Reproduce “Few-step Flow for 3D Generation via Marginal-Data Transport Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zanwei Zhou, Taoran Yi, Jiemin Fang, Chen Yang, Lingxi Xie, Xinggang Wang, Wei Shen, Qi Tian
Researchers from Shanghai Jiao Tong University, Huazhong University of Science and Technology, and Huawei Inc. developed MDT-dist, a framework that accelerates 3D flow models to produce high-fidelity 3D assets in 2-4 inference steps, achieving up to a 9.0x speedup over the baseline while preserving visual and geometric fidelity.
Key insights
- Directly learning marginal-data transport can offer a more stable and effective optimization pathway for few-step generation compared to indirect consistency constraints or direct noise-to-data mapping.
- The inherent intractability of integrating velocity fields for direct marginal-data transport can be circumvented by deriving equivalent, optimizable objectives at the velocity (VM) and distribution (VD) levels.
- Velocity Matching and Velocity Distillation are complementary mechanisms; VM provides stable gradient estimates across all time steps, while VD refines one-step generation by addressing distribution discrepancies, leading to enhanced overall performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.