Reproduce “Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target 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

Kim Yong Tan, Yueming Lyu, Ivor Tsang, Yew-Soon Ong

The Fast Direct algorithm introduces a query-efficient online black-box guidance method for diffusion models, utilizing a novel Guided Noise Sequence Optimization with a dynamically updated pseudo-target. This approach demonstrated 6x to 44x improvements in query efficiency for generating high-quality images aligned with complex prompts and optimizing 3D molecular structures for binding affinity.

Key insights

  • Guiding diffusion models by optimizing the noise sequence rather than model parameters offers an efficient and robust method, particularly effective even when the guidance target is noisy.
  • Updating noise along the data manifold using a universal direction (`x* - xK`) is crucial for preserving the high quality and semantic coherence of generated outputs.
  • A dynamically evolving "pseudo-target," inferred from black-box objective feedback, enables query-efficient online guidance without requiring differentiable objectives or large pre-collected datasets.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.