Reproduce “Unifying Layout Generation with a Decoupled Diffusion Model” on OpenResearch

About this paper

Mude Hui, Zhizheng Zhang, Xiaoyi Zhang, Wenxuan Xie, Yuwang Wang, Yan Lu

This paper introduces the Layout Diffusion Generative Model (LDGM), a unified diffusion model that addresses the versatility and control limitations in graphic layout generation. LDGM employs a decoupled diffusion process for different attribute types and a joint transformer-based denoising process, enabling it to handle a comprehensive range of tasks from unconditional generation to complex combinations of missing, coarse, or precise attributes. The model achieves superior performance on various layout generation subtasks and produces more aesthetically pleasing layouts compared to prior state-of-the-art methods.

Key insights

• Missing or coarse layout attributes can be effectively modeled as corrupted states in a forward Markov diffusion process, allowing layout generation to be framed as a reverse denoising task.
• Decoupling the diffusion processes for different attribute types (e.g., uniform noise for categories, Gaussian noise for geometric attributes) is crucial for accurately modeling their distinct semantics and generating diverse, high-quality layouts.
• A confidence-based inference strategy, which adaptively re-marks low-confidence predictions as corrupted for further refinement, is effective in preventing error propagation during the denoising process.

Summary by alphaXiv.