Structure Alignment Resampler: More Structure for the Specific ID Generation

The text-to-image generation methods based on diffusion models have generated a strong response in both industry and academia. As an important application scenario, ID-Driven Generation has become the focus of many researchers' efforts. However, most current methods often encounter the issue of structural misalignment. This is specifically demonstrated by the reference ID image's structural information influencing the structural details of the target image during the generation process. To enhance structural control over generated images, we introduce a novel structure in the image generation process called the Structure Alignment Resampler (SAR). This structure aligns the main subject ID information from the reference image with the structural information contained in the target text description, thereby improving structural control during image generation. To fully utilize the potential of this structure, we also propose a dataset construction method to avoid the occurrence of "copy-paste" phenomena. Additionally, we incorporate a structure-loss during training to further align the ID injection effects in subject-related regions. We evaluated our method on a self-constructed test dataset, and the results indicate that our approach shows promising performance in both extracting ID subject features and controlling target structures.