3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing
Authors: Jiahua Dong, Yu-Xiong Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate the effectiveness of 3DGS-Drag in various scenes, achieving state-of-the-art performance in geometry-related 3D content editing. Notably, the editing is efficient, taking 10 to 20 minutes on a single RTX 4090 GPU. Our code is available at https://github.com/Dongjiahua/3DGS-Drag. ... 4 EXPERIMENT ... 4.2 QUALITATIVE EVALUATION ... 4.3 QUANTITATIVE EVALUATION ... Ablation Study |
| Researcher Affiliation | Academia | Jiahua Dong Yu-Xiong Wang University of Illinois Urbana-Champaign EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and figures (Figure 2 and Figure 3 illustrate the framework and multi-view edits, respectively), but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/Dongjiahua/3DGS-Drag. ... Our code is released at https://github.com/Dongjiahua/3DGS-Drag. |
| Open Datasets | Yes | our experiments include edits on eight scenes, using the published datasets from Instruct Ne RF2Ne RF Haque et al. (2023), PDS Koo et al. (2024), Mip-Ne RF360 Barron et al. (2022), and Tank and Temple Knapitsch et al. (2017). |
| Dataset Splits | No | The paper mentions using published datasets and selecting views for editing but does not provide specific train/validation/test dataset splits. For example, it states: 'During editing, 50 views are selected to enable efficient editing by default.' and 'The pretrained 3D Gaussians are trained with original 3D Gaussian Splatting (Kerbl et al., 2023)'. This implies usage of pre-trained models and a dynamic 'dataset' for view correction rather than traditional fixed train/test/validation splits for their own model training. |
| Hardware Specification | Yes | Notably, the editing is efficient, taking 10 to 20 minutes on a single RTX 4090 GPU. ... The running time is tested on a single RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions using 'Dreambooth model (Ruiz et al., 2023) with Lo RA (Hu et al., 2022)' and that 'The pretrained 3D Gaussians are trained with original 3D Gaussian Splatting (Kerbl et al., 2023)'. However, it does not specify version numbers for these models or any underlying software libraries (e.g., Python, PyTorch, CUDA). |
| Experiment Setup | Yes | During editing, 50 views are selected to enable efficient editing by default. ... We use batch size 4 and train for 200 iterations. After each dragging step, we continue fine-tuning the diffusion model for 50 iterations with the updated image buffer in each interval. ... The loss weight of λ1, λssim and λlpips are set to 8, 2, 1 respectively. |