Real3D: The Curious Case of Neural Scene Degeneration
Authors: Dengsheng Chen, Jie Hu, Xiaoming Wei, Enhua Wu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results demonstrate that Real3D outperforms all existing state-of-the-art text-to-3D generation methods, providing valuable insights to facilitate the development of learning-based 3D scene generation approaches.Experiments We focus on comparing our method with existing text-to-3D methods, i.e., Dream Fusion (Poole et al. 2022a), SJC (Wang et al. 2022) and Magic3D (Lin et al. 2022) on the text prompts taken from the online website1. |
| Researcher Affiliation | Collaboration | Dengsheng Chen1, Jie Hu1, Xiaoming Wei1, Enhua Wu2,3,4 1Meituan 2State Key Laboratory of Computer Science, ISCAS 3University of Chinese Academy of Sciences 4University of Macau {chendengsheng, hujie39, weixiaoming}@meituan.com, ehwu@um.edu.mo |
| Pseudocode | No | The paper describes methods using mathematical equations and textual explanations, but no explicitly labeled 'Pseudocode' or 'Algorithm' block is present. |
| Open Source Code | No | As some algorithms do not have publicly-available 3D models or opensource code, we compared the performance of various methods int the rest four rows using replicated results from their original papers. |
| Open Datasets | Yes | We focus on comparing our method with existing text-to-3D methods, i.e., Dream Fusion (Poole et al. 2022a), SJC (Wang et al. 2022) and Magic3D (Lin et al. 2022) on the text prompts taken from the online website1. ... We randomly selected 100 prompts from the 397 prompts provided by Dream Fusion for generation and ensured that each prompt generated reasonable results for all algorithms. |
| Dataset Splits | No | The paper refers to using prompts for generation and comparison, but it does not specify any training, validation, or test dataset splits in terms of percentages or sample counts for the model's development. |
| Hardware Specification | Yes | The entire process was carried out on a single A100 GPU, and it took approximately 25 minutes to complete. |
| Software Dependencies | No | The paper mentions using 'Adan (Xie et al. 2022) optimizer' and a 'frozen stable diffusion model pϕ' but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We use Adan (Xie et al. 2022) optimizer with an initial learning rate of 5 10 3. A cosine annealing learning rate scheduler is applied with a minimum learning rate of 1 10 6. In all experiments, we set λSDS = 1 and λadv = 1 10 2. ... The SDF threshold ε is set as 0.1. |