From data captured on working film sets, to understanding and simulating world state, to offline content generation and real-time interaction.
A video can end. A world has to keep updating, remembering and responding — and stay physically consistent.
ScenesAI is DAWA's real-time world model. Instead of treating a world as a video waiting to be generated, it organizes space, people, objects, motion, lighting, physics and history into a world state that is continuously updated.
Persistent World = Memory + Physics + Sparse Diffusion + Rendering + AI Lighting
In a continuous world, only 10–20% of each frame actually changes — people, faces, hands, fabric, smoke and fire. Static regions reuse the latent cache; only the state delta is generated locally.
In a single latent space, scene, characters, lighting, camera, wardrobe and props, and physics are separated into stages that evolve independently under joint constraints. Turn the light to sunset, and the set, the cast and the camera stay frozen.
Live-action footage is decomposed into six editable components — character, environment, camera, lighting, objects and motion — and recomposed into new worlds. Used on 20+ major film and TV productions at Yongchuan Tech Film Studio.
From an ordinary camera feed to 3D skeletons and spatial relations — and from an intent like “walk to the sofa and sit down” to motion sequences and character states that respect spatial boundaries and physical constraints.
Neural Physics solves rigid bodies, cloth, fluids, particles, collisions and gravity, and outputs world state. The Neural Renderer combines geometry, materials and lighting to turn that state into real-time frames.
The question shifts from “how good is the generated video” to “how well does it generate a world in real time”. These are the evaluation dimensions and target metrics for SCENESAI.
Not a video dataset — a state dataset. Multimodal production data from Yongchuan Tech Film Studio is captured, cleaned, annotated and structured into 180 TB of world-state data.
Camera paths, focal lengths, optical parameters and shot language, recorded in sync with real shoots.
Cumulative capture volume, licensable data and deliverable datasets are reported separately. Data samples, dataset partnerships and capture services are available on request.
Worlds you can understand, generate and edit.
Built around LiveStage: persistent world state, latent perturbation with sparse diffusion, real-time video generation and lighting, neural physics and digital twins.
An AI film crew for everyone.
Script breakdown, storyboarding, scheduling, set building, virtual previs, shooting and post — each stage handled by collaborating agents, cutting total cycles from 6–12 months to hours or days.
Motion as a universal interface for intelligence.
A 1B-parameter motion foundation model that turns camera footage into 3D motion in real time, driving MetaHuman, Unreal Engine, Unity, XR and robots.
Every Salesperson Has an AI Partner.
A domain-knowledge model with long-term, cross-session memory, covering outreach, conversation, intent detection, follow-up reminders and content. Featured by NVIDIA.
Understands the structure of a world and builds spaces and states that keep evolving.
Professional content generation and creative workflows for the film industry.
A real-time world model connected to the physical world.
One stack underneath: CUDA · TensorRT · multimodal inference · distributed training
Our data comes from the working film industry, not lab demos. From national projects to cinematic virtual production, every real shoot adds simulation, spatial, motion and camera data.
Virtual production. Related titles include Water Gate Bridge and Panda Plan.
With R&D centers in Beijing and Guangzhou, our work appears at international venues in computer graphics, ubiquitous computing and human–computer interaction.
DAWA ScenesAI Team · Submission ID real_110s1
Live demo on December 4; program details to be announced
Ze Gao, Zongxing Li, Jinda Han, Yiran Zhang, Mengyao Guo
Included in ACM UbiComp Companion ’26 · DOI pending
Mengyao Guo, Yu Nie, Zongxing Li, Jinda Han, Ze Gao
Mengyao Guo, Yu Nie, Ze Gao
Coverage from NVIDIA's official blog, Xinhua, People's Daily, CCTV and other national and regional media, plus government sources. Every item links to the original.
At the mid-term round in Suzhou on October 15, DAWA won the Excellence Award for its Scenes AI real-time world model. Thirty companies were selected from more than 580 to present; it is DAWA's second award at the Showcase this year, after Outstanding Company in Shenzhen in August.
Explains the 34B-parameter (4B active) Scene Foundation Model and sparse micro-diffusion behind LiveStage, with 25 FPS and 3,600 consecutive consistent frames.
On August 5, DAWA presented the LiveStage real-time world model at the 2026 NVIDIA Inception Showcase in Shenzhen and received an Outstanding Company certificate.
Yongchuan District's official account announces DAWA's 2026 NVIDIA Inception Showcase award for LiveStage.
An NVIDIA case study: DealCat runs LLM inference on NVIDIA DGX Spark to cut conversation latency and improve concurrency and memory retrieval.
A world model isn't a single-model problem: it takes science, engineering, product and industry working together. Our core engineers come from Tencent, ByteDance, Alibaba, Huawei and others.
