HyperErase: Scale-Calibrated Hypernetwork for Multi-Concept Erasure in Text-to-Image Models
HyperErase synthesizes prompt-specific LoRAs to erase multiple concepts from text-to-image models.
HyperErase treats concept erasure as prompt-conditioned parameter amortization, training a hypernetwork to map text to prompt-specific LoRA updates. A decoupled rectification separates LoRA tokens into pattern and scale subspaces, applies a square-root transform to limit over-scaling, and uses teacher-derived priors at inference. Across major concept categories, erasure effectiveness, image quality, and semantic alignment match gold-standard single-concept baselines. Each prompt variation receives a specialized LoRA in one forward pass without inference-time gradient updates.
- A hypernetwork emits prompt-specific LoRA updates in one forward pass.
- No per-prompt gradients or manual LoRA merging are required.
- Scale rectification uses a square-root transform and teacher priors.
- Erasure quality matches single-concept baselines across major categories.
From the original article
1 min read at arxiv.orgRecent advances in text-to-image (T2I) generation have substantially improved visual synthesis, but have also raised increasing safety concerns due to their potential to generate harmful or undesirable content.
By Yi Sun · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.