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..
VisDiff: SDF-Guided Polygon Generation for Visibility Reconstruction, Characterization and Recognition
Authors: Rahul Moorthy Mahesh, Jun-Jee Chao, Volkan Isler
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We compare Vis Diff with existing methods on the Visibility Reconstruction problem and further demonstrate its effectiveness on the Visibility Characterization problem. We also present preliminary results for the Visibility Recognition problem. In addition, we evaluate the generalization capability of Vis Diff to other combinatorial graph structures, such as triangulation graphs. Finally, we demonstrate the ability of Vis Diff to perform high-diversity data sampling over the space of all polygons. |
| Researcher Affiliation | Academia | Rahul Moorthy University of Minnesota EMAIL Jun-Jee Chao University of Minnesota EMAIL Volkan Isler The University of Texas at Austin EMAIL |
| Pseudocode | No | Detailed architecture specifications are provided in Appendix J. |
| Open Source Code | Yes | We provide open access to data and code which has been added as link in the paper. |
| Open Datasets | Yes | We design a carefully curated dataset that captures a wide range of combinatorial properties of polygons and make it publicly available for further research. |
| Dataset Splits | Yes | The test dataset for validating our approach is generated in two splits: in-distribution and out-of-distribution. In-distribution samples are obtained by setting aside 100 unique polygons per link diameter from the larger dataset, ensuring they are not included in the training set. ... The final training dataset consists of 370,000 polygons along with their corresponding visibility and triangulation graphs. |
| Hardware Specification | Yes | All models were trained on a single NVIDIA A100 GPU using 10 workers, with a total training time of approximately 16 hours. |
| Software Dependencies | No | We train the SDF diffusion model for 60 epochs using the Adam optimizer with a learning rate of 10-4 and a batch size of 128. |
| Experiment Setup | Yes | We train the SDF diffusion model for 60 epochs using the Adam optimizer with a learning rate of 10-4 and a batch size of 128. We employ a log-linear noise scheduler with ฯmin = 0.005 and ฯmax = 10. |