ConceptACP erases diffusion concepts with sparse autoencoders
Original title: Concept Acupuncturist: An Efficient and Precise Concept Erasure Method for Diffusion Models via Sparse Autoencoders
On SDXL-Turbo one-step generation, forgetting success rises 36.34 points over the best baseline.
Concept Acupuncturist maps semantic concepts to U-Net cross-attention blocks and fine-tunes only the most relevant block instead of the full diffusion trajectory. Sparse autoencoder features supply correction signals for concept removal without extra inference-time modules. On SDXL-Turbo the authors report a 36.34 percentage-point gain in forgetting success for one-step generation, with image quality and prompt fidelity preserved, and generalization to other Stable Diffusion backbones.
Unconfirmed: The forgetting-rate gain is the authors' average over their baselines.
- Optimization targets the most concept-relevant cross-attention block.
- SAE activations are projected back as supervision for that block.
- No external inference-time module is required after fine-tuning.
- Code is released in the ConceptACP GitHub repository.
- AI models
- SDXL-TurboStable Diffusion
From the original article
1 min read at arxiv.orgDiffusion-based image generation models produce high-quality images but may generate harmful or policy-violating content, motivating machine unlearning (MU) to selectively remove target concepts while preserving general capabilities.
By Yiran Xiahou · 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.