Novel View Enhancement for 3D Gaussian Splatting via Stable Video Diffusion

This repository contains an implementation of 3DGS-Enhancer, a method that improves low-quality novel views rendered from 3D Gaussian Splatting (3DGS) models using Stable Video Diffusion (SVD). The approach reframes 3D-consistent image enhancement as a pose-aware, temporally consistent video generation task, enabling high-quality view synthesis even from severely degraded sparse-view reconstructions.

Figure 1: Low-quality (left) and enhanced (right) novel views.

Overview

• Enhances blurry or artifact-ridden 3DGS renderings.

• Uses camera-pose–aware conditioning to maintain 3D consistency.

• Trained with renderings generated from varying numbers of input views to improve robustness.

• Tested on both DL3DV outdoor scenes and THuman human meshes.

• Produces noticeable improvements in visual quality compared to raw 3DGS outputs.

What's Included

• Re-implementation of the SVD-guided enhancement pipeline; in folder 3DGS_Enhancer_SVD. Please refer to train.sh and srun_inference.sh for training and inference instructions.
• Code for generating low-quality conditional views; in folder 3dgs_dataset_generator.
• Auxilliary code required for rendering multi-views from meshes; in folder THuman_dataset_generator.
• calc_eval_metrics for calculating SSIM and PSNR scores on enhanced view predictions.

Miscellaneous

Downloading an example scene from DL3DV dataset using DL3DV_download.py

python DL3DV_download.py --odir DL3DV-10K --subset 2K --resolution 960P --file_type images+poses --hash e2cedefea8a0ed2d0ffbd5bdc08acbe7e1f85c96f72f7b790e9dfe1c98963047 --clean_cache