Reproduce “AttriCtrl: Fine-Grained Control of Aesthetic Attribute Intensity in Diffusion Models” 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

Die Chen, Zhongjie Duan, Zhiwen Li, Cen Chen, Daoyuan Chen, Yaliang Li, Yingda Chen

AttriCtrl presents a lightweight and generalizable framework for diffusion models, enabling precise, continuous control over the intensity of semantic attributes in generated images. The framework quantitatively improves control accuracy for aesthetic attributes like brightness, detail, and realism, achieving an average absolute difference of 0.175, and enhances content suppression with a 57.7% removal rate for unsafe content.

Key insights

  • It is feasible to quantify complex aesthetic attributes into continuous values and translate them into a model-interpretable format for precise control in diffusion models, overcoming limitations of discrete token processing.
  • A modular, plug-and-play value encoder design enables disentangled and compositional control of multiple semantic attributes simultaneously, allowing for flexible 'mixing-console'-like manipulation.
  • The framework provides superior fine-grained control and smoother transitions, addressing the limitations of global preference alignment and interpolation-based methods in generative AI.

Summary by alphaXiv.

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