LCGen: Mining in Low-Certainty Generation for View-consistent Text-to-3D

Authors: Zeng Tao, Tong Yang, Junxiong Lin, Xinji Mai, Haoran Wang, Beining Wang, Enyu Zhou, Yan Wang, Wenqiang Zhang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this Section, we apply LCGen to several baseline methods of SDS-based text-to-3D, including Dream Fusion [25], Magic3D [16], and Prolific Dreamer [34], and conduct corresponding experiments. We also compare with other methods that address the Janus Problem. Sec. 5.2 presents the effects of original methods and LCGen, including qualitative and quantitative assessments. Sec. 5.3 demonstrates the ablation of hyperparameters. Furthermore, Sec. 5.4 presents visualization results of LCGen s impact on generation certainty.
Researcher Affiliation Academia 1Shanghai Engineering Research Center of AI & Robotics, Academy for Engineering & Technology, Fudan University, Shanghai, China 2Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University, Shanghai, China 3School of Computer Science, Fudan University, Shanghai, China 4Engineering Research Center of AI & Robotics, Ministry of Education, Academy for Engineering & Technology, Fudan University, Shanghai, China {ztao19,yanwang19,wqzhang}@fudan.edu.cn
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. It provides mathematical equations and descriptions.
Open Source Code No Code will be available [here] if accepted.
Open Datasets Yes We selected two sets of text prompts from the library [25] and conducted experiments on three SDS-based text-to-3D baseline methods, including Dream Fusion-sd [25], Magic3D-sd coarse [16], and Prolific Dreamer [34]... We selected 30 sets of text prompts from the library and calculate the mean score.
Dataset Splits No The paper does not explicitly specify training, validation, and test dataset splits by percentages or counts.
Hardware Specification Yes We implement original methods and LCGen based on threestudio [8] and a single A100 GPU.
Software Dependencies Yes For the sake of experimental consistency, we have chosen the Stable Diffusion 2.1 base [27] as guidance and Ne RF [24] as the 3D representation in the SDS-based method.
Experiment Setup Yes In the experiment, we set G(c) = |ϕ| and γ to 10, and obtained the results after a maximum of 10,000 steps. For Dreamfusion, Magic3D, and Prolific Dreamer, the hyperparameters are all kept consistent with the default settings in config files in the repository. Specifically, according to the instructions in threestudio, In Dreamfusion, to prevent scene being stuffed with floaters/becoming empty, we set system.loss.lambda_sparsity=0.1. In Dreamfusion and Magic3D, to prevent the model incorrectly treating the background as part of the object, we replace the background with random colors with a probability 0.5 by setting system.background.random_aug=true.