Shengcai Liao

Diffusion editors routinely allocate less than 1% of their attention mass to conditioning images, a major identity bottleneck that can be resolved on the fly across multiple architectures using training-free dynamic logit offsets.

Achieving a harmonious balance between facial fidelity and overall appearance consistency could redefine personalized image generation.

FADRA achieves unprecedented facial detail recovery and temporal consistency in video face restoration, outperforming state-of-the-art techniques.