Search papers, labs, and topics across Lattice.
This paper introduces a plug-and-play guided diffusion sampling method that uses adaptive moment estimation (Adam) to stabilize noisy likelihood score approximations. By applying Adam to the score estimates, the method effectively mitigates gradient noise during the reverse diffusion process. The proposed approach achieves state-of-the-art results on image restoration and class-conditional generation, surpassing more complex and computationally intensive techniques.
Ditch the complex diffusion sampling methods: a simple Adam optimizer can give you state-of-the-art image generation and restoration results.
Guided diffusion sampling relies on approximating often intractable likelihood scores, which introduces significant noise into the sampling dynamics. We propose using adaptive moment estimation to stabilize these noisy likelihood scores during sampling. Despite its simplicity, our approach achieves state-of-the-art results on image restoration and class-conditional generation tasks, outperforming more complicated methods, which are often computationally more expensive. We provide empirical analysis of our method on both synthetic and real data, demonstrating that mitigating gradient noise through adaptive moments offers an effective way to improve alignment.
