FactorSplat
An N-dimensional Gaussian splatting proxy model for medical volume rendering that uses factorization to decouple scene geometry from appearance. It supports real-time switching of transfer functions at inference time without retraining and is applicable to 3D visualization of medical images such as CT and MR scans.
Timeline (last 90 days)
FactorSplat is a per-scene-trained N-dimensional Gaussian splatting (N-DGS) proxy that can directly accept region-specific intensity-to-RGBA curves at inference time, allowing appearance changes without retraining.
FactorSplat decomposes rendering appearance into composable parts: local lookups, a shared functional encoder paired with per-Gaussian low-rank factors, and geometry and directional appearance shared across presets.
FactorSplat's design decouples the geometry component from the color-visibility component, enabling a single checkpoint to support multiple TF edits.
FactorSplat retains visibility control and TF-aware pruning, ensuring that users can still hide or reveal specific anatomical structures when switching transfer functions.
The paper conducted evaluations on 7 CT and MR scan datasets, with the region-aware VEG method as the comparison target.
Testing covered four settings: validation, interpolation, unseen combinations, and out-of-distribution editing.
Across the four splits, the average PSNR improvement over VEG across seven scans ranged from 1.10 dB to 1.52 dB.
Only a single checkpoint per scan is needed to support unseen edits without retraining.
FactorSplat takes only 1.17 milliseconds to switch transfer functions.
As a per-scene-trained method, FactorSplat still requires training a separate model for each scan.
All Facts (10)
FactorSplat is a per-scene-trained N-dimensional Gaussian splatting (N-DGS) proxy that can directly accept region-specific intensity-to-RGBA curves at inference time, allowing appearance changes without retraining.
50%UnverifiedFactorSplat decomposes rendering appearance into composable parts: local lookups, a shared functional encoder paired with per-Gaussian low-rank factors, and geometry and directional appearance shared across presets.
50%UnverifiedFactorSplat's design decouples the geometry component from the color-visibility component, enabling a single checkpoint to support multiple TF edits.
50%UnverifiedFactorSplat retains visibility control and TF-aware pruning, ensuring that users can still hide or reveal specific anatomical structures when switching transfer functions.
50%UnverifiedThe paper conducted evaluations on 7 CT and MR scan datasets, with the region-aware VEG method as the comparison target.
50%UnverifiedTesting covered four settings: validation, interpolation, unseen combinations, and out-of-distribution editing.
50%UnverifiedAcross the four splits, the average PSNR improvement over VEG across seven scans ranged from 1.10 dB to 1.52 dB.
50%UnverifiedOnly a single checkpoint per scan is needed to support unseen edits without retraining.
50%UnverifiedFactorSplat takes only 1.17 milliseconds to switch transfer functions.
50%UnverifiedAs a per-scene-trained method, FactorSplat still requires training a separate model for each scan.
50%