SCION: Scene Composition with Instanced Neural Primitives

William Koch; Amogh Joshi; Cyrus Vachha; Cheng Zheng; Felix Heide

cs.CVOct 1, 2026

SCION: Scene Composition with Instanced Neural Primitives

Authors: William Koch, Amogh Joshi, Cyrus Vachha, Cheng Zheng, Felix Heide

Organizations: Princeton University

Abstract

Real-world scenes are compositional: bricks, blades of grass, pebbles, and tree leaves recur across human-built and natural environments. Existing neural scene representations model these elements independently. Most 3D Gaussian Splatting and follow-up abstraction and compression methods treat each element as unique, fitting millions of independent Gaussians per scene. Prior methods like Splat and Replace fit template objects, but they require mostly manual selection of repeated elements. As a result, these representations store redundant parameters and provide weak manipulation handles for downstream tasks. We introduce SCION, a hierarchical compositional scene representation that replaces independent Gaussians with a compact vocabulary of reusable primitives and lightweight world-space instances that place transformed copies throughout the scene. We fit this representation to multi-view captures via a joint optimization over discrete and continuous scene parameters, combining two-level densification over splats and instances with an adversarial loss that preserves detail across shared primitives. The recovered structure yields a compact, controllable representation while maintaining high quality even at 1.2 MB. SCION achieves rate-distortion favorable to existing Gaussian compression methods, and it enables instance-level scene editing and animation without retraining. Our results show that neural scene representations need not memorize scenes as independent primitives; they can discover reusable parts. Project webpage: https://light.princeton.edu/SCION

Figures & tables

Figure 1: Overview of SCION . We initialize instances from a sparse point cloud, group them into reusable primitives, assemble the selected local splats for rendering, and update the representation through image losses, projected adversarial supervision, and active densification.

Panel of Figure 2 — Figure 2: Neural primitive assembly. The (a) learned primitives are (b) instantiated multiple times throughout the scene, resulting in (c) assembled primitive structures that (d) construct the scene.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$
ContextGS	20.79	0.559	0.511	1.760
HAC++	26.51	0.756	0.299	1.699
OMG	26.25	0.770	0.271	1.353
Ours-Small	22.01	0.605	0.431	1.204

Table 1: Matched-storage comparison on MipNeRF360. Each baseline is rerun at the lowest available compression point point near our method.

Figure 3: Qualitative matched-storage comparison. At approximately 1.5 MB, SCION retains coherent repeated structure as each reused primitive is represented once and instantiated many times.

Ablation	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$
	Ours-Small ( $n{=}3$ )				Ours-Large ( $n{=}3$ )
Splat DC only	$21.98_{\pm 0.07}$	$0.56_{\pm 0.00}$	$0.45_{\pm 0.01}$	1.48	$23.23_{\pm 0.02}$	$0.65_{\pm 0.00}$	$0.34_{\pm 0.00}$	3.81
$+$ instance DC	$21.87_{\pm 0.15}$	$0.56_{\pm 0.01}$	$0.46_{\pm 0.02}$	1.45	$23.20_{\pm 0.05}$	$0.65_{\pm 0.00}$	$0.34_{\pm 0.00}$	3.76
$+$ LPIPS	$21.90_{\pm 0.08}$	$0.56_{\pm 0.00}$	$0.45_{\pm 0.01}$	1.45	$23.09_{\pm 0.07}$	$0.65_{\pm 0.00}$	$0.33_{\pm 0.00}$	3.76
Full ( $+$ adv)	$21.74_{\pm 0.12}$	$0.55_{\pm 0.01}$	$0.37_{\pm 0.02}$	1.44	$22.88_{\pm 0.02}$	$0.63_{\pm 0.00}$	$0.27_{\pm 0.00}$	3.72

Table 2: Component ablation. Rows cumulatively add the active SCION components on the MipNeRF360 garden anchor scene. “DC” stands for density control.

Figure 4: Component ablation renders. Example render of the Ours-Large variant of our method on the Garden scene from MipNeRF360. “DC” stands for density control.

Figure 5: Animation. Flow maps for the same wind animation applied to (a) 3DGS and (b) SCION . Per-splat transformations in 3DGS produce incoherent local motion, while SCION moves repeated scene elements coherently through instance transforms.

Explore similar work

ZipSplat: Fewer Gaussians, Better Splats
Jun 3, 2026Alexander Veicht, Sunghwan Hong, Dániel Baráth +1Feed-Forward 3D Gaussian Splatting Gaussian Splatting
ETH Zürich · Microsoft
MLP Splatting: Object-Centric Neural Fields
Jun 2, 2026Shinjeong Kim, Yuzhou Cheng, Xin Kong +2Novel View Synthesis 3D Representation
Department of Computing, Imperial College London
Compact Feed-Forward 3D Gaussians via Saliency-Guided Primitive Merging
Aug 11, 2026Tim-Felix Faasch, Jochen Kall, Cyrill Stachniss3D Gaussian Feed-Forward
Bosch Research Hildesheim, Germany · University of Bonn Bonn, Germany · Lamarr Institute for Machine Learning and Artificial Intelligence Bonn, Germany

Jun 3, 2026, cs.CVJ/K move · Enter open · S save

ZipSplat: Fewer Gaussians, Better Splats

Alexander Veicht, Sunghwan Hong, Dániel Baráth, Marc Pollefeys

ETH Zürich · Microsoft

Feed-Forward 3D Gaussian Splatting Gaussian Splatting Feed-Forward 3D Gaussian Gaussian Primitives Geogs-Slam

Open paper AskPDF arXiv

Feed-forward 3D Gaussian Splatting methods reconstruct a scene from posed or pose-free images in a single forward pass, yet current approaches predict one Gaussian per input pixel, tying the representation budget to camera resolution rather than scene complexity. A flat wall and a richly textured object thus produce equally many Gaussians despite very different geometric needs. We propose ZipSplat, a token-based feed-forward model that decouples Gaussian placement from the pixel grid. A multi-view backbone extracts dense visual tokens, and k-means clustering compresses them into a compact set of scene tokens. Cross- and self-attention refine these tokens, and a lightweight MLP decodes each into a group of Gaussians with unconstrained 3D positions. Because clustering is applied at inference, a single trained model spans the quality-efficiency curve without retraining. ZipSplat operates without ground-truth poses or intrinsics, yet sets a new state of the art on DL3DV and RealEstate10K with

{\sim}6{\times}

fewer Gaussians than pixel-aligned methods, surpassing the best pose-free baseline by 2.1dB and 1.2dB PSNR, respectively. It further generalizes zero-shot to Mip-NeRF360 and ScanNet++, outperforming all comparable baselines. Our project page is at https://veichta.com/zipsplat.

Abstract

Explore similar work

ZipSplat: Fewer Gaussians, Better Splats

MLP Splatting: Object-Centric Neural Fields

Compact Feed-Forward 3D Gaussians via Saliency-Guided Primitive Merging