摘要
In this paper, we propose TextDestroyer, the first training- and annotation-free method for scene text destruction using pre-trained diffusion models. Unlike existing scene text removal approaches that require extensive annotation and retraining, our method leverages the inherent capabilities of pretrained diffusion models to achieve complete text obliteration while preserving background integrity. TextDestroyer employs a three-stage hierarchical process for precise text localization: introductory text capturing through cross-attention aggregation, continuous text adjustment via cropped sub-image processing, and meticulous text delineation using clustering analysis. For text destruction, we replace identified text regions with Gaussian noise in the latent space before reconstruction. During the diffusion denoising process, we introduce a novel background restoration mechanism that combines self-attention key-value guidance with strategic latent code replacement to ensure seamless background recovery. The key advantages of TextDestroyer include: (1) reducing the need for labor-intensive annotation and resource-intensive training requirements; (2) effective suppression of recognizable textual traces through a destroy-then-reconstruct paradigm; and (3) strong extensibility for rapid iteration and integration with other image editing methods.