ScenesAI

DAWA's real-time world model. It understands a real space from a single camera, rewrites objects, materials and events on instruction, and leaves the rest of the frame untouched. Runs in real time on one NVIDIA RTX GPU.

Click the chair, curtain, pot or floor in the scene

This 3D scene is reconstructed from recording frames and depth estimation to illustrate local updates. The recording below shows the model's actual output.

Scroll →

Unedited recording

One office, eight consecutive rewrites in 63 seconds. Click a thumbnail or type an instruction to jump to that moment.

00:00:00·Changed—

No rewrites yet

More rewrites

A second recording, 193 seconds.

All clips are taken from the same recording. The Chinese subtitles are part of the original recording.

Applications

ShootReal-time previs

Actors perform in a white studio while the frame is rebuilt live as an alpine meadow and a beach. Motion and camera come from the real shoot.

RenderScene rebuild

The walls are rebuilt in wood, the curtain burns from the bottom up, petals cover the floor and a cherry tree grows, with everything else left in place.

StreamLive costume change

A host performs in an ordinary office; the character, costume and garden are generated and streamed directly through NVIDIA NVENC.

PlayProp control

A lightsaber is generated in the actor's hand and follows the swing. Only that part of the frame is diffused; the desert and the actor stay unchanged.

How it works

Video generation repaints the whole frame every time, including walls and floors that never change. ScenesAI splits the scene into world state that updates independently, and hands only the changing regions to the generation model.

Video generationRepaints every frame

ScenesAIUpdates changed regions only

Regenerated this frame 100%

Regenerated this frame —

↔

Drag the dividerThe right-hand value is computed live from pixel differences between consecutive frames of this recording.

10–20%

Of each frame changes in typical scenes

40–60%

Less GPU

3–5×

Faster inference

Compared with conventional video diffusion

How one contact event is handled: a person touches the curtain

World state

Space, people, camera, lighting, materials and physics are separate variables that can be queried and edited independently.

Latent Stage

World building and precise control

34B parameters, ~4B active per step
98.2% target-edit success
93.3% identity retention

Live Stage

Real-time local generation

720p · 25 FPS
3,600 consecutive consistent frames
One NVIDIA RTX GPU

Underneath, neural physics and neural rendering constrain motion, collision, gravity and material response, and firelight and shadows act on the environment. The system is built on CUDA, TensorRT and NVENC.

“We shoot real films and short dramas every day. Every lighting setup and every camera move tells the model how the world works. The next race in world models belongs to the teams that own real-world process.”

Chen Liang, Co-founder, DAWA

Training data from working film sets

Internet video keeps only the final picture. Set data also records camera motion, blocking, object contact and every adjustment the director makes, which is exactly where real-time generation tends to fail.

180TB

Structured world-state data

300+

Commercial productions

300+

Annotation dimensions per scene

84,000m²

Fully digital studio in Yongchuan

Awards and papers

Request ScenesAI early access

For film studios, live-streaming teams and interactive-content developers. Tell us about your scene by email and we will arrange a demo and evaluation.

Request access→Partnerships