Jie Zhou

Ditch blurry FID scores: evaluating generative models directly in discrete token space reveals perceptual quality nuances that continuous feature metrics miss.

WaterVIB achieves superior zero-shot watermark robustness against generative AI attacks by learning a minimal sufficient representation, sidestepping the fragility of existing methods that entangle watermarks with easily-rewritten high-frequency cover textures.

By dynamically steering attention heads based on semantic context, LVLMs can significantly reduce hallucinations without any additional training.