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..
MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference
Authors: Wenyuan Zhang, Jimin Tang, Weiqi Zhang, Yi Fang, Yu-Shen Liu, Zhizhong Han
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on widely used benchmarks show that our method faithfully recovers both illumination and geometry, achieving state-of-the-art rendering quality in novel views synthesis. Project Page: https://wen-yuan-zhang.github.io/Material Ref GS. 4 Experiments 4.1 Experiment Settings Datasets & Metrics. We evaluate the performance of our method on widely used benchmarks, including two synthetic datasets, Shiny Blender [47] and Glossy Synthetic [33], as well as two real-world datasets, Ref-Real [47] and Mip-Ne RF 360 [1]. All of these datasets contain challenging scenes with prominent reflective surfaces. To evaluate the quality of novel view synthesis, we report PSNR, SSIM [49] and LPIPS [59]. We also evaluate the accuracy of the predicted normals using Mean Angular Error (MAE). 4.3 Ablation Study Effectiveness of Each Module. We conduct ablation studies to evaluate the effectiveness of each module in our framework on both synthetic and real-world datasets. |
| Researcher Affiliation | Academia | School of Software, Tsinghua University, Beijing, China1 Center for AI and Robotics (CAIR), NYU Abu Dhabi, UAE2 Department of Computer Science, Wayne State University, Detroit, USA3 |
| Pseudocode | No | The paper describes methods using equations and textual explanations, but no distinct pseudocode or algorithm blocks are provided. |
| Open Source Code | No | We do not include our code or data at this stage, but we will definitely release our code and data in the future. |
| Open Datasets | Yes | We evaluate the performance of our method on widely used benchmarks, including two synthetic datasets, Shiny Blender [47] and Glossy Synthetic [33], as well as two real-world datasets, Ref-Real [47] and Mip-Ne RF 360 [1]. |
| Dataset Splits | No | The paper mentions using widely used benchmarks but does not specify the training, validation, or test splits used for these datasets. It refers to standard benchmarks without detailing their specific splits. |
| Hardware Specification | No | The paper states, "We report our training time with baseline methods in the supplementary materials." However, the main paper does not provide specific hardware details (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with their version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x) that would be needed to replicate the experiments. |
| Experiment Setup | Yes | We train our method for a total of 30k iterations. We begin by training a 2DGS [18] with normal priors during the first 3k iterations to ensure geometric stability. After that, we incorporate PBR and our environment illumination modeling into the training process. At 10k iteration, we remove the normal prior to avoid potential bias from inaccurate predictions, and introduce our multi-view regularization terms. We also adopt normal propagation [24, 56] to propagate reliable normals to neighboring Gaussians for consistency and stability. The loss function can be written as L = Lc + λn d Ln d + λn Ln + λmv Lmv + λref Lref, (10) where Lc = 0.8 Lrgb + 0.2 LD SSIM is the photometric loss commonly used in Gaussian-based methods [22, 56], Ln d denotes the depth-normal consistency loss used in 2DGS [18], and Ln = |1 N T N| is the normal prior loss. Lmv, Lref correspond to our multi-view consistency loss (Eq. 5) and reflection strength loss (Eq. 7), respectively. |