

One approach has an upstream engine or program explicitly write out rules and state, with the model using reference images as an interface for rendering, suitable for games and prototypes with clear rules;

Scenes AI targets real shooting and live-streaming spaces, where world state must be parsed by the model from camera images, then continuously maintained, inferred, and written back.

Precise control lands in the control layer, whose core module is Latent Stage.

First, variable decoupling.

In latent space, Latent Stage splits the scene, characters, camera, lighting, clothing, props, and physics into independent "stages" that can evolve separately or be jointly constrained.

Every edit has an explicit "change set" and "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.

In the screenshot, "the art style is completely changed while the furniture stays put" is the result of the change set and keep set being separated cleanly enough.

Second, the non-linear video matrix.

The non-linear view matrix fusion method proposed by Dawa turns footage from "linear time" into an "editable matrix": characters, environment, camera, lighting, objects, and motion are split into visual elements that can be recombined, making it easier for the control layer to understand camera movement and fuse real and generated content.

The related demo 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 completed jointly by the Dawa AIGC Research Institute with the Hong Kong Polytechnic University, the University of Illinois Urbana-Champaign, and Newcastle University.

Third, world memory.

World Memory persistently caches historical state segments, providing queryable, updatable context for subsequent generation.

When the camera leaves and comes back, the model reads the state record of that area rather than reimagining a wall.

Dawa regards long-horizon memory as the most critical challenge at present: identity, space, physics, and story state must all remain consistent across time.

Without causal memory, no amount of GPUs can solve the problem.

Fourth, physical constraints.

Neural Physics covers rigid bodies, cloth, fluids, and particles, organizing generation conditions around collisions and gravity, determining where fire should stop burning and how cloth should be pushed aside; Neural Renderer then turns geometry, materials, and lighting into images.

Physical state thus becomes a common reference for generation, training, and evaluation, rather than treating the video itself as the answer.

In continuous editing tests, Scenes AI achieved a target edit success rate of 98.2% and a character identity preservation rate of 93.3%.

The rendering layer is handled by the real-time generation model LiveStage. The problem it solves is very specific:

within the compute and VRAM budget of a consumer-grade graphics card, keep the image continuously following real people with low latency.

Autoregression handles continuous evolution over time, and sparse computation limits the update scope of each step, meaning the model predicts state increments rather than an entire new image.

Based on analysis of data from more than 300 real commercial shooting projects, Dawa found that in typical scenes, the area that actually changes per frame accounts for only 10%–20% of the image.

Thus, stable areas such as walls and floors are directly reused through latent space caching (Latent Cache), and compute is concentrated on people, faces, hands, cloth, smoke, lighting and shadows, and interaction areas.

The rendering layer receives a set of conditions issued by the control layer, including event type, position, speed, intensity, object identity, material, and local history, and performs few-step sampling only on the changed areas, then composites the local results back into the persistent world image.

"Causalization" means each step depends only on history that has already occurred, suitable for streaming output;

"micro-diffusion" means processing only a small local area each time, rather than the entire image.

The model has 34B total parameters, activating only about 4B per step, an activation ratio of about 11.8%.

Local updates in space reduce "how much to compute," while sparse activation inside the model controls "how many parameters to use per computation." Together these constitute the feature of "the capacity of a large model, the computational path of real time."

The combined result of the three is: compared with traditional video diffusion approaches, GPU consumption is reduced by 40%–60%, and inference speed is increased by 3–5 times.

When the frame needs to switch as a whole—for example, when the entire room in a screenshot turns into a grassland—the scope of change far exceeds 20%, and the system dynamically allocates computation according to the scope of change; local editing and full-scene transformation share the same pipeline.

Ultimately, on a single NVIDIA GeForce RTX consumer-grade graphics card, Scenes AI achieved real-time output at 720P and 25 frames per second.

Dawa's technical collaboration 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 updates reduce wasted computation, Latent Cache facilitates memory reuse, and the layered architecture naturally lends itself to layered deployment; this mechanism itself is designed around GPU parallel computing and computational reuse.

With the route clear, the real challenge is training it. Dawa's accumulated foundation can be broken down into three layers.

To learn "how the world changes," video alone is not enough.

Most internet video consists of finished, edited products: how the camera moves, how the lighting is set, when props are picked up—almost all of this process information has been cut out, and the pixels themselves cannot explain why the world changes as it does.

As Dawa puts it: it is not a Video Dataset, but a State Dataset.

Dawa operates an 84,000-square-meter fully digital technology studio in Yongchuan, Chongqing.

Over the past five years, the team has continuously collected, annotated, and organized data from more than 300 real commercial projects, accumulating approximately 180TB 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 camera, optics, space, performance, and motion onto the same timeline.

The studio also naturally concentrates the long-tail samples that are hardest to handle in real-time generation: wide sleeves and skirt hems deforming with movement, hands repeatedly contacting props, multiple characters occluding one another, the camera moving continuously during a performance, and lighting changes affecting both characters and environment simultaneously.

Beyond real data, Dawa uses NVIDIA Omniverse and PhysX to build simulation scenes and generate annotated data with precise physical states at scale; and uses NVIDIA Cosmos on H200s to generate "process-type" data such as ice cream melting, trees growing, and snow melting.

The three types of data have clear divisions of labor: real footage provides the distribution, simulation provides precise states, and generated data fills in coverage.

What Dawa is training is not a "video generator," but a continuously evolving world system, divided overall into five stages:

Dawa itself produces real film and short-drama content. Client revisions, feedback from interactions, and cases where the model makes mistakes flow back as entry points for the next round of data selection, supplementary collection, and training—in other words, the model is tested on the studio floor, and the studio floor in turn feeds the model.

"We shoot real films and short dramas every day, and every lighting setup, every camera move, tells the model how this world operates," said Chen Liang, co-founder of Dawa Technology. "The next round of competition in world models belongs to teams that possess real-world processes."

A world model should not be a floating concept; applied to concrete use cases, this route corresponds to several clear scenarios:

They share the same set of requirements: changes must be controllable, output must be timely, and costs must be estimable.

It can run on a single consumer-grade graphics card and a local host without relying on cloud compute—this is what truly marks the step from lab demo to an environment closer to actual deployment.

"Video generation is already powerful enough, but the world has not yet begun to run. A video can end; a world must keep running." At the halfway 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 lab 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."

Looking back at Dawa's path over the past few years, several key pieces are already in place.

The data is there: an 84,000-square-meter studio, more than 300 real projects, 180TB of world-state data—and it is still accumulating;

The model works: a 34B world model runs in real time on a single consumer-grade graphics card, with precise control and real-time rendering both holding true;

Recognition has come too: an outstanding enterprise at the Shenzhen stop in August, a UbiComp paper in October, and in the same month the NVIDIA Halfway Showcase Excellence Enterprise Award.

As the industry's attention begins to shift from "generating a video" to "running a world," this route grown out of the studio floor is being seen by more and more people.

Videos end; the world does not. For Scenes AI, this world has only just begun to run.

English text is being prepared.