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..
Fix False Transparency by Noise Guided Splatting
Authors: Aly El Hakie, Yiren Lu, Yu Yin, Michael Jenkins, Yehe Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments across multiple datasets show that NGS substantially reduces false transparency while maintaining competitive performance on standard rendering metrics (e.g., PSNR), demonstrating its overall effectiveness. |
| Researcher Affiliation | Collaboration | Aly El Hakie1 , Yiren Lu2 , Yu Yin2, Michael Jenkins1,2, Yehe Liu1,2 1Opsi Clear LLC 2Case Western Reserve University |
| Pseudocode | Yes | Figure 3: 3DGS training schedule with NGS. NGS is an add-on to the standard 3DGS pipeline. Initialization SFM + initial training Growth (~15k iters) Densify, prune & reset opacity Noise Initiation Freeze surface points Add multi-scale, subsurface noise Fine-tuning Stop density control & opacity reset Noise Finalization Freeze noise Continue regular training 3DGS (generic workflow) NGS (as an add on) Noise optimization (@ ~6k iter) Randomize color every iteration Train noise opacity and prune transparent noise |
| Open Source Code | No | All data created for this study will be made available to the community. This release includes the Stone Dataset, foreground segmentation masks, the generated noise infills, and a supplementary dataset featuring a mixture of everyday objects (Fig. S4). Justification: We will provide open access to the data and code. |
| Open Datasets | Yes | We used public available object-centric datasets, DTU [37] and Omni Object3D [38], to evaluate NGS. ... All data created for this study will be made available to the community. This release includes the Stone Dataset, foreground segmentation masks, the generated noise infills, and a supplementary dataset featuring a mixture of everyday objects (Fig. S4). |
| Dataset Splits | Yes | A standard 7:1 train-test split was used for all datasets, with the test set forming the basis for all quantitative metrics. |
| Hardware Specification | Yes | We conducted all experiments on NVIDIA L40S GPUs using the GSplat framework [41]. |
| Software Dependencies | No | We conducted all experiments on NVIDIA L40S GPUs using the GSplat framework [41]. Unless explicitly stated, our base implementation of NGS used the default variant of Gsplat. |
| Experiment Setup | Yes | Noise Gaussians were introduced at iteration 6,000 during adaptive density control, allowing sufficient time for the initial surface reconstruction to establish before applying our transparency guidance strategy (Fig. 3). We refined the noise Gaussians for 1,000 iterations. The surface Gaussians were frozen during noise refinement. After the noise refinement, the noise Gaussians means, opacity, scale and rotations were frozen. Until the end of training, each noise Gaussian s color was randomized from RGBCMY at each iteration, preventing the surface Gaussians from fitting a fixed noise pattern. We reset the learning rate of Gaussian means to compensate for the sudden change to the blending. The rest of training follows the default GSplat parameters. |