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..
Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling
Authors: Zhihao Li, Yufei Wang, Heliang Zheng, Yihao Luo, Bihan Wen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The paper includes a dedicated "4 Experiments" section, where it details implementation, datasets, comparison methods, and presents quantitative and qualitative results using tables (Table 1, Table 2) and figures (Figure 4, Figure 5, Figure 6). It also includes "4.3 Ablation Studies" to validate components of the proposed method. For example: "We implement all Sparcubes as custom CUDA kernels. Following TRELLIS [32], we train both the Sparconv-VAE and its latent flow model on 500 K high-quality assets from Objaverse [8] and Objaverse-XL [7]. The VAE runs on 32 A100 GPUs (batch size 32) with Adam W (initial LR 1 10 4) for two days." and "Table 1: Quantitative comparison of watertight remeshing across the ABO [6], Objaverse [8], and In-the-Wild datasets." |
| Researcher Affiliation | Collaboration | The authors list affiliations from "Nanyang Technological University, Singapore" (academic), "Math Magic" (private company/lab, indicated by generic email and lack of city/country), and "Imperial College London, UK" (academic). The presence of both academic and non-academic institutions indicates a collaboration. For example: "Zhihao Li1,2 , Yufei Wang1, Heliang Zheng2, , Yihao Luo2,3, , Bihan Wen1, 1Department of EEE, Nanyang Technological University, Singapore 2Math Magic 3Imperial-X, Imperial College London, UK EMAIL, EMAIL, EMAIL EMAIL, EMAIL" |
| Pseudocode | No | The paper describes the Sparcubes reconstruction pipeline in Section 3.2 using numbered steps (Step 1, Step 2, Step 3, Step 4) and mathematical equations. However, these steps are presented as descriptive paragraphs and equations, not in a formalized pseudocode block or algorithm environment. |
| Open Source Code | No | The paper states in the NeurIPS Paper Checklist under "5. Open access to data and code": "Justification: We will release the code once the paper has been accepted and has successfully passed internal review." This indicates a plan for future release, not current concrete access to the code. |
| Open Datasets | Yes | The paper explicitly mentions and cites several well-known public datasets used for training and evaluation: "We train both the Sparconv-VAE and its latent flow model on 500 K high-quality assets from Objaverse [8] and Objaverse-XL [7]." and "Following Dora [2], we curated a VAE test set by selecting the most challenging examples from the ABO [6] and Objaverse [8] datasets" |
| Dataset Splits | No | The paper mentions curating a "VAE test set" from ABO and Objaverse datasets and assembling a "Wild dataset" that is "disjoint from both ABO and Objaverse" for generation benchmarks. However, it does not provide specific details on the dataset splits such as exact percentages, sample counts for training/validation/test sets, or a detailed methodology for generating these splits beyond selecting challenging examples or ensuring disjointness. |
| Hardware Specification | Yes | The paper explicitly states the types and quantities of GPUs used for training: "The VAE runs on 32 A100 GPUs (batch size 32)... We then fine-tune the TRELLIS latent flow model on our VAE latents using 64 A100 GPUs (batch size 64)" |
| Software Dependencies | No | The paper mentions using "Adam W" as an optimizer, but it does not specify version numbers for any other software dependencies such as programming languages, deep learning frameworks (e.g., PyTorch, TensorFlow), or CUDA versions. |
| Experiment Setup | Yes | The paper provides concrete experimental setup details including batch sizes, optimizer type, initial learning rate, and inference parameters: "The VAE runs on 32 A100 GPUs (batch size 32) with Adam W (initial LR 1 10 4) for two days. We then fine-tune the TRELLIS latent flow model on our VAE latents using 64 A100 GPUs (batch size 64) for ten days. At inference, we sample with a classifier-free guidance scale of 3.5 over 25 steps, matching TRELLIS settings." |