SCION
A compositional 3D scene representation method proposed by Princeton University that constructs a reusable vocabulary of neural primitives and lightweight world-space instances, enabling compact and efficient neural scene representation and rendering with support for instance-level scene editing and animation.
Timeline (last 90 days)
SCION (Scene Composition with Instanced Neural Primitives) was proposed by a team from Princeton University.
SCION replaces mutually independent Gaussian primitives with a compact, reusable vocabulary of primitives, and then uses lightweight world-space instances to place transformed copies of the primitives throughout the scene.
SCION adopts a joint optimization strategy for discrete and continuous parameters, combined with two-level densification at both the splat and instance levels.
SCION introduces an adversarial loss to preserve fine details across shared primitives, preventing details from being averaged out and blurred when multiple instances share a template.
SCION maintains relatively high rendering quality even when compressed to an extremely small size of 1.2 MB.
SCION outperforms existing Gaussian compression methods in rate-distortion performance.
SCION's representation supports instance-level scene editing and animation without requiring retraining.
The SCION project homepage is publicly available at the Princeton Computational Imaging Lab (light.princeton.edu/SCION).
All Facts (8)
SCION (Scene Composition with Instanced Neural Primitives) was proposed by a team from Princeton University.
50%UnverifiedSCION replaces mutually independent Gaussian primitives with a compact, reusable vocabulary of primitives, and then uses lightweight world-space instances to place transformed copies of the primitives throughout the scene.
50%UnverifiedSCION adopts a joint optimization strategy for discrete and continuous parameters, combined with two-level densification at both the splat and instance levels.
50%UnverifiedSCION introduces an adversarial loss to preserve fine details across shared primitives, preventing details from being averaged out and blurred when multiple instances share a template.
50%UnverifiedSCION maintains relatively high rendering quality even when compressed to an extremely small size of 1.2 MB.
50%UnverifiedSCION outperforms existing Gaussian compression methods in rate-distortion performance.
50%UnverifiedSCION's representation supports instance-level scene editing and animation without requiring retraining.
50%UnverifiedThe SCION project homepage is publicly available at the Princeton Computational Imaging Lab (light.princeton.edu/SCION).
50%