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..
Optimized Minimal 3D Gaussian Splatting
Authors: Joo Chan Lee, Jong Hwan Ko, Eunbyung Park
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that OMG reduces storage requirements by nearly 50% compared to the previous state-of-the-art and enables 600+ FPS rendering while maintaining high rendering quality. Our source code is available at https://maincold2.github.io/omg/. 4 Experiment 4.2 Performance evaluation 4.3 Ablation study |
| Researcher Affiliation | Academia | Joo Chan Lee Sungkyunkwan University Suwon, South Korea EMAIL Jong Hwan Ko Sungkyunkwan University Suwon, South Korea EMAIL Eunbyung Park Yonsei University Seoul, South Korea EMAIL |
| Pseudocode | No | The paper describes its methodology in prose and mathematical formulations within the main text and supplementary materials but does not include any explicit section or figure labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Our source code is available at https://maincold2.github.io/omg/. |
| Open Datasets | Yes | Following the previous works, we evaluated our approach on three real-world datasets, Mip-Ne RF 360 [2], Tanks&Temples [31], and Deep Blending [23]. |
| Dataset Splits | No | The paper refers to using 'training views' and 'training rays' for importance scoring and evaluating on specific datasets (Mip-Ne RF 360, Tanks&Temples, Deep Blending). However, it does not explicitly provide specific details such as percentages, sample counts, or methodology for splitting these datasets into training, validation, and test sets for its experiments. |
| Hardware Specification | Yes | The rendering results were obtained using an NVIDIA RTX 3090 GPU. Our rendering performance was measured using the same GPU, with the values in parentheses obtained from an NVIDIA RTX 4090 GPU. All experiments were conducted using an NVIDIA RTX 4090. |
| Software Dependencies | No | The paper mentions several tools and algorithms like 'G-PCC [51]', 'Huffman encoding [25]', 'LZMA [1]', 'K-means', and bases its implementation on 'Mini-Splatting [14]'. However, it does not provide specific version numbers for key software dependencies or programming environments (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | All experiments were conducted using an NVIDIA RTX 4090. Our method was implemented within the Mini-Splatting [14] framework and trained for 30K iterations. At the 20K iteration simplification process, local distinctiveness scoring was incorporated where the factor λ was set to 2. The dimension of appearance features D was set to 3. Scale and rotation were trained from the initial training, while appearance features were introduced at 15K iterations. From 29K iterations (last 1K iterations), SVQ (Sub-Vector Quantization) was applied to per-Gaussian features. For SVQ, different bit allocations were assigned. Scale: length 1, 26 codes for each sub-vector Rotation: length 2, 29 codes for each sub-vector Appearance features: length 2, 210 codes for each sub-vector All codes in the codebook are stored with 16-FP precision. The model storage for each variant was determined only by the importance score threshold τ, which is used for simplification at the 20K iteration, set to 0.96, 0.98, 0.99, 0.999, and 0.999, respectively. |