Back to search

University of North Texas

Interpretable Concept Evolution Flow in Diffusion Models

Abstract

dc:description

Diffusion- and flow-based generative models can be viewed as trajectory processes that transport samples from a base distribution toward a data distribution through a sequence of local updates. Although these updates determine the final sample, their step-wise roles remain difficult to interpret from the final output alone. Our work introduces Concept Evolution Flow, a training-free, trajectory-level perturbation framework for analyzing how individual guided update components influence generation over time. We decompose each guided update into a target-related component and a generic-related component, perturb one component at a selected sampling step, keep the remaining trajectory unchanged, and use the resulting final sample as a human-readable visual output of that local perturbation. Sweeping the perturbation step produces an ordered sequence of final readouts that reveals how structures emerge, transform, stabilize, or become unstable along the trajectory. We evaluate Concept Evolution Flow on Stable Diffusion v1.5, Stable Diffusion v2.1, and Stable Diffusion 3.5 Medium under a shared prompt and seed protocol. For generic-related perturbations, the readouts suggest recurring phase-dependent behavior: early perturbations expose competing or unstable structures, middle perturbations reveal transformations of coexisting structures, and late perturbations mainly affect local refinement. Target-related perturbations show a different response: in score-based diffusion models, coarse structure often remains comparatively stable despite strong artifacts, whereas the flow-matching transformer exhibits more model-specific and non-monotonic readouts. Finally, we observe a near-terminal unstable phase in score-based diffusion models, where perturbing either component near the final steps can produce sudden non-semantic artifacts followed by abrupt recovery. These findings suggest that temporal update-wise perturbation provides a direct and interpretable lens for studying generative dynamics, while also exposing open questions that require future update-space analysis.

Degree

thesis:*
Grantor dc:publisher
University of North Texas
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Liu, Yuxuan 1997-
Contributors dc:contributor
  • Yuan, Jing
  • Choudhury, Sagnik Ray
  • Cui, Xinrui

Subjects

dc:subject × 3

Rights

dc:rights
Statement dc:rights
  • Public
  • Liu, Yuxuan 1997-
  • Copyright
  • Copyright is held by the author, unless otherwise noted. All rights Reserved.
Language dc:language
English

Identifiers

dc:identifier.*
Identifier
local-cont-no: submission_5051
ark: ark:/67531/metadc2652541

Chain of custody

source
Harvested from
University of North Texas
Base URL
digital.library.unt.edu/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Liu, Yuxuan 1997-. Interpretable Concept Evolution Flow in Diffusion Models. University of North Texas, 2026. https://digital.library.unt.edu/ark:/67531/metadc2652541/