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..
LinPrim: Linear Primitives for Differentiable Volumetric Rendering
Authors: Nicolas von Lützow, Matthias Niessner
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through experiments on real-world datasets, we demonstrate comparable performance to state-of-the-art volumetric methods while requiring fewer primitives to achieve similar reconstruction fidelity. |
| Researcher Affiliation | Academia | Nicolas von Lützow Technical University of Munich EMAIL Matthias Nießner Technical University of Munich EMAIL |
| Pseudocode | No | The paper describes algorithms and gradient computations but does not include a figure, block, or section explicitly labeled 'Pseudocode' or 'Algorithm', nor are there structured steps formatted like code blocks in the main text or appendices. |
| Open Source Code | No | We will open-source our training, evaluation, and rendering code (and include full instructions) upon acceptance. An anonymized version of the code can be provided upon request. |
| Open Datasets | Yes | We evaluate our results on Mip-Ne RF 360 [2] and Scan Net++ v2 [29]. (In Appendix I: Mip-Ne RF 360 Data [1]: Available from https://jonbarron.info/mipnerf360/ and does not provide license terms. Scan Net++ v2 Data [29]: Available from https://kaldir.vc.in.tum.de/scannetpp/ and uses a custom license also found on the website.) |
| Dataset Splits | Yes | We train and test on the official splits provided by each dataset and maintain consistent parameters across all evaluated scenes. |
| Hardware Specification | Yes | Experiments were performed using either a single RTX A6000 or RTX 3090, rendering speed was evaluated using the same RTX 3090 setup for all approaches. |
| Software Dependencies | No | The paper mentions using code from 3DGS [13], Mip-Splatting [30], and GS-MCMC [14], specifying commit IDs in Appendix I. However, it does not explicitly list specific version numbers for general software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | In this section, we give specific values for the hyperparameters used and the experimental setting to ensure the reproducibility of our results. Experiments were performed using either a single RTX A6000 or RTX 3090, rendering speed was evaluated using the same RTX 3090 setup for all approaches. We use the following learning rates for both our Octahedron and Tetrahedron approaches, the position learning rates and the corresponding schedule are consistent with 3DGS [13]. Specific values were chosen empirically. Table 16: The learning rates used for our Octahedron and Tetrahedron approaches. The distance learning rate is scaled by the maximum distance between a known camera pose and the mean camera pose to ensure consistent behavior even in scale-ambiguous settings. Other notable paradigms and parameters concerning the initialization and population of primitives are as follows: Primitives are initialized with distances equal to the nearest Sf M point, clamped between 10 5 and 0.5. Opacities are set to 0.1 and rotations are sampled uniformly at random. The population is adjusted every 250 iterations. Primitives with position gradients larger than 1.5 10 4 are considered for densification, and ones with opacity smaller than 0.025 are removed. Similarly, primitives with a size larger than 40% of the scene or 20 pixels are removed. During densification, primitives smaller than 1% of the scene are duplicated. Larger ones are replaced with two newer ones with a relative size of 1.2 1. For octahedra, the position is chosen normally at random with a standard deviation aligned with, and equal to, the distances in each dimension. For tetrahedra, we sample normally with standard deviations equal to half of the largest distance. We do not propagate gradients through the density normalization factor (see Equation 1). For our anti-aliasing efforts, we set the kernel size of the 2D filter to 0.1. For octahedra, the ray space vector is moved by half that amount, as the adjustment is mirrored on both sides. The 3D filter remains consistent with Mip-Splatting. |