UniRestore3D: A Scalable Framework For General Shape Restoration
Authors: Yuang Wang, Yujian Zhang, Sida Peng, Xingyi He, Haoyu Guo, Yujun Shen, Hujun Bao, Xiaowei Zhou
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The capabilities of our proposed method are demonstrated across multiple shape restoration subtasks and validated on various datasets, including Objaverse, Shape Net, GSO, and ABO. Our approach achieves SOTA results in tasks including noise-free shape completion, noisy shape refinement and completion. Related ablation studies have confirmed the effectiveness of key modules in our model. |
| Researcher Affiliation | Collaboration | 1State Key Lab of CAD&CG, Zhejiang University 2Ant Group |
| Pseudocode | No | The paper describes the methods, models, and training processes in detail across sections such as "3 GENERAL SHAPE RESTORAION" and "4.1 HIERARCHICAL SHAPE ENCODING", and "A.2 NETWORK ARCHITECTURE" but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The proposed method is built on publicly available codebases, including latent diffusion1 and other open-source diffusion model implementations2. Main network architectures are built upon the publicly available Torch Sparse codebase3. This information facilitates the reproduction of our method. |
| Open Datasets | Yes | The effectiveness of our method is validated on datasets including Objaverse (Deitke et al., 2023b;a), Shape Net (Chang et al., 2015), ABO (Collins et al., 2022), GSO (Downs et al., 2022) and Scan Net (Rao et al., 2022). |
| Dataset Splits | Yes | The training set (Objaverse subset) contains approximately 120K samples, and the training process takes about 14 days. The training set contains approximately 800K samples, and the entire training process takes about 3.5 days. For known categories, we use the Shape Net-13 (Liu et al., 2020) test set. |
| Hardware Specification | Yes | All our trainings conducts on 8 A100 GPUs. For inference on an NVIDIA A100 GPU, the 1st stage denoiser runs approximately 10 timesteps per second, resulting in about 10 seconds per sampling and a VRAM usage of 8 GB. |
| Software Dependencies | No | The proposed method is built on publicly available codebases, including latent diffusion1 and other open-source diffusion model implementations2. Main network architectures are built upon the publicly available Torch Sparse codebase3. This information facilitates the reproduction of our method. (Footnotes: 1https://github.com/Comp Vis/latent-diffusion 2https://github.com/lucidrains/denoising-diffusion-pytorch 3https://github.com/mit-han-lab/torchsparse) |
| Experiment Setup | Yes | The batch size is set to 8 8=64, and each iteration takes about 5 seconds. We use 8 A100 GPUs to train for 90 epochs. The batch size is 8 8=64, each iteration takes about 4.5 seconds... We employ the DDIM sampler with 100 timesteps for diffusion sampling in all stages. |