Reproduce “DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling” 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

Xin Xie, Dong Gong

Researchers at UNSW Sydney developed DyMO, a training-free framework that dynamically aligns text-to-image diffusion models with contextual semantics and human preferences during inference. DyMO consistently improves image quality and prompt alignment, achieving higher PickScore, HPSv2, and ImageReward metrics compared to existing training-free and many training-based methods, without requiring additional model training.

Key insights

  • Semantic content and layout are largely determined in early denoising steps, where cross-attention maps provide effective supervision through LLM-derived semantic graphs.
  • Dynamic scheduling of multi-objective guidance (semantic and preference) across different denoising stages optimizes image generation by focusing on structural formation early and detail refinement later.
  • Inference-time guidance effectiveness is enhanced by adaptive mechanisms like Polyak step size and dynamic recurrent iterations, which improve convergence and efficiency.

Summary by alphaXiv.

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