Related Experiment Video
Updated: Aug 6, 2026
08:01
Fabrication of Zero Mode Waveguides for High Concentration Single Molecule Microscopy
Published on: May 12, 2020
Safe Image Generation via Lightweight Concept Erasure in Diffusion Models
Summary
This study introduces the Singular Value Eraser (SVEraser), a novel method for safe image generation. SVEraser precisely removes sensitive concepts from diffusion models without affecting general image quality or unrelated content.
Area of Science:
- Artificial Intelligence
- Computer Vision
- Machine Learning
Background:
- Text-to-image diffusion models excel at image synthesis but pose safety risks due to potential misuse.
- Existing concept erasure methods struggle with precise control, leading to incomplete or excessive erasure and degradation of non-target concepts.
Purpose of the Study:
- To develop a safe and effective method for concept erasure in diffusion models.
- To selectively suppress sensitive concepts while maintaining the model's overall generative capabilities.
Main Methods:
- Proposed the Singular Value Eraser (SVEraser), a lightweight module optimizing singular-value offsets of weight matrices.
- Implemented an eraser activation mechanism for adaptive SVEraser selection during inference.
- Enabled flexible combination of multiple SVErasers for multi-concept erasure.
Main Results:
- SVEraser achieved precise concept removal with minimal side effects on unrelated content.
- Demonstrated accurate suppression of copyrighted objects, artistic styles, and explicit content.
- Preserved non-target semantics effectively, even in multi-concept scenarios.
Conclusions:
- SVEraser offers a practical and reliable solution for safe diffusion-based image generation.
- The method effectively balances concept suppression with the preservation of general generative capabilities.