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..
ViewCraft3D: High-fidelity and View-Consistent 3D Vector Graphics Synthesis
Authors: Chuang Wang, Haitao Zhou, Ling Luo, Qian Yu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive experiments demonstrate that VC3D outperforms previous methods in both qualitative and quantitative evaluations, while significantly reducing computational overhead. The resulting 3D sketches maintain view consistency and effectively capture the essential characteristics of the original objects. |
| Researcher Affiliation | Academia | 1Beihang University 2Ningbo University |
| Pseudocode | Yes | Algorithm 1 shows the details of Salient Point Cloud Extraction (in main paper Sec. 3.2.1). |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: The code and data need to be organized and then open-sourced so that others can use or follow them. |
| Open Datasets | No | We collected 40 images from prior works and online sources as inputs. |
| Dataset Splits | No | We collected 40 images from prior works and online sources as inputs. All generated 3D vector graphics are rendered into 12 views using identical camera parameters, upon which the metrics are computed. ... We selected a subset of 20 images from the inputs in Section 4.2.1, where all examples were optimized with SDS loss. ... For the user study, each questionnaire contained 20 cases randomly sampled from these 40. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | Our VC3D framework is implemented in Py Torch. |
| Experiment Setup | Yes | Our VC3D framework is implemented in Py Torch. For primary structure fitting stage, we set the distance threshold dthresh = 0.05 and angle threshold θthresh = 50 . Each Bézier curve is defined by 4 control points, with s = 64 sample points per curve for optimization. For detail refinement stage, we employ Tripo SG [15] as our pretrained image-to-3D model, with SDS loss weight set to 2 10 4. We use the SGD optimizer [29] with a learning rate of 5 10 3. ... Our full method takes about 30 minutes to generate a vector graphic, with 100 optimization steps for Stage I and 200 steps for Stage II. ... By default, N is set to 128, but it can be adjusted based on the geometric complexity of missing areas. The intricate-to-approximate regions are determined by the mesh vertices and the point cloud Ps. Specifically, for a point p, if its distance to all points in the point cloud Ps exceeds a threshold (set to 0.03 in our experiments), it is considered to belong to the intricate-to-approximate regions. |