Baoquan Zhao

Treating pretrained vision embeddings as temporary scaffolding to inject and internalize—rather than passive alignment targets—slashes diffusion transformer training budgets by over 43×.

Unconstrained sampling offsets routinely degrade dynamic spatial modeling—freezing the kernel's center and filtering deformations with high-frequency edge priors consistently outperforms DCNv4 without adding architectural bloat.

PhysScene reveals that structured representations of experimental scenes can significantly enhance relational reasoning capabilities in AI systems, challenging current benchmarks.