Youngjin Shin

MORPHOS achieves unprecedented temporal consistency in 3D asset generation by leveraging a unified 4D representation that adapts to evolving topologies.

Achieve superior real-world super-resolution by adaptively emphasizing detail during training and refining reconstruction with frequency-aware enhancers, all without retraining.

By explicitly aligning attention with external correspondences, CORAL significantly improves detail preservation in virtual try-on, addressing a key limitation of existing Diffusion Transformer methods.