Lingyu Zhu

By transforming the Retinex decomposition into an additive problem in latent space, this method achieves more stable and precise low-light image enhancement.

AI-generated educational videos often miss the mark on accurately representing the concepts they're supposed to teach, and this new benchmark dataset and evaluation framework can help close that gap.

Achieve better image compression without sacrificing quality by adapting diffusion models to image content using uncertainty-guided quantization and bitrate-free textual conditioning.