Reproduce “Asynchronous Denoising Diffusion Models for Aligning Text-to-Image 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
Zijing Hu, Yunze Tong, Fengda Zhang, Junkun Yuan, Jun Xiao, Kun Kuang
Zhejiang University and Nanyang Technological University researchers introduced Asynchronous Denoising Diffusion Models (AsynDM), a framework that improves text-to-image alignment by enabling pixel-level timestep allocation and adaptive scheduling in diffusion models. This method yielded consistently higher scores across four text-to-image alignment metrics compared to existing baselines, while maintaining similar image quality and sampling efficiency.
Key insights
- Synchronous denoising, which treats all pixels uniformly, is a fundamental factor contributing to text-to-image misalignment by limiting clear contextual information during semantic refinement.
- Differentiating denoising rates, allowing prompt-unrelated regions to clarify faster, provides superior contextual cues for more precise refinement of prompt-related elements.
- Pixel-level timestep allocation can be achieved within the existing diffusion framework by replacing scalar timesteps with a timestep tensor, maintaining the Markov property.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.