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..
Bézier Splatting for Fast and Differentiable Vector Graphics Rendering
Authors: Xi Liu, Chaoyi Zhou, Nanxuan Zhao, Siyu Huang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that Bézier Splatting significantly outperforms existing methods with better visual fidelity and significant optimization speedup. The project page is xiliu8006.github.io/Bezier_splatting_project. |
| Researcher Affiliation | Collaboration | Xi Liu1 , Chaoyi Zhou1 , Nanxuan Zhao2, Siyu Huang1 1Clemson University, 2Adobe Research |
| Pseudocode | Yes | Algorithm 1 Adaptive Pruning and Densification for Closed Bézier Curves ... Algorithm 2 Convert Bézier Splatting to Standard SVG |
| Open Source Code | No | The project page is xiliu8006.github.io/Bezier_splatting_project. ... The code will be publicly available. |
| Open Datasets | Yes | We use the publicly available DIV2K [24] dataset for evaluating natural images. ... Additionally, we test our method on non-photorealistic images, including the artwork images from Clipart1K dataset [10] and cartoon images from Danbooregions dataset [32] ... We further evaluate our method on another natural image dataset, Kodak [14]. |
| Dataset Splits | Yes | Due to the high computational cost of baseline method LIVE [18], we uniformly subsample the DIV2K dataset by selecting one out of every four images, resulting in a final evaluation set of 200 images from the original 800-image dataset. |
| Hardware Specification | Yes | This work proposes Bézier Splatting, a new differentiable vector graphics (VGs) renderer that achieves an order-of-magnitude computational speedup in comparison with the state-of-the-art method Diff VG [15] (tested on a NVIDIA RTX 4090 GPU). |
| Software Dependencies | No | We implement Bézier Splatting in Py Torch [20] and optimize it by using the Adam optimizer [13] with a Step LR learning rate scheduler. |
| Experiment Setup | Yes | We implement Bézier Splatting in Py Torch [20] and optimize it by using the Adam optimizer [13] with a Step LR learning rate scheduler. The learning rate is initialized at 0.01 for color, 2e-4 for Bézier curve control points, and 0.1 for opacity. For the pruning and densification strategy, the opacity threshold is set to 0.02, and the overlap threshold based on Axis Aligned Bounding Boxes (AABB) is set to 0.9. Following the approach in LIVE [18], new curves are initialized in a circular pattern, and the number of added curves matches the number of removed ones to maintain a constant total curve count. Open curves are optimized for 15,000 iterations, and closed curves are optimized for 10,000 iterations. Pruning and densification are applied every 400 steps until the last but 1,000 steps, after which they are halted to stabilize the representation. |