Xinyuan Chen

DeforM boosts video generation realism by directing attention to physics-critical regions, outperforming traditional models in both quality and consistency.

A unified definition of world models could catalyze breakthroughs across AI subfields by clarifying what these internal simulators should predict and how they should be constructed.

I2C-3D achieves unprecedented multi-view consistency in compositional 3D generation, outperforming existing methods and enabling realistic object interactions.

Forget rigid shot-music mappings: BEAT's elastic alignment framework finally captures the dynamic rhythm of professional movie trailer editing.

By intelligently pruning attention heads based on their spatial or temporal roles and adaptively routing denoising steps through the network, PARE achieves significant computational savings in video generation without sacrificing quality.

A single model now handles both image and video sketch colorization with high fidelity and temporal consistency, outperforming specialized methods.