Parallel Diffusion Solver for Faster Image Generation

This paper addresses the critical issue of slow sampling in diffusion models, a major bottleneck for their practical application. It proposes a novel ODE solver, EPD-Solver, that leverages parallel gradient evaluations to accelerate the sampling process while maintaining image quality. The use of a two-stage optimization framework, including a parameter-efficient RL fine-tuning scheme, is a key innovation. The paper's focus on mitigating truncation errors and its flexibility as a plugin for existing samplers are also significant contributions.