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..

LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering

Authors: Jonas Kulhanek, Marie-Julie Rakotosaona, Fabian Manhardt, Christina Tsalicoglou, Michael Niemeyer, Torsten Sattler, Songyou Peng, Federico Tombari

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To validate our method, we conduct experiments on large-scale indoor and outdoor datasets, comparing our method to other SOTA approaches. We further analyze individual components of our method and evaluate rendering speed on various mobile devices. We report standard PSNR, SSIM, and LPIPS (VGG) metrics, but also FPS and the number of Gaussians loaded in GPU memory (#G).
Researcher Affiliation Collaboration 1 Google, 2 Google Deep Mind, 3 Technical University of Munich, 4 Czech Technical University in Prague, Faculty of Electrical Engineering, 5 Czech Technical University in Prague, Czech Institute of Informatics, Robotics and Cybernetics
Pseudocode No The paper describes its methodology through detailed text, equations (e.g., Eq. (1), (3), (4), (6)), and figures, but does not include a distinct block explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code No Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: Paper code cannot be released at the time of submission due to institutional policies.
Open Datasets Yes We use two larger-scale datasets to validate our approach, ie., two outdoor scenes from Hierarchical 3DGS dataset [11] and three indoor scenes from Zip-Ne RF dataset [2].
Dataset Splits Yes In our experiments, we use the official train/test split provided by the authors. We also use the official segmentation masks provided with the dataset to remove license plates, pedestrians, and moving cars. Same as Zip-Ne RF, we take each 8th image as testing image (when sorted alphabetically).
Hardware Specification Yes All baselines and our method use a single NVIDIA A100 SXM4 40GB GPU for training and evaluation. For mobile experiments, we used two i Phones (13 Mini and 15 Pro), as well as two lower-end laptops without a powerful GPU (Mac Book Air M3 and HP Chromebook).
Software Dependencies No The paper mentions using a "web-based 3DGS renderer by Mark Kellogg [9]" but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes All models were trained for for 30 000 iterations. We reset opacity every 3 000 steps until iteration 15 000 and apply densification every 300 steps from step 600 to 15 000. Importance score pruning is performed at steps 8 000, 16 000, and 24 000 with threshold of 0.02 for all scenes, except Campus, where we use lower threshold of 0.01. We use the same loss function (DSSIM+L1) as during the standard optimization.