Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization

Authors: Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, Tat-Seng Chua

ICML 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that Dream DPO achieves competitive results, and provides higher-quality and more controllable 3D content compared to existing methods. In this section, a series of experiments are conducted to justify our claims.
Researcher Affiliation Academia 1State Key Laboratory of Brain-machine Intelligence, Zhejiang University, China 2Re LER, CCAI, Zhejiang University, China 3National University of Singapore, Singapore. Correspondence to: Yi Yang <EMAIL>, Xiaobo Xia <EMAIL>.
Pseudocode Yes Algorithm 1 Pseudo-code for Dream DPO
Open Source Code Yes Code is publicly available at: https://github.com/Zhenglin Zhou/Dream DPO.
Open Datasets Yes We here evaluate the proposed method with 110 prompts from GPTEval3D (Wu et al., 2024a)
Dataset Splits No The paper evaluates on 110 prompts from GPTEval3D (Wu et al., 2024a) but does not specify how these prompts, or any other data, are split into training, validation, or testing sets. The training process for Dream DPO involves online generation of pairwise examples rather than using predefined dataset splits.
Hardware Specification Yes The optimization process takes around two hours on a single NVIDIA RTX A6000 GPU.
Software Dependencies No We conduct experiments using Py- Torch (Paszke et al., 2019) and threestudio (Guo et al., 2023), with MVDream (Shi et al., 2023) as the backbone of our method. Specific version numbers for PyTorch or other key libraries are not provided.
Experiment Setup Yes We follow the training strategy of MVDream and use HPSv2 (Wu et al., 2023) as the default reward model. The optimization process takes around two hours on a single NVIDIA RTX A6000 GPU. In Equation (6), τ = 0.001 is a pre-defined threshold. To speed up the process, we can adopt a simple yet effective strategy: performing SDS for the first 6000 iterations and then switching to Dream DPO.