DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention | Read Paper on Bytez