Researchers have developed a method to improve the performance of AI-generated images in terms of both visual fidelity and query retrieval. The approach, called CERES, creates a three-level semantic pyramid that helps the generator produce images with scale-specific concepts. This is particularly important when dealing with scenes containing entities at vastly different scales. CERES has been tested on four pansharpening benchmarks and outperforms existing methods in terms of both visual quality and query retrieval accuracy. --- Why it matters: This matters to researchers in AI because it addresses a critical issue in multimodal information systems, where generated images need to remain retrievable by the queries they were meant to serve. CERES' ability to preserve queryable content rather than self-referential feature consistency has significant implications for applications such as image search and retrieval. Source: https://arxiv.org/abs/2608.20810