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

Muhammad H. Ashiq, Samanyu Arora, Abhinav N. Harish, Ishaan Kharbanda, Hung Yun Tseng, Grigorios G. Chrysos · 2026

Read Why DDIM Hallucinates More Than DDPM: A Theoretical Analysis of Reverse Dynamics by Muhammad H. Ashiq, Samanyu Arora, Abhinav N. Harish, Ishaan Kharbanda, Hung Yun Tseng, Grigorios G. Chrysos (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06831.txt.

Introduction

Muhammad H. Ashiq and five co-authors explain theoretically why the deterministic DDIM sampler hallucinates more than the stochastic DDPM sampler in diffusion models. Both reverse a noising process, one by an ordinary differential equation and one with added randomness. Analyzing both for a Gaussian mixture target, the authors prove that after a critical time the deterministic sampler can become stuck on the segment between the two nearest modes, fabricating in-between content.

The stochastic sampler's noise, by contrast, knocks it free of that trap region. Experiments confirm a significantly lower hallucination rate for the stochastic method whenever the trap is entered. Adding extra stochastic steps to the deterministic sampler also helps it escape, informing better sampler design.

Catalog record

CREATORMuhammad H. Ashiq, Samanyu Arora, Abhinav N. Harish, Ishaan Kharbanda, Hung Yun Tseng, Grigorios G. Chrysos
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE157 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06831

Browse the full library · About the Tower of Babel