Atlas3D: Physically Constrained Self-Supporting Text-to-3D for Simulation and Fabrication
Authors: Yunuo Chen, Tianyi Xie, Zeshun Zong, Xuan Li, Feng Gao, Yin Yang, Ying Nian Wu, Chenfanfu Jiang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We verify Atlas3D s efficacy through extensive generation tasks and validate the resulting 3D models in both simulated and real-world environments. ... In this section, we devise comprehensive experiments (both virtual and real-world) to demonstrate the efficacy of our method. |
| Researcher Affiliation | Collaboration | 1University of California, Los Angeles, 2 Amazon, 3 University of Utah |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not include an explicit statement providing a link to open-source code for the described methodology. |
| Open Datasets | Yes | We randomly select 150 prompts from [58] and manually exclude 43 prompts deemed unfeasible (for instance, it does not make sense to require a swan and its cygnets swimming in a pond to be standable), leaving a total of 107 prompts. |
| Dataset Splits | No | The paper describes a two-stage training strategy and uses terms like 'refine stage' and 'evaluation', but it does not specify explicit training/validation/test dataset splits (e.g., percentages or sample counts) for reproducibility. |
| Hardware Specification | Yes | We train our models using a single NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | We implement our pipeline in Py Torch with Adam optimizer. For differentiable simulation, we adopt the semi-implicit Euler simulator in Warp [44]... We utilize Nvdiffrast [26] as the differentiable renderer and Stable Diffusion v2.1 [64] for guidance. The paper only provides a specific version for Stable Diffusion (v2.1); PyTorch, Adam optimizer, Warp, and Nvdiffrast are mentioned without specific version numbers. |
| Experiment Setup | Yes | We use a two-stage training strategy: the coarse stage and the refine stage. ... In our quantitative evaluation of a batch of prompts, we observed an average refinement time of 36 minutes for each training step, with a default setting of 5,000 iterations. ... For the rigid body simulator in Warp, we set dt = 10 3s. Contact stiffness and damping are set to 103 and 2.0; friction coefficient is set to 0.5; stiffness of friction force is set to 103. Density of the 3D objects is set to 103. ... In our experiments, we use the following default weights for the loss terms: {λSDS = 1, λnormal = 104, λstand = 105, λstable = 105, λb-lap = 107}. |