Reproduce “Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Anh Bui, Long Vuong, Khanh Doan, Trung Le, Paul Montague, Tamas Abraham, Dinh Phung
This research from Monash University, VinAI Research, and the Australian Defence Science and Technology Group develops a method for erasing undesirable concepts from text-to-image diffusion models by adaptively identifying and preserving the specific concepts most sensitive to the unlearning process. The approach achieves a superior balance between effective concept erasure and the preservation of the model's ability to generate other, desirable content, outperforming existing sanitization techniques.
Key insights
- Erasing different concepts from diffusion models leads to distinct degradation patterns on other concepts, and generic "neutral" concepts are typically not the most sensitive to these changes.
- Explicitly identifying and preserving concepts most adversely affected by the unlearning process, rather than generic concepts, is crucial for maintaining overall model generation quality.
- The proposed adversarial search for sensitive concepts, enabled by the Gumbel-Softmax trick, allows for adaptive preservation that outperforms methods relying on fixed or textually similar concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.