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..
Efficient Part-level 3D Object Generation via Dual Volume Packing
Authors: Jiaxiang Tang, Ruijie Lu, Max Li, Zekun Hao, Xuan Li, Fangyin Wei, Shuran Song, Gang Zeng, Ming-Yu Liu, Tsung-Yi Lin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Section 4 is titled "Experiments". It includes subsections like "Qualitative Comparisons" and "Quantitative Comparisons", discussing evaluation metrics (ULIP, Uni3D), visual comparisons (Figure 4, Figure 5), and an "Ablation Study" (Section 4.4, Figure 7). For example: "Experiments show that our model achieves better quality, diversity, and generalization than previous image-based part-level generation methods." (Abstract) and "We also evaluate the generation quality of image-to-3D methods from single-view images. Following Hunyuan3D-2.0 [59], we adopt ULIP [48, 49] and Uni3D [61] as evaluation metrics." (Section 4.3). |
| Researcher Affiliation | Collaboration | The authors are affiliated with: "1State Key Laboratory of General AI, Peking University" (academic), "2NVIDIA Research" (industry), and "3Stanford University" (academic). The presence of both university (Peking University, Stanford University) and industry (NVIDIA Research) affiliations indicates a collaborative effort. |
| Pseudocode | Yes | The supplementary materials (Appendix A.1) explicitly include "Algorithm 1: Greedy Odd Cycle Contraction", which provides a structured algorithmic description. |
| Open Source Code | No | The NeurIPS Paper Checklist for "Open access to data and code" states: "Answer: [No] Justification: The dataset is publicly available. The code will be released later." |
| Open Datasets | Yes | The paper explicitly states: "We use the Trellis500k subset [46] of the Objaverse-XL dataset [12, 11] (ODC-BY v1.0 license)." References [11] and [12] provide detailed information for the Objaverse-XL dataset. |
| Dataset Splits | No | The paper mentions: "After applying our part extraction process (Section 3.2), approximately 386K meshes with more than one part remain. Further filtering results in around 254K meshes with well-balanced SDF grids, which are used for training our model." (Section 4.1) and "To perform the evaluation, we curate a test set of 40 images sourced from diverse domains..." (Section 4.3). While it states the number of meshes for training and the size of the test set, it does not specify explicit percentages or a methodology for splitting the main dataset into training, validation, and test sets beyond the filtering process. |
| Hardware Specification | Yes | The paper specifies the hardware used: "The VAE is trained at multiple latent sizes on 64 A100 GPUs over the course of approximately one week." and "This progressive training phase spans about two weeks using at most 256 A100 GPUs." (Section 4.1). |
| Software Dependencies | No | The paper mentions general software concepts like VAEs, rectified flow models, and specific techniques like Marching Cubes, but it does not provide specific version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, CUDA versions) used in its implementation. |
| Experiment Setup | Yes | The paper provides several specific details about the experimental setup. For instance, in Section 4.1, it states: "All meshes are normalized to the [ 0.95, 0.95]3 cube, and watertight conversion is performed at a resolution of 5123, with the dilation threshold set to the voxel size." In Appendix A.2, it further details: "we sample 32768 uniformly distributed surface points and 16384 salient edge points as input to the VAE. The dihedral angle threshold for salient edge detection is set to less than 165." and "During VAE training, we randomly select 16384 uniform samples, 8192 near-surface samples, and 8192 near-salient-edge samples in each iteration." Also, in Section 4.3, it mentions: "we fix the number of denoising steps to 50, extract meshes at a grid resolution of 5123, and simplify them to 50000 faces via decimation." |