Dawa Technology's Scenes AI Wins the 2026 NVIDIA Startup Showcase Halfway Showcase Excellence Enterprise Award

On October 15, the 2026 NVIDIA Startup Showcase — Halfway Showcase was held at the Suzhou Jinji Lake International Convention Center. Dawa Technology won the Halfway Showcase Excellence Enterprise Award with its real-time world model product Scenes AI. This is the second time Dawa Technology has won an award at an NVIDIA Startup Showcase event, after being named an Outstanding Enterprise at the Shenzhen stop in August.

This year's Halfway Showcase selected 30 companies from more than 580 tech startups across the city stops of the NVIDIA Startup Showcase, which presented roadshows centered on AI agents, physical AI, and AI industry applications. Li Zongxing, Chief Product Officer of Dawa Technology, took the stage on behalf of the team and shared Dawa's technical judgment on real-time world models: video generation answers "what the next frame will look like," whereas a world model must also answer "how the world will evolve after one variable is changed, and which parts should remain unchanged."

The next threshold for video generation: from generating images to controlling the world

Dawa Technology's core positioning is the video generation control layer. It does not hand every problem to a single large model for frame-by-frame regeneration; instead, it adds a control system between creativity and image generation: first, scenes, characters, cameras, lighting, materials, actions, and object relationships are organized into a world state that can be queried, modified, and continuously updated, and then the generation layer renders the parts that need to change.

The key to this architecture is "a heavy control layer and a light generation layer." The control layer is responsible for planning actions, maintaining state, and deducing outcomes; the generation layer is responsible for turning state into video frames. For offline content production, the control layer lets creators modify only the camera position, action, or lighting while preserving other objects and relationships in the same scene; for real-time interaction, the control layer identifies the areas that actually change, and the generation layer updates only those local areas, thereby reducing repeated computation of the entire frame.

To this end, Dawa proposes a new paradigm of "world state maintenance": instead of redrawing the entire frame frame by frame, the model understands the world behind the image, decomposes space, people, cameras, lighting, materials, and physics into states that can be updated independently, and computes only the parts that actually change. Dawa Technology summarizes this technical approach as: build the world once, and it can be filmed, rendered, broadcast, and played.

[Li Zongxing, Chief Product Officer of Dawa Technology, at the Halfway Showcase roadshow]

One consumer-grade graphics card, rewriting a real office in real time

This paradigm has already been validated in real spaces. In September this year, Dawa debuted a live-action demonstration of Scenes AI: a camera pointed at an ordinary office continuously received brand-new prompts and generated multiple scenes in real time. All core generation computation was carried by a single local NVIDIA RTX graphics card. The office could become an animated grassland, and snow monsters, flames, and new furniture forms could appear; but every change followed the same world state rather than redrawing the entire frame from scratch. The running frame rate was displayed in real time in the upper right corner of the image, and the entire process was "no presets, no stuttering, 100% real-time continuous generation."

In the video, the floor collapses, a snow monster breaks through the ground and climbs the wall to leave, yet the pit in the ground remains in place; the wrought-iron chair beside it first turns into a red spiked seat and then into a wolf-headed iron throne, while the host can continue sitting in the same position; the hem of the curtain ignites with open flame, and the fire spreads upward along the fabric, yet the doll and flowerpot right next to it remain unharmed; finally, the entire office transforms into an animated grassland, with the air conditioner, curtains, and cabinet all remaining in place.

Behind this are the three hardest challenges of real-time generation: inference latency, error accumulation, and dynamic scene control. The camera continuously inputs new frames, and the person's actions change the generation conditions at any time; the model must determine, within limited compute and video memory, which content needs updating and which states should be preserved; and any deviation in facial appearance or spatial structure may be continuously amplified in subsequent frames. For Dawa, real-time performance is not simply compressing an offline video model to be faster; rather, the control layer first gives the world state and control, and then rendering computation occurs as events happen.

[The Scenes AI real-time demonstration turns the entire office into an animated grassland]

Latent Stage precise control: three key technologies that bring a 34B world model into the live stream

Scenes AI's underlying model, Latent Stage, has a total of 34B parameters, with about 4B activated per step. Running a model of this scale continuously in real time on a single consumer-grade graphics card relies on three self-developed technologies:

• Sparse autoregressive diffusion + latent space cache: compute only the parts that are changing. Based on analysis of Dawa's 300+ real commercial shoots, in typical scenes the area that actually changes per frame accounts for only 10% to 20% of the image. Scenes AI predicts "state increments" only for these areas, while stable areas such as walls and floors are directly reused through a latent space cache (Latent Cache). Local updates in space, combined with sparse activation inside the model, reduce GPU consumption by 40% to 60% compared with traditional video diffusion schemes and increase inference speed by 3 to 5 times.

• Latent Stage + World Memory: change one place, everything else stays the same, and it remembers. Latent Stage decomposes scenes, characters, cameras, lighting, clothing, props and physics into independently controllable variables in latent space; World Memory continuously preserves the changes that have already occurred. In continuous editing tests, the target edit success rate reached 98.2%, and the character identity preservation rate reached 93.3%. After the snow monster leaves, the pit remains—precisely because the model "remembers what has happened in the world."

• Neural physics + neural rendering: from generating images to simulating physics. Neural physics constrains generation around motion, collision, gravity and material response, while neural rendering makes firelight and shadows act on the environment in sync. The model does not just generate flames; it also determines "what will burn and what will not."

In addition, the non-linear visual matrix proposed by Dawa organizes characters, environments, cameras, lighting, objects and motion into visual elements that can be recombined. The related paper on NVD technology has been accepted by the international pervasive computing conference UbiComp / ISWC 2026.

Live Stage real-time rendering: the control layer decides where changes occur, and local diffusion is responsible for generating those changes

Scenes AI's real-time capability is handled by the real-time generative model Live Stage. It adopts a causal sparse micro-diffusion design: rather than performing full-frame diffusion on the entire image, it receives conditions such as events, position, velocity, intensity, object identity, material and local history emitted by the control layer, performs few-step sampling on the changed regions, and then composites the local results back into the persistent world.

A typical runtime pipeline is:

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

For example, when a character touches a curtain, the system first uses the rule and event layer to identify the contact position, motion direction and material conditions, and then the real-time diffusion model handles cloth deformation and contact boundaries; stable walls, floors and unaffected objects reuse their existing representations. After the character leaves, short-lived fabric sway can end, while results that need to be retained long-term, such as footprints in snow and ground depressions, are written back to the world state for subsequent shots and interactions to continue reading.

In typical scenes, the regions that actually change account for about 10%–20% of the image. Local updates, sparse expert activation and latent space caching together reduce repeated inference, bringing real-time generation closer to the computational approach of traditional rendering.

The dataset that truly determines the training quality of the next generation is not a video dataset—it is a state dataset

Unlike most models that learn from internet video, Scenes AI's training data comes from real film and television production. Dawa operates an 84,000-square-meter full-process digital technology studio in Yongchuan, Chongqing. From more than 300 real commercial projects, it has accumulated about 180TB of structured world state data, covering camera, motion, object, lighting, physics and narrative states.

This kind of data directly corresponds to the long-tail situations that real-time world models find hardest to handle: wide sleeves and skirts 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 the environment at the same time. Internet video can usually only tell the model "what the final image looks like," whereas real studio data also records "how the camera shoots, how the actors move, how objects make contact, and why the director made a particular change." These are precisely the areas where real-time generation is most prone to drift, penetration and inconsistency, and they are also where the greatest value of Dawa's data lies. The dataset that truly determines the training quality of the next generation is not a video dataset—it is a state dataset.

Over the past five years, Dawa has continued to collect, annotate and organize this data. In a single scene, more than 300 dimensions of structured annotation are performed, so that camera, optics, space, performance and motion can be aligned on the same timeline. These process data serve world construction, offline video generation and real-time local diffusion respectively, and enter creative and interactive scenarios through FilmOS and Scenes AI. Customer revisions, interaction feedback and model failure cases further provide entry points for subsequent data screening, supplementary collection and training.

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

"NVIDIA provides a solid computing foundation for real-time world models," said Li Zongxing, Chief Product Officer of Dawa Technology. "We hope to work with the NVIDIA ecosystem to bring world models out of the lab and 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."

From real film sets to the international academic stage

Dawa Technology's technical approach has also gained academic recognition. On October 13, the paper "NVD: Non-Linear View Matrix Fusion for 2D Video Using Multimodal Foundation Models," jointly completed by the Dawa Technology AIGC Research Institute with The Hong Kong Polytechnic University, the University of Illinois Urbana-Champaign, and Newcastle University, was accepted by ACM UbiComp / ISWC 2026 and published in the ACM UbiComp Companion '26. The paper focuses on the virtual production challenges in real film and television production, proposing a non-linear view matrix fusion method based on multimodal foundation models that incorporates video, depth, audio, and motion capture information into a single computational framework, and validates multi-view real-time visual synthesis on consumer-grade NVIDIA RTX hardware and ordinary cameras. Comparative experiments reported in the paper show that across ten categories of virtual production tasks—including scene replacement, relighting, recomposition, and digital doubles—the system reduces hardware costs by 72% and shortens deployment time by 63%.

The significance of this research lies in translating the industrial problems Dawa has long faced into reproducible research questions: how to understand the relationships among camera, characters, motion, lighting, and environment; how to keep fusion results stable during camera movement and character motion; and how to move professional virtual production from dedicated soundstages and server clusters toward lighter-weight combinations of software and hardware. At the ACM UbiComp / ISWC 2026 demonstration site, scholars and engineers from Tsinghua University, ShanghaiTech University, The Hong Kong University of Science and Technology, and other universities exchanged views on the system architecture and real-time demonstrations. For Dawa, academic papers provide external validation at the methodological and experimental level, while real film sets provide the problems, data, and scenarios for continuous iteration; together, the two support its technical roadmap from video generation to video generation control foundation models and real-time world models.

Built natively on the NVIDIA technology stack

As early as 2018, Dawa had already adopted NVIDIA technology. That year, Dawa purchased its first NVIDIA DGX-1 system, and it has since conducted model research and development on NVIDIA computing platforms.

Today, Scenes AI is built natively on the NVIDIA ecosystem: NVIDIA CUDA accelerates model training and inference, NVIDIA TensorRT optimizes inference, and NVIDIA NVENC supports real-time video output. Dawa's sparse update and state caching mechanisms are themselves designed around the parallel computing and computational reuse of NVIDIA GPUs. A single GeForce RTX graphics card can handle local editing, physical simulation, and full-scene transformation in real time, providing a clear performance and cost reference for film and television studios deploying real-time world models.

For film and television studios, livestreaming teams, and interactive content developers, continuous operation on consumer-grade graphics cards means real-time world models can move from laboratory demonstrations to hardware environments closer to actual deployment. Dawa Technology hopes that through Scenes AI and FilmOS, it can extend the same world state to professional shooting, offline rendering, livestreaming output, and interactive experiences, enabling video generation to move from "generating a result" to "continuously running a world."

About Dawa Technology

Dawa Technology is a generative world foundation model company that builds controllable, deliverable, lightweight, and easily deployable AI video generation infrastructure, headquartered in Yongchuan, Chongqing. Leveraging high-quality exclusive process datasets formed through real film and television production, the company has built three major model systems—Latent Stage, CineStage, and LiveStage—and connects professional film and television, short dramas, livestreaming, marketing, and other creative and commercial scenarios through FilmOS and Scenes AI, realizing "build a world once, and you can shoot it, render it, broadcast it, and play with it," driving continuous iteration across models, products, project delivery, and data feedback.