DreamTime: An Improved Optimization Strategy for Diffusion-Guided 3D Generation

Authors: Yukun Huang, Jianan Wang, Yukai Shi, Boshi Tang, Xianbiao Qi, Lei Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our simple redesign significantly improves 3D content creation with faster convergence, better quality and diversity.We conduct experiments on the generation of 2D images and 3D assets for a comprehensive evaluation of the proposed time prioritized score distillation sampling (TP-SDS).
Researcher Affiliation Collaboration 1International Digital Economy Academy (IDEA) 2The University of Hong Kong
Pseudocode Yes Algorithm 1: Time Prioritized SDS (TP-SDS).
Open Source Code Yes Our implementation are based on the publicly-accessible threestudio codebase (Guo et al., 2023)
Open Datasets Yes Given the 153 text prompts from object-centric COCO validation set, we show in Figure 8 the R-Precision scores of 2D generation results at different iteration steps using the vanilla SDS and our TP-SDS.
Dataset Splits No The paper mentions using the 'COCO validation set' for evaluation, but does not provide explicit details about its own training, validation, and test splits for model reproduction (e.g., percentages, counts, or splitting methodology).
Hardware Specification Yes Magic3D: 5.3 A100 hours Fantasia3D: 6 RTX3090 hours
Software Dependencies No The paper mentions basing its implementation on the 'publicly-accessible threestudio codebase' but does not provide specific version numbers for this or any other software dependencies.
Experiment Setup Yes Input: A differentiable generator g with initial parameters θ0 and number of iteration steps N, pre-trained diffusion model ϕ, prior weight function W(t), learning rate lr, and text prompt y.We emphasize that we adopt the same hyper-parameter configuration {m = 500, s = 125} for both TP-SDS and TP-VSD throughout this paper.