Zanyi Wang

Fine-tuning efficient few-step diffusion models no longer requires sacrificing their speed, thanks to a self-distillation approach that preserves inference capabilities.

Time conditioning, a seemingly crucial component of diffusion models like DDIM, can be entirely bypassed without sacrificing generation quality by carefully shaping the evolution of noisy data manifolds.

Achieve strong spatio-temporal video grounding with only 10M trainable parameters by smartly adapting pre-trained 2D visual-language models.