Anatomy of a Hallucination: How Diffusion Solvers Shape Generation Errors

Diffusion models consistently produce stunningly realistic images, but building a reliable pipeline free of structural hallucinations forces us to look closely at our methods: do we rely on stochastic or deterministic sampling? Since stochastic methods inject noise that can cause the generation path to drift, deterministic methods often feel like the obvious fix. But does the empirical data actually support this intuition? This talk explores whether choosing a deterministic method genuinely prevents errors, or if it simply alters the specific hallucination profile. By dissecting the two sampling methods, we ask an open question: if we cannot eliminate generation errors entirely, which failures are we actually willing to accept?

(This talk is part of the weekly Machine Learning Lunch Meetings (MLLM), held every Tuesday from 12:15 to 1:15 p.m.  Professors from Computer Sciences, Statistics, ECE, the iSchool, and other departments will discuss their latest research in machine learning, covering both theory and applications. This is a great opportunity to network with faculty and fellow researchers, learn about cutting-edge research at our university, and foster new collaborations. For the talk schedule, please visit https://sites.google.com/view/wiscmllm/home. To receive future weekly talk announcements, please subscribe to our UW Google Group at https://groups.google.com/u/1/a/g-groups.wisc.edu/g/mllm.)