Appearance Pointers: Multimodal Region Control of Diffusion Transformers

This paper introduces Appearance Pointers, compact control tokens for localized multimodal guidance in Diffusion Transformers (DiTs). A region correspondence network aligns text or image inputs with user-provided masks, while spatial aggregation refines the resulting pointers and supports multiple regional descriptions without substantially increasing token load. The method is designed as a modality-agnostic interface that works without retraining the base model from scratch. According to the abstract, a single model matches or surpasses modality-specific state-of-the-art methods across several metrics, although the supplied material does not include the model configuration, benchmark names, numerical results, or implementation details.