WonderZoom:

Multi-Scale 3D World Generation

  • Anonymous Authors

Teaser Image

WonderZoom enables interactive exploration across spatial scales. Users can zoom into any region and specify prompts to generate new fine-scale content while maintaining cross-scale consistency. Here we show three zoom-in sequences

Generated Multi-Scale Worlds

Interactive Viewing

Keyboard: Move by "W/A/S/D", look around by "I/J/K/L". Rotate around the center point by "F". Click the button "Scale 0/1/2" to zoom-in/out.
Note: Clicking an image will load a generated world example from local files. After loading, please click on the canvas to activate control. The rendering here is done on your device in real-time. Due to CVPR supplementary material size limitations, we provide two interactive demos (ladybug and fish) for demonstration. Important: The web renderer used here cannot perform our proposed Level-of-Detail (LoD) rendering and applies depth quantization, which may cause visual inconsistencies across scales that are not present in our GPU-based renderer. The videos above demonstrate the true quality of our method using the full GPU renderer with proper LoD support. Any scale inconsistencies observed in this interactive demo are artifacts of the web renderer limitations, not our method.

Drag & drop .splat/.spz/.ply files
or double-click to open
FOV: 18.9°
Focal: 1441px
FPS: 0
Position:
  X: 0.00
  Y: 0.00
  Z: 0.10
Rotation:
  X: 0.00°
  Y: 0.00°
  Z: 0.00°
Splats: 0
Rendered: ?
Speed Factor: 1.00
Orbit: OFF
Split: OFF
Move
W
Orbit
A
S
D
F
Look
I
J
K
L

Approach

From an input image, we first reconstruct an initial 3D scene. Users specify prompts and camera viewpoints to generate finer-scale content. Our progressive detail synthesizer creates new-scale images, registers depth to maintain geometric consistency, and synthesizes auxiliary views for complete 3D scene creation. Our scale-adaptive Gaussian surfels enable dynamic updates without re-optimization, seamlessly integrating new content while preserving real-time rendering.