ReVersion proposes the task of Relation Inversion—extracting relational patterns from example image pairs and applying them to generate new images. The method uses a diffusion model backbone with specialized relation modeling to learn reusable relation prompts. It enables generalization from few examples to novel scenes while preserving the extracted relationship pattern.