TopoDiff: Training-free image generation with topological layout control

Abstract

Recent diffusion models can generate high-quality images from text, but their spatial control remains limited. To address this, the goal is to enhance layout control in text-to-image generation without requiring retraining of existing models. Specifically, the proposed TopoDiff framework is a training-free approach that leverages topological guidance to enable precise spatial control during inference. It leaves the original architecture and parameters of Stable Diffusion unmodified. This approach employs a graph-based topological language to explicitly capture object spatial relationships while integrating topological loss into the diffusion model’s denoising process. Additionally, a dynamic offset mechanism is designed to adjust spatial positions during generation, balancing topological structure consistency with the flexibility required for complex generation. Experimental results demonstrate that TopoDiff achieves over 10 % higher Average Precision (AP) than the Stable Diffusion. The source codes are publicly available at https://github.com/marcocst/TopoDiff.