Researchers have conducted the first zero-shot benchmark of unified multimodal large language models (MLLMs) for assessing and restoring distortions in shallow-water unmanned aerial vehicle (UAV) imagery. The study evaluated five open-weight MLLMs on 1,000 image pairs, finding significant variation in distortion assessment accuracy based on the model and prompt used, with a best pooled balanced accuracy of 0.654 for distortion presence. Restoration capabilities proved limited, with the best models showing minimal improvement in image quality metrics and some even altering scene structure or hallucinating content. The findings highlight that while MLLMs can recognize degradation, their ability to reliably restore imagery is not yet established. AI
IMPACT Highlights limitations of current multimodal LLMs in practical image restoration tasks, suggesting future research directions.
RANK_REASON Academic paper detailing a novel benchmark for multimodal LLMs on a specific computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
