Optimization-Free Style Transfer for 3D Gaussian Splats

TL;DR

This work tackles the challenge of stylizing 3D Gaussian splats without reconstructing or retraining the scene. It introduces a surface-based graph pipeline that converts a 3DGS into an oriented graph, applies a pretrained image-style transfer network via the Interpolated SelectionConv framework, and then interpolates the stylized colors back to each splat, enabling direct stylization on .splat/.ply files. The method runs primarily on CPU, achieving under ~2 minutes for large scenes and does not require original camera views, COLMAP outputs, or re-rendered views. Qualitative results show competitive visual fidelity, especially for single-object splats, with ablations confirming the importance of accurate normals, and the approach offers favorable speed compared to reconstruction-based stylization methods. The work includes an ablation on normals and discusses limitations such as lack of geometric editing and reliance on a pseudo implicit surface, outlining directions toward transformer- or diffusion-based stylization in the future.

Abstract

The task of style transfer for 3D Gaussian splats has been explored in many previous works, but these require reconstructing or fine-tuning the splat while incorporating style information or optimizing a feature extraction network on the splat representation. We propose a reconstruction- and optimization-free approach to stylizing 3D Gaussian splats, allowing for direct stylization on a .ply or .splat file without requiring the original camera views. This is done by generating a graph structure across the implicit surface of the splat representation. A feed-forward, surface-based stylization method is then used and interpolated back to the individual splats in the scene. This also allows for fast stylization of splats with no additional training, achieving speeds under 2 minutes even on CPU-based consumer hardware. We demonstrate the quality results this approach achieves and compare to other 3D Gaussian splat style transfer methods. Code is publicly available at https://github.com/davidmhart/FastSplatStyler.

Figures (8)

• Figure 1: A 3D Gaussian splat with geometrically complex content and an image containing a distinct artistic style (Left) and a 3D Gaussian splat of the same content after style transfer using our approach (Right).
• Figure 2: Overview of our approach. A graph construction pipeline takes an existing Gaussian Splat scence and samples it as a point cloud. Normal vectors are approximated and points in the point cloud are connected using a K-Nearest-Neighbors approach. A local planar approximation of the graph at each point in the graph allows for convolution on the splat using 2D image CNN weights. A style transfer network is then converted into the graph space using SelectionConv hart2023interpolated. The resulting graph is interpolated back to existing splat center to modify the color values, generating the final stylized result. The pipeline is shown in both a technical version (Top) and full splat visualization (Bottom).
• Figure 3: A 3D Gaussian splat under two different style transfers using our approach.
• Figure 4: A toy example 3D Gaussian splat scene (Left) and high resolution scene of toys (Right). The originally rendered scene (Top) is spatially misrepresented when selecting only the centers of the provided splats (Middle). A higher density point cloud can be generated by extensive sampling within each Gaussian (Bottom). The graph created from this new point cloud provides higher quality results when fed into the style transfer network.
• Figure 5: A high quality 3D Gaussian splat (Left) after style transfer using a point cloud created from selecting the centers of individual Gaussians (Middle) and using a point cloud with Gaussian Sampling to reduce information loss (Right) demonstrating the effectiveness of applying Gaussian Sampling. Lettering detail shown in bottom row.