TranSplat: Instant Object Relighting in Gaussian Splatting via Spherical Harmonic Radiance Transfer

Boyang (Tony) Yu *, Yanlin Jin *, Yun He, Akshat Dave, Ravi Ramamoorthi, Guha Balakrishnan · 2025

Read TranSplat: Instant Object Relighting in Gaussian Splatting via Spherical Harmonic Radiance Transfer by Boyang (Tony) Yu *, Yanlin Jin *, Yun He, Akshat Dave, Ravi Ramamoorthi, Guha Balakrishnan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22676.txt.

Introduction

Yu, Ramamoorthi, Balakrishnan and their co-authors present TranSplat, a fast way to relight objects in Gaussian Splatting scenes in 2025. Older relighting needs costly inverse rendering or retraining, which slows editing work.

TranSplat works as a post step and shifts the light data of each small surface piece using ratios from source and target light maps. A dual path handles glossy, view-dependent looks, and a light module adds self-shadowing without mesh tracing. Tests on fake and real objects beat recent methods in most settings. Each relight took under one second, the authors report.

Catalog record

CREATORBoyang (Tony) Yu *, Yanlin Jin *, Yun He, Akshat Dave, Ravi Ramamoorthi, Guha Balakrishnan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-22676

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TranSplat: Instant Object Relighting in Gaussian Splatting via Spherical Harmonic Radiance Transfer Boyang (Tony) Yu *, Yanlin Jin *, Yun He, Akshat Dave, Ravi Ramamoorthi, Guha Balakrishnan Abstract We present TranSplat , a method for instant, accurate object relighting within the Gaussian Splatting (GS) framework. Rather than relying on costly inverse rendering routines, we propose a BRDF-free radiance transfer strategy that analytically modulates the spherical harmonic (SH) appearance coefficients of an object’s 2D Gaussian surfels using per-normal irradiance ratios derived from source and target environment maps. To handle view-dependent and glossy appearances without explicit material estimation, we introduce a specularity-aware dual-path SH transfer strategy that adapts higher-order SH bands in the reflection domain. Additionally, we propose a lightweight SH-domain self-shadowing module to ensure physically realistic occlusion without explicit mesh raycasting. Operating as a post-processing step, TranSplat requires no additional GS retraining for a pair of source and target scenes. Evaluations on synthetic and real-world objects demonstrate state-of-the-art accuracy, outperforming recent inverse-rendering and diffusion-based GS relighting methods across most conditions, all while completing relighting operations in under one second. Although bounded by radially-symmetric BRDF approximations and the low-pass nature of the SH basis, TranSplat produces perceptually realistic renderings even for glossy, complex materials, establishing a valuable, lightweight path forward for GS relighting. I Introduction The Gaussian Splatting (GS) framework has demonstrated remarkable power and computational speed for high-quality, real-time neural rendering applications [ 1 ] . Building on these rendering capabilities, there is a growing demand for realistic 3D scene editing and composition applications with GS, such as object relighting. Given a pretrained Gaussian asset alongside high dynamic range (HDR) environment maps for both the source and target environments, the goal of object relighting is to re-render the object so that its shading and color are consistent with the target illumination settings. The key challenge of this task is that the object’s perceived illumination must change as a function of both environments in a physically consistent manner, without being given any knowledge of its material properties. Several recent GS relighting methods offer solutions to this problem, but rely on computationally heavy paradigms. These include explicit inverse rendering pipelines that attempt to explicitly decompose the scene into geometry, materials (BRDFs), and lighting, which requires slow, iterative per-scene optimization [ 2 , 3 , 4 ] . Alternatively, some methods rely on generative diffusion priors that distill 2D image-space relighting into 3D representations. These often suffer from low resolution, multi-view inconsistencies, and slow inference [ 5 ] . For example, even the fastest of these optimization-based inverse rendering methods [ 3 , 6 ] take approximately 60 minutes to jointly optimize and relight a typical scene. This computational burden precludes real-time applications such as gaming, VR, and interactive 3D scene editing. We propose TranSplat , a novel framework that instantly and realistically relights off-the-shelf Gaussian assets in novel target environments, running in under 1 second on an NVIDIA A6000 GPU. To achieve this, we take inspiration from a theoretical identity relating the spherical harmonic (SH) coefficients of surfaces with radially symmetric BRDFs under different lighting conditions [ 7 ] . For diffuse surfaces, this identity states that relighting reduces to multiplying the SH appearance coefficients by the per-normal irradiance ratios between the target and source environments. Crucially, this radiance transfer bypasses explicit BRDF estimation, avoiding a major computational bottleneck. While radiance transfer has been leveraged in classical 2D ratio image methods [ 8 , 9 ] , it has not been integrated into GS frameworks for lightweight relighting. Our key insight is that the spherical harmonic coefficients of a standard pretrained GS asset, despite entangling source illumination, material response, and view-dependent appearance, can be directly transformed to fit the target environment without retraining or explicit BRDF decomposition. TranSplat introduces three main technical contributions. First, we formulate an analytical diffuse irradiance transfer approach in the GS framework. We enable this by building upon the 2D Gaussian Surfels representation [ 10 ] , exploiting its high-quality, flat surface normals which are essential for accurate, per-surfel irradiance computation. Second, we extend this formulation beyond the diffuse case. By performing a reflection-domain reparameterization, we extract an approximate BRDF attenuation profile directly from the pretrained SH coefficients, allowing us to modulate glossy appearances without parametric material estimation. Third, we introduce an SH-domain self-shadowing module utilizing smooth continuous visibility functions and the Gaunt tensor, ensuring physically realistic occlusion without explicit raycasting. The entire pipeline operates as a single, closed-form analytical pass over the Gaussian assets. To our knowledge, TranSplat is the first framework to enable real-time relighting of off-the-shelf Gaussian assets. While this lightweight method makes certain simplifications at the cost of modeling higher-order, subtle material details, it achieves fast, perceptually realistic relighting of both diffuse and glossy objects directly in the SH domain. We quantitatively evaluate TranSplat on several synthetic objects. TranSplat achieves superior relighting accuracy compared to baseline inverse-rendering [ 3 , 11 , 2 , 6 ] and diffusion-based [ 5 ] methods, all while operating in under one second on a GPU. We further demonstrate its robustness through qualitative cross-scene object insertions on real-world captures under diverse lighting. Ultimately, despite relying on radially symmetric BRDF approximations and a low-pass SH basis, TranSplat presents a highly practical trade-off, prioritizing instant, perceptually realistic relighting over the heavy computational costs of explicit modeling. II Related work Neural-Based Inverse Rendering and Radiance Transfer. Classical inverse rendering, the estimation of physical scene properties from images, relies on iterative optimizations using physics and geometry [ 12 , 13 , 14 , 15 ] , often requiring controlled setups or known scene properties due to the ill-posed nature of the task. Modern inverse rendering approaches leverage deep learning and neural representations to decompose geometry, materials, and lighting from unstructured images. Early methods used feed-forward networks trained on large datasets to predict material attributes [ 16 , 17 , 18 ] . Neural Radiance Fields (NeRFs) [ 19 , 20 , 21 ] introduced per-scene optimization via MLP encodings, enabling methods like NeRF-in-the-Wild [ 22 ] to separate transient lighting from constant scene properties. Subsequent NeRF-based works integrated physically-based rendering to further disentangle object and lighting properties [ 23 , 24 , 25 , 26 ] . Recent advancements include TensoIR [ 27 ] , which leverages tensor factorization to efficiently model illumination, and Neural-PBIR [ 28 ] , a hybrid pipeline that pairs neural representations with explicit classical physics-based rendering. In contrast, radiance transfer and ratio image methods [ 7 , 29 ] perform image-based relighting without explicit BRDF estimation by manipulating spherical harmonic (SH) identities. While these methods trade the ability to explicitly model complex, high-frequency material properties for computational efficiency, they provide a practical alternative to iterative optimization. TranSplat brings these theoretical foundations into the Gaussian Splatting (GS) framework. Inverse Rendering and Relighting in GS. Recent methods apply physically-based rendering to the 3D Gaussian Splatting (GS) [ 1 ] framework to estimate per-Gaussian surface properties [ 2 , 3 , 30 , 11 ] . Because single-environment captures inherently entangle lighting and materials, ReCap [ 6 ] requires cross-environment captures. Other works utilize precomputed transport: PRTGS [ 4 ] adapts Precomputed Radiance Transfer but requires heavy precomputation for dense transport

Read the complete text