Reproduce “PROUD: PaRetO-gUided Diffusion Model for Multi-objective Generation” 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
Yinghua Yao, Yuangang Pan, Jing Li, Ivor Tsang, Xin Yao
The PROUD framework, developed by researchers from A*STAR and Lingnan University, enables multi-objective generation by adaptively guiding a pre-trained diffusion model towards Pareto optimal solutions while preserving high data quality. It achieved superior generation quality with an FID of 31.39 and an HV of 5.21 for 2-objective image generation on CIFAR10, alongside effective protein sequence design.
Key insights
- Treating generation quality as a hard constraint rather than an additional objective in multi-objective generation is crucial for maintaining sample realism and avoiding meaningless outputs.
- Dynamically adjusting the generative guidance based on a sample's proximity to the Pareto front allows for an adaptive strategy: prioritizing Pareto improvement when far, and focusing on generation quality when near.
- Leveraging post-hoc guidance on pre-trained unconditional diffusion models enables versatile multi-objective generation without requiring costly retraining or architectural modifications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.