Why DDIM Hallucinates More than DDPM: A Theoretical Analysis of Reverse Dynamics

Researchers provide theoretical analysis demonstrating that DDIM (deterministic diffusion model) generates more hallucinations than DDPM (stochastic diffusion model) when sampling from multi-modal distributions. The study proves that stochastic noise in DDPM helps escape local modes, while DDIM can become trapped between modes, with implications for improving generative AI sampling algorithms.

This theoretical paper addresses a fundamental challenge in diffusion-based generative models, which have become central to modern AI image generation and other applications. The researchers dissect why DDIM, despite being computationally faster through deterministic sampling, produces more hallucinations—synthesized content that doesn't match the target distribution—compared to its stochastic counterpart DDPM. Through mathematical analysis of Gaussian mixture models, they identify a critical transition point after which DDIM's deterministic trajectory can trap the sampling process between modes, creating false or corrupted outputs. DDPM avoids this through inherent stochasticity that provides escape routes from these local minima. This finding bridges theory and practice, as hallucination in generative models directly impacts real-world applications from medical imaging to content generation, where accuracy is paramount. The broader context involves the ongoing tension between computational efficiency (favoring deterministic methods) and sample quality (favoring stochastic approaches). The research provides actionable insights for algorithm designers seeking to improve diffusion samplers by strategically incorporating stochastic steps into otherwise deterministic frameworks. For practitioners deploying these models, understanding these failure modes enables better architectural choices when optimizing for either speed or quality. Looking forward, this work establishes a foundation for hybrid sampling strategies that combine the efficiency advantages of DDIM with the robustness of DDPM's stochasticity, potentially advancing the reliability of generative AI systems across commercial applications.

• →DDIM's deterministic approach can trap sampling between distribution modes, causing hallucinations after critical time points
• →DDPM's stochasticity provides the mechanism to escape local minima and maintain fidelity to target distributions
• →Theoretical framework reveals mathematical conditions under which diffusion samplers fail or succeed
• →Hybrid approaches incorporating strategic stochastic steps could improve sampler design without sacrificing computational efficiency
• →Findings directly applicable to improving generative model reliability in production systems