He has recently been focusing on sparse attention and quantization techniques in Diffusion Transformers (DiT) to enable efficient, high-quality generative models.