DDIM Hallucinates More Than DDPM: Theoretical Analysis of Reverse Dynamics

A recent study published on arXiv in the Computer Science > Machine Learning category reveals that DDIM exhibits a higher rate of hallucination compared to DDPM in diffusion samplers. The research focuses on the reverse ordinary differential equation (ODE) of DDIM and the stochastic differential equation (SDE) of DDPM when applied to a Gaussian mixture target. After a critical time τ, DDIM tends to become stuck between two modes, while DDPM's inherent stochasticity allows it to escape this state effectively. The findings suggest that incorporating stochastic steps could enhance DDIM's performance and mitigate hallucinations, offering valuable insights for future sampler designs.