Boston University
Diffusion models as representation learners: interpreting and leveraging their latent features
Abstract
dc:description.abstractDiffusion models are powerful generative frameworks capable of producing highly detailed visual content. Beyond synthesis, their intermediate representations encode rich semantic structure, making them valuable for downstream tasks. This thesis explores diffusion models as representation learners through two complementary lenses: interpretability and domain generalization. In the first part, we present Revelio, a framework that uses k-Sparse Autoencoders (k-SAE) to extract monosemantic, human-interpretable features and analyze how semantic concepts emerge across layers, timesteps, and architectures. We complement this with a discriminative analysis using lightweight classifiers (Diff-C) trained on frozen features, revealing how semantic features are packed in diffusion models. In the second part, we compare how different vision architectures and pretraining objectives structure class and domain-specific information. We find that diffusion models disentangle domain signals from class semantics. Building on this, we propose a domain generalization method that infers pseudo-domains via unsupervised clustering and integrates them into a standard classifier - achieving strong generalization without domain labels, fine-tuning, or test-time access.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Thomas, Xavier
- Advisor dc:contributor.advisor
-
- Ghadiyaram, Deepti
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53420