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..

Anti-Aliased 2D Gaussian Splatting

Authors: Mae Younes, Adnane Boukhayma

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate AA-2DGS on standard novel view synthesis datasets, including Mip-Ne RF 360 [2] and Blender [43], as well as the DTU [29] mesh reconstruction benchmark. Our results demonstrate that AA-2DGS consistently outperforms the original 2DGS method, particularly under challenging conditions such as varying sampling rates and mixed resolution training.
Researcher Affiliation Academia Mae Younes Adnane Boukhayma INRIA France, University of Rennes, CNRS, IRISA
Pseudocode No The paper describes methods using mathematical formulations and descriptive text, but no explicit pseudocode or algorithm block is present.
Open Source Code No Code will be available at AA-2DGS.
Open Datasets Yes We evaluate AA-2DGS on standard novel view synthesis datasets, including Mip-Ne RF 360 [2] and Blender [43], as well as the DTU [29] mesh reconstruction benchmark.
Dataset Splits Yes The Blender dataset [43] includes 8 synthetically rendered scenes... Each scene has 100 training views and 200 test views, rendered at 800 x 800 resolution.
Hardware Specification Yes We conduct all experiments on NVIDIA RTX A6000 GPUs.
Software Dependencies No The paper mentions 'custom CUDA kernels' for implementation but does not specify version numbers for any software dependencies or libraries.
Experiment Setup Yes We train our models for 30K iterations across all scenes and use the same loss function, Gaussian density control strategy, schedule, and hyperparameters. ... We follow the Mip-Splatting approach and recompute the sampling rate of each 2D Gaussian primitive every m = 100 iterations. Similarly, we choose the variance of our object-space Mip filter as 0.1, approximating a single pixel, and the variance of the flat smoothing filter as 0.2.