Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

EGGS: Exchangeable 2D/3D Gaussian Splatting for Geometry-Appearance Balanced Novel View Synthesis

Authors: Yancheng Zhang, Guangyu Sun, Chen Chen

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that EGGS outperforms existing methods in rendering quality, geometric accuracy, and efficiency, providing a practical solution for high-quality NVS.
Researcher Affiliation Academia Yancheng Zhang , Guangyu Sun , and Chen Chen Institute of Artificial Intelligence, University of Central Florida EMAIL
Pseudocode Yes Algorithm 1: Frequency-Decoupled Optimization
Open Source Code Yes Code and demo available at https://github.com/Fobow/EGGS.
Open Datasets Yes We evaluate EGGS on several widely used benchmarks. For appearance evaluation, we use Mip-Ne RF360 [25], LLFF [41], Tanks&Temples [42], and DTU [43].
Dataset Splits Yes We follow the standard train/test splits used in prior work [25], with additional dataset statistics provided in Table 6.
Hardware Specification Yes All experiments are conducted on an A5000 GPU.
Software Dependencies No We implement our hybrid rasterizer based on the CUDA rasterization code of 3DGS [11]. We used the Haar filter for the DWT [37, 48]. The paper mentions software components like CUDA and Haar filter, but does not provide specific version numbers for key software dependencies like Python, PyTorch, or CUDA toolkit versions.
Experiment Setup Yes Our training setup closely follows 3DGS [11]. All methods, including EGGS and baselines, are trained for 30K iterations. Learning rates for Gaussian parameters follow the default configurations from 3DGS and 2DGS. We adopt the densification strategy from 3DGS, which refines Gaussian distributions by pruning or duplicating them in underor over-reconstructed regions. Densification begins at iteration 500, ends at iteration 15K, and is performed every 100 iterations. ... During type switching, we set the effective rank threshold θe to 2.05. For scale modulation, ... we use θz = 1.05 and temperature T = 0.001 in Eq. 7. Additionally, we incorporate frequency-based supervision... we set the weight for the frequency components as λlow = 0.2 and λlow = 0.4.