GSGAN: Adversarial Learning for Hierarchical Generation of 3D Gaussian Splats

Authors: Sangeek Hyun, Jae-Pil Heo

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that ours achieves a significantly faster rendering speed ( 100) compared to state-of-the-art 3D consistent GANs with comparable 3D generation capability.
Researcher Affiliation Academia Sangeek Hyun Sungkyunkwan University Jae-Pil Heo Sungkyunkwan University
Pseudocode No The paper describes its methods using text and figures but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes First of all, please refer to the attached code in supplementary for a more detailed implementation of the proposed method.
Open Datasets Yes Following the experimental settings of previous 3D GANs [12, 25], we use FFHQ [35] and AFHQ-Cat [50] datasets with 256 256 and 512 512 resolutions.
Dataset Splits No The paper states using FFHQ [35] and AFHQ-Cat [50] datasets and follows experimental settings of previous 3D GANs [12, 25], but it does not explicitly provide specific training/validation/test dataset split percentages or sample counts within its text.
Hardware Specification Yes Rendering time is measured on a single RTX A6000 GPU. ... Additionally, the training time of the proposed methods is 28 RTX A6000 days, while the state-of-the-art Mimic3D requires 64 A100 days on the FFHQ-512 dataset.
Software Dependencies No The paper mentions various models and techniques like StyleGAN2 and Neus2 but does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes For the coefficients of objective functions, we use λ = 1, λpose = 1, and λknn = 10. We train the model until the discriminator sees 10-15M images. ... For the architectural details, we use an upsampling ratio r = 4 and the initial number of Gaussians N = 256. For the number of hierarchical levels, we adopt L = 5 for the dataset with 256 256 resolution, and L = 6 for 512 512 resolution.