TranSplat: Lighting-Consistent Cross-Scene Object Transfer with 3D Gaussian Splatting

Boyang Yu, Yanlin Jin, Ashok Veeraraghavan, Akshat Dave, Guha BalakrishnanarXiv 2025 · DOI · arXiv

Abstract

TranSplat introduces a novel framework for the near-instant, realistic relighting of 3D objects when transferred between disparate 3D Gaussian Splatting (3DGS) scenes. By leveraging a mathematical radiance transfer identity within the Spherical Harmonics (SH) domain, the method bypasses the computationally expensive and often ill-posed process of explicit material decomposition. TranSplat achieves high-fidelity relighting in under 10 seconds—orders of magnitude faster than existing optimization-based or diffusion-based inverse rendering techniques.

Introduction

3D Gaussian Splatting (3DGS) has emerged as a transformative technique for real-time neural rendering, offering high-quality reconstructions of complex scenes. However, editing these scenes—specifically transferring a digital asset from a "Source" environment to a "Target" environment—remains a significant challenge. A primary hurdle is ensuring lighting consistency; an object captured in a sunny outdoor setting will look jarringly out of place if moved into a dimly lit interior without its appearance being adjusted to match the new environment.

Overview of TranSplat's cross-scene object transfer and relighting pipeline.

Traditionally, this problem is addressed via inverse rendering, which attempts to disentangle an object's geometry, material properties (BRDFs), and the surrounding illumination. This process is notoriously difficult and slow, often requiring hours of optimization. TranSplat addresses this by providing an analytical solution that enables "instant" relighting without explicit material estimation.

Methodology

The core of TranSplat is a radiance transfer identity based on Spherical Harmonics (SH). The pipeline consists of four key components:

1. Gaussian Surfels

Standard 3DGS uses ellipsoidal Gaussians, which lack well-defined surface normals. TranSplat instead represents objects using "surfels" (flat Gaussians). This geometric refinement is crucial because accurate surface normals are required to calculate how light interacts with the object's surface.

2. Automated Lighting Estimation

To relight an object, the system must understand the lighting of the target scene. Rather than relying on external HDR environment maps, TranSplat renders a "cube map" directly from the 3DGS target scene at the specific location where the object will be placed. This captures spatially varying lighting, including local shadows and reflections, which global maps often miss.

3. Spherical Harmonic Transfer (The "Ratio" Shortcut)

The most significant insight of the paper is the use of a theoretical identity for surfaces with radially symmetric BRDFs (such as matte or moderately shiny materials). The authors demonstrate that relighting can be approximated by multiplying the object's original SH appearance coefficients by the ratio of the target lighting to the source lighting. This "ratio" acts as a universal transfer function, allowing for relighting without knowing the specific material properties of the object.

4. Shading and Shadow Baking

To enhance realism, TranSplat computes per-Gaussian shading maps to handle self-occlusion. Additionally, a lightweight "shadow baking" step projects shadows from the transferred object onto the target scene's geometry, based on the dominant light directions identified in the environment.

Key Results

TranSplat was evaluated against state-of-the-art (SOTA) methods using both synthetic benchmarks (e.g., TensoIR-Synthetic, Toys4K) and real-world captures (e.g., MipNeRF360).

Efficiency and Speed

The most striking result is the computational efficiency. While optimization-based methods like GS-IR can take nearly 50 minutes and diffusion-based models like Neural Gaffer can take up to 16 hours, TranSplat completes the entire relighting process in approximately 6 seconds. This near-instantaneous performance makes it suitable for interactive editing workflows.

Side-by-side comparison of object transfer without relighting versus TranSplat's relighting.

Accuracy and Fidelity

Quantitatively, TranSplat outperformed baselines in synthetic benchmarks, achieving PSNR (Peak Signal-to-Noise Ratio) values exceeding 30dB, compared to the low 20s for many competing methods. Qualitatively, the system accurately captures color shifts—such as an object taking on a reddish hue under warm lighting—and generates plausible dynamic shadows that respond to changes in the environment.

Limitations and Future Work

Despite its speed and accuracy, TranSplat has limitations. The reliance on radially symmetric BRDFs means the method struggles with highly specular, mirror-like surfaces or complex inter-reflections. Furthermore, the quality of the output is tied to the accuracy of the initial 3DGS reconstruction; noisy geometry leads to blotchy shading. Future research will focus on improving normal estimation and extending the framework to support more complex, non-Lambertian material types.

Conclusion

TranSplat represents a significant step forward in 3D scene editing. By replacing slow, iterative optimization with a direct analytical calculation based on Spherical Harmonics, it enables realistic, lighting-consistent object transfer in seconds. This efficiency opens new possibilities for real-time digital asset management and virtual scene composition.

Summary updated

BibTeX
@misc{yu2025transplat,
  author = {Boyang Yu and Yanlin Jin and Ashok Veeraraghavan and Akshat Dave and Guha Balakrishnan},
  title = {{TranSplat: Lighting-Consistent Cross-Scene Object Transfer with 3D Gaussian Splatting}},
  year = {2025},
  eprint = {2503.22676},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.22676},
}
Figure