Zi-Yu Wan

Few-step diffusion models can achieve state-of-the-art 4-NFE generation without auxiliary score models or adversarial training, matching or exceeding multi-step teachers.

OPSD-V shows that incorporating real video context during training can dramatically enhance the performance of autoregressive video generators, leading to superior visual quality and motion fidelity.

Autoregressive Transformers can now generate high-fidelity, animatable 4D Gaussian avatars from single portrait images, offering a new paradigm for controllable avatar creation.