Why DDIM Hallucinates More Than DDPM: A Theoretical Analysis of Reverse Dynamics
A new theoretical analysis examines hallucination phenomena in diffusion models, specifically comparing the Denoising Diffusion Probabilistic Model (DDPM) and the Denoising Diffusion Implicit Model (DDIM). The study proves that DDIM can become stuck between modes after a critical time, while DDPM's inherent stochasticity helps it avoid this issue. The research suggests that introducing additional stochastic steps could improve DDIM's performance and reduce hallucinations. AI
IMPACT Provides theoretical insights into diffusion model behavior, potentially guiding the development of more robust generative models.