New framework fixes GenAI image editing hallucinations and content shifts

Generative AI image editors like Nano Banana let users retouch photos via text prompts, but often introduce spatial misalignment, texture distortion, or hallucinated content that break pixel-level accuracy needed for professional workflows. To solve this, Luxi Zhao and Michael S. Brown introduce a new problem setting called “structure-preserving GenAI fusion.” They propose a post-processing framework that takes both the original input image and the GenAI-enhanced output, establishes coarse spatial and photometric correspondences between them, and then performs a fusion stage that transfers the desired enhancements (e.g., improved lighting or color) while suppressing hallucinated details. The method works as a black-box solution, meaning it can be applied to any GenAI editor without modifying the underlying model.

In experiments, the researchers compared their framework against representative methods from photorealistic style transfer and image fusion. Their approach outperformed baselines in preserving aesthetic quality while maintaining pixel-level structural consistency and the original input resolution. The framework also does not require specialized hardware or retraining, making it practical for integration into existing pipelines. This addresses a critical gap: as GenAI image tools become common, ensuring that enhancements do not compromise fidelity is essential for applications like advertising, forensics, and medical imaging where every pixel matters.