Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models

Arian Komaei Koma, Seyed Amir Kasaei, Ali Aghayari, AmirMahdi Sadeghzadeh, Mohammad Hossein Rohban

Abstract

Post-hoc unlearning has emerged as a practical mechanism for removing undesirable concepts from large text-to-image diffusion models. However, prior work primarily evaluates unlearning through erasure success; its impact on broader generative capabilities remains poorly understood. In this work, we conduct a systematic empirical study of concept unlearning through the lens of compositional text-to-image generation. Focusing on nudity removal in Stable Diffusion 1.4, we evaluate a diverse set of state-of-the-art unlearning methods using T2I-CompBench++ and GenEval, alongside established unlearning benchmarks. Our results reveal a consistent trade-off between unlearning effectiveness and compositional integrity: methods that achieve strong erasure frequently incur substantial degradation in attribute binding, spatial reasoning, and counting. Conversely, approaches that preserve compositional structure often fail to provide robust erasure. These findings highlight limitations of current evaluation practices and underscore the need for unlearning objectives that explicitly account for semantic preservation beyond targeted suppression.

Erasure or Erosion? Evaluating Compositional Degradation in Unlearned Text-To-Image Diffusion Models

Abstract

Post-hoc unlearning has emerged as a practical mechanism for removing undesirable concepts from large text-to-image diffusion models. However, prior work primarily evaluates unlearning through erasure success; its impact on broader generative capabilities remains poorly understood. In this work, we conduct a systematic empirical study of concept unlearning through the lens of compositional text-to-image generation. Focusing on nudity removal in Stable Diffusion 1.4, we evaluate a diverse set of state-of-the-art unlearning methods using T2I-CompBench++ and GenEval, alongside established unlearning benchmarks. Our results reveal a consistent trade-off between unlearning effectiveness and compositional integrity: methods that achieve strong erasure frequently incur substantial degradation in attribute binding, spatial reasoning, and counting. Conversely, approaches that preserve compositional structure often fail to provide robust erasure. These findings highlight limitations of current evaluation practices and underscore the need for unlearning objectives that explicitly account for semantic preservation beyond targeted suppression.

Paper Structure

This paper contains 11 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Analysis: Compositional Alignment In Unlearned Models
  4. Experimental Setup
  5. Compositional Performance of Unlearning Methods
  6. Performance on T2I-CompBench++
  7. Performance on GenEval
  8. Unlearning vs. Compositional Generalization
  9. Qualitative Analysis: Evidence of Manifold Collapse
  10. Conclusion
  11. Qualitative Visualization

Figures (2)

  • Figure 1: Qualitative comparison of unlearning methods trained to remove nudity, evaluated on a distant, safe prompt ("a green banana and a brown dog"). While the base model preserves structure, several unlearning methods exhibit object omission or attribute leakage, indicating that safety constraints have corrupted fundamental semantic syntax.
  • Figure 2: Qualitative comparison of compositional generation behavior across different unlearning methods. While ACE and SPM preserve structure, aggressive methods (e.g., EraseDiff, SalUn) exhibit mode collapse (repetitive scenes despite differing seeds) or object loss, failing to bind attributes correctly.