loading
loading
Dawa Technology, a Chongqing company that grew out of virtual film production, has detailed Scenes AI, a real-time world model that takes video from an ordinary camera and turns the room in front of it into an editable, persistent world. The product won an excellence award at the NVIDIA startup showcase mid-term event in Suzhou on October 15. Thirty NVIDIA Inception members reached that stage from more than 580 applicants. Most video models try to predict what the next frame should look like. Scenes AI instead maintains a running world state. Scene, people, camera, lighting, objects and physics are each stored as variables that can be read and written separately. When something changes, the model updates only that part of the state and writes it back, instead of redrawing the whole picture. In Dawa's demos, a room changes into a hand-drawn meadow while the chair, a plush bear and a cabinet keep their positions. Words written on a wall stay in place while people come and go and the weather turns from snow to rain. The system has a heavy control layer and a light generation layer. The control layer, built around a module called Latent Stage, separates scene, character, camera, lighting, costume and props in latent space and applies physics constraints for rigid bodies, cloth, fluids and particles. The rendering layer, LiveStage, uses causal sparse diffusion to sample only the regions that change. The model has 34B parameters in total but activates about 4B per step, around 11.8 percent. Dawa says this cuts GPU use by 40 to 60 percent and speeds up inference three to five times compared with conventional video diffusion. On a single consumer NVIDIA GeForce RTX card, it reaches real-time output at 720p and 25 frames per second. In continuous-editing tests, the company reports a 98.2 percent success rate for targeted edits and 93.3 percent identity retention for people. The system uses CUDA, TensorRT and NVENC, and it runs on a local machine without cloud compute. Dawa says its data comes from 84,000 square meters of studio space, more than 300 real productions and 180TB of world-state data. Its chief product officer, Li Zongxing, summed up the approach this way: a video can end, but a world has to keep running. All performance figures are company-reported. SEE ALSO: VAST Open-Sources EdenLive · Qianjue Unveils Polibrain
document fetched 2026-10-28 13:04