Synapse
⌘+K
Synapse
PulseExploreClubsResearchersJournals
Instagram
HomeClubsExplore
June 11, 2026Scientia Sinica Informationis

ShallowUnlearn: Rethinking concept unlearning through disentangled component-level erasure in text-guided diffusion models

View Full Paper
Ask AI
Bookmark
Share

Authors

MZMengnan ZhaoTZTianhang ZhengBWBo Wang

Discussion

Loading...

Member takes

Overview

Randomized trial investigates component-level erasure in text-guided diffusion models, suggesting improved concept unlearning efficacy.

Key Points

  • The study aims to improve concept unlearning in text-guided diffusion models by addressing ethical concerns related to harmful content generation.
  • Developed a shallow unlearning approach combining a component extraction module and a decoupling exchange strategy.
  • Pre-trained the component extraction module to decompose concept embeddings into distinct components.
  • Fine-tuned model weights to selectively remove critical erasure components while preserving non-essential components.
  • Shallow unlearning effectively balances the erasure of harmful concepts and the retention of model performance.
  • Demonstrated improved efficiency in forgetting harmful concepts with minimal impact on the model's utility.

Cite This Study

Zhao et al. (2026) studied this question.

synapsesocial.com/papers/6a2a52d980c8f91e7f39eb14https://doi.org/10.1360/ssi-2025-0408
View Full Paper
Ask AI
Bookmark
Share

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

  1. 1Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient2024 · 3 citations
  2. 2Concept Unlearning by Modeling Key Steps of Diffusion Process2025
  3. 3ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning2024 · 1 citations
  4. 4Erasing Concepts from Text-to-Image Diffusion Models with Few-shot Unlearning2024 · 2 citations
  5. 5T2VUnlearning: A Concept Erasing Method for Text-to-Video Diffusion Models2025