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 [1].
GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation
Authors: Dingdong Yang, Yizhi Wang, Konrad Schindler, Ali Mahdavi Amiri, Hao Zhang
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4 EXPERIMENTS AND RESULTS. We test the reconstruction and generation ability of our methods on the whole Shape Net (Chang et al., 2015) dataset. We follow the dataset split of previous works (Zhang et al., 2022; 2023; Yariv et al., 2023) and train conditional generation on the whole Shape Net classes. 4.3 RECONSTRUCTION RESULTS. 4.4 GENERATION RESULTS. 4.5 ABLATION STUDIES. |
| Researcher Affiliation | Academia | Dingdong Yang1 Yizhi Wang1 Konrad Schindler2 Ali Mahdavi-Amiri1 Hao Zhang1 1Simon Fraser University. 2ETH Zurich. |
| Pseudocode | Yes | Algorithm 1 Rescaling the Grid with Histogram. Algorithm 2 Flip Signs of the Interior Parts. |
| Open Source Code | Yes | Details of the pure C++/CUDA implementation and anonymous code release link will be found at A.3. |
| Open Datasets | Yes | We test the reconstruction and generation ability of our methods on the whole Shape Net (Chang et al., 2015) dataset. To validate the ability of out GALA representation on more complicated datasets, we conduct the textto-3D generation experiment on the Objaverse dataset (Deitke et al., 2023) and show the preliminary results here. We show reconstruction results on large shapes with more than a million triangles from Stanford 3D Scanning dataset (Turk & Levoy, 1994; Curless & Levoy, 1996; Krishnamurthy & Levoy, 1996; Gardner et al., 2003). |
| Dataset Splits | Yes | We follow the dataset split of previous works (Zhang et al., 2022; 2023; Yariv et al., 2023) and train conditional generation on the whole Shape Net classes. |
| Hardware Specification | Yes | The measurements were conducted under the configuration of 6 virtualized logical cores (hyper-thread) of AMD EPYC 7413 @2.65GHz and 1 Nvidia A100. In a 30 logical cores (virtualized) AMD EPYC 7J13 and 1 Nvidia A100 machine, the total fitting time will be less than 6 seconds. We also provide a reference record on a lower-end local desktop machine, configured with 1 Nvidia RTX 3060 and Intel i9-12900K, showing approximately 3 seconds for extraction and 27 seconds for refinement. |
| Software Dependencies | No | We show the algorithm flowchart Figure 17 of the initialization stage of GALA, which is built upon CUDA, Libtorch, a third-party R-Tree based SDF query library2 and few head only utility libraries such as C++ argparse3. The specific version numbers for these software components are not explicitly stated. |
| Experiment Setup | Yes | For the hyperparameter of GALA, we set tree root number No = 256, child node expanding ratio α = 0.2, grid resolution m = 5 and depth d = 1. The number of histogram bins H is 2m. For the refinement process, 8192 points are queried near surfaces on each run for 400 iterations. For the cascaded generation, each network for every generation step is consisted of 24 transformer layers with model channel 1024. Adam W (Loshchilov & Hutter, 2017) with lr = 8e 5, β1 = 0.9, β2 = 0.999 is adopted. Classifier-free guidance (Ho & Salimans, 2022) is used, with an additional null label trained as y during diffusion training, and w = 0 is applied during classifier-free guidance in diffusion inference. |