DAWA says ScenesAI runs at 720p and 25 fps on one RTX GPU
DAWA describes ScenesAI as a real-time world model that can understand a space from one camera and edit objects, materials or events while leaving the rest of the frame untouched. The company says it runs on one NVIDIA RTX GPU, with Live Stage performance of 720p at 25 frames per second.

Key takeaways · 5
- 01
DAWA lists Live Stage performance at 720p and 25 frames per second on one NVIDIA RTX GPU.
- 02
The model updates changing regions rather than regenerating the whole frame; DAWA says typical scenes change in 10% to 20% of each frame.
- 03
DAWA reports 40% to 60% lower GPU use and three to five times faster inference than conventional video diffusion.
- 04
The company reports a 98.2% target-edit success rate and 93.3% identity retention.
- 05
DAWA invites film studios, live-streaming teams and interactive-content developers to request a demo and evaluation by email.
A model built around local edits
DAWA says ScenesAI understands a real space from a single camera and can rewrite objects, materials and events without changing the rest of the frame.[1] Its stated approach updates only the regions that change, rather than regenerating every pixel in each frame.[1] DAWA says a typical scene changes in just 10% to 20% of each frame.[1] For teams assessing the system, this makes the distinction between local edits and full-frame generation central to understanding the product’s design. The company describes ScenesAI as a real-time world model.[1]
Performance and quality claims
DAWA lists ScenesAI Live Stage at 720p and 25 frames per second, and says the model runs in real time on one NVIDIA RTX GPU.[1] Compared with conventional video diffusion, the company reports 40% to 60% lower GPU use and inference that is three to five times faster.[1] DAWA also reports a 98.2% target-edit success rate and 93.3% identity retention.[1] These are company-reported figures; professionals evaluating the tool can treat them as benchmarks to test against their own footage and production requirements.
Continuity across frames
DAWA says ScenesAI has produced 3,600 consecutive consistent frames and describes one office undergoing eight consecutive rewrites in 63 seconds.[1] The company says changes that actions make to an environment carry into every later frame.[1] It attributes motion, collision, gravity and material response constraints to neural physics and neural rendering.[1] Those claims point to continuity as a feature to examine alongside image quality: teams considering a workflow can assess whether edits remain consistent as scenes and actions evolve.
Target users and company context
DAWA says ScenesAI is intended for film studios, live-streaming teams and interactive-content developers, and invites those groups to email for a demo and evaluation.[1] The company lists 180 terabytes of structured world-state data and more than 300 commercial productions.[1] DAWA also says its Yongchuan digital studio covers 84,000 square meters.[1] At an NVIDIA Inception Showcase mid-term event in Suzhou on October 15, 2026, DAWA received the Excellence Award; the event selected 30 companies from more than 580 applicants to present.[2]
For teams exploring AI-assisted video, the main question is whether local edits, frame continuity and the reported GPU performance hold up on their own material. The stated one-GPU setup and performance figures give production and technical teams concrete claims to evaluate before seeking a demo.
Why it matters
Bring this to your team — one session a day, built for your industry.
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
For teamsHow this developed
24 October 2026
DAWA says ScenesAI runs at 720p and 25 fps on one RTX GPU
Sources
- ScenesAI — Real-Time World Model by DAWAscenesai.cn
- 一张5090跑起34B实时世界模型,国产玩家拿下英伟达大奖! – 量子位qbitai.com
- 永川达瓦科技凭AI世界模型获NVIDIA卓越企业奖-中新网重庆cq.chinanews.com.cn