GaussianCube: A Structured and Explicit Radiance Representation for 3D Generative Modeling
Authors: Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, Baining Guo
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on unconditional and classconditioned object generation, digital avatar creation, and text-to-3D synthesis all show that our model achieves state-of-the-art generation results both qualitatively and quantitatively, underscoring the potential of Gaussian Cube as a highly accurate and versatile radiance representation for 3D generative modeling. |
| Researcher Affiliation | Collaboration | 1University of Science and Technology of China 2Tsinghua University 3Microsoft Research Asia |
| Pseudocode | No | The paper does not contain any sections or figures explicitly labeled as 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Project page: https://gaussiancube.github.io/. We also release our code and model checkpoints at https://github.com/Gaussian Cube/Gaussian Cube. |
| Open Datasets | Yes | The model s capability for unconditional and class-conditioned generation is evaluated on the Shape Net [9] and Omni Object3D [64] datasets. |
| Dataset Splits | No | The paper describes training steps and mentions evaluating on 'test' data, but it does not explicitly provide details for a separate 'validation' split (e.g., specific percentages or counts for training, validation, and testing phases) across all datasets used. |
| Hardware Specification | Yes | Our approach begins with the aim of maintaining a constant number of Gaussians g RNmax C across different objects during the fitting. ... It takes about one week to train our model on Shape Net Car, Shape Net Chair, and Omni Object3D, and approximately two weeks for the Synthetic Avatar and Objaverse datasets. ... Inference time is measured on a single A100 GPU. ... We deploy 16 Tesla V100 GPUs for the Shape Net Car, Shape Net Chair, Omni Object3D, and Synthetic Avatar datasets, whereas 32 Tesla V100 GPUs are used for training on the Objaverse dataset. |
| Software Dependencies | No | The paper mentions software components and tools like 'ADM U-Net', 'Adam W optimizer', 'DPM-solver', 'DINO ViT-B/16', and 'CLIP-L/14' but does not specify their version numbers or the versions of underlying software environments like Python or PyTorch. |
| Experiment Setup | Yes | For Gaussian Cube construction, we set Nmax to 32,768 and C to 14 across all datasets. We perform the proposed densification-constrained fitting for 30K iterations...The timesteps of diffusion models are set to 1, 000 and we train the models using the cosine noise schedule [41] with loss weight λ set to 10. We train our model using Adam W optimizer [33], and apply exponential moving average (EMA) with a rate of 0.9999 during training. ... For unconditional generation on Shape Net, we train the model with a base learning rate 5e 5 for 850K iterations and then decay the learning rate to 5e 6 for another 150K iterations. |