Agentic AI Pipeline Improves Shadow Removal in Image Editing

A recent paper published on arXiv (2608.06075) explores the potential of commercial vision-language models to lessen reliance on traditional, physics-based low-level vision techniques, particularly for shadow removal. The findings indicate that although a commercial generative editor can create shadow-free images that maintain surface texture and local appearance, it introduces a new type of error: it may regenerate scene elements, fabricate objects, or misinterpret shadows as material or geometry, resulting in edits that, while plausible, are physically inaccurate. To tackle this issue, the authors suggest an agentic candidate-selection pipeline, where the editor produces a guided probe, an evaluator checks for significant errors, retries as necessary, and samples various candidates. This research emphasizes both the advantages and drawbacks of employing large vision-language models in low-level vision tasks, indicating that they can compete with specialized systems but still need oversight to ensure physical correctness.