Zihan Wang Golisano College of Computing and Information Sciences, Rochester Institute of Technology

Abstract: A real-time 3D style transfer pipeline based on 3D Gaussian Splatting (3DGS). This project addresses the "flickering" and "visual noise" artifacts common in existing methods (like StyleSplat) by introducing a hybrid loss function that combines Perceptual Loss (LPIPS) for structural coherence and Reprojection Consistency Loss for temporal stability.

1. Smooth Geometry (Hotdog Scene) Our method eliminates high-frequency noise and creates coherent brushstrokes.

|

2. Complex Geometry (Chair Scene) Our method maintains style consistency even on thin structures and complex occlusions.

| |

We achieve a ~4% reduction in temporal warping error compared to the baseline, significantly reducing the "flickering" effect in videos.

We achieve an average LPIPS score of 0.72, indicating robust perceptual similarity to the target style across diverse geometries.

• Hybrid Loss Function: * $L_{total} = \lambda_{nnfm} L_{nnfm} + \lambda_{lpips} L_{lpips} + \lambda_{reproj} L_{reproj}$
• Perceptual Loss: Uses VGG-16 features to enforce global structure (removing splotchy artifacts).
• Reprojection Consistency: Uses depth-based warping to enforce temporal stability (removing flickering).
• Robust Segmentation: Integration with DEVA for precise temporal object masking.
• Colab Optimization: Custom patches for NumPy 1.x/2.x compatibility and CUDA compilation.

This project relies on custom CUDA kernels. Please follow the installation steps strictly, especially regarding package versions.

Prerequisites:

• Linux (Ubuntu 20.04 recommended or Google Colab)
• NVIDIA GPU (Tesla T4, A100, etc.)
• CUDA Toolkit 11.3
• Python 3.10

git clone [https://github.com/YOUR_USERNAME/StyleSplat-Consistency.git](https://github.com/YOUR_USERNAME/StyleSplat-Consistency.git)
cd StyleSplat-Consistency
# Important: Clone submodules recursively
git submodule update --init --recursive