Research map / Image generation / Text-to-image diffusion
CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CURE, a training-free, closed-form concept unlearning method for text-to-image diffusion models that edits cross-attention key/value weights via an SVD-based orthogonal projection with a spectral expansion parameter.
Problem
Removing unsafe, copyrighted, or privacy-violating concepts from text-to-image diffusion models while remaining efficient, robust to red-teaming attacks, and avoiding collateral damage to unrelated concepts.
Approach
Builds forget and retain subspaces via SVD of concept token embeddings, applies an energy-based spectral expansion function f(r;α)=αr/((α-1)r+1) to weight singular directions, forms an orthogonal unlearning projector I - (P_f - P_f P_r), and folds it into cross-attention W_k and W_v in a single closed-form edit.
Key results
- On removing Van Gogh style, CURE achieves Acc_e 0.30 (lower is better) versus 0.35 for SAFREE and 0.95 for UCE, with LPIPS_e 0.44.
- On I2P nudity removal, CURE yields a total of 11 nude body parts versus 66 for RECE and 646 for SD v1.4.
- CURE reports under 2 seconds for erasure, compared with ~4500s and ~500s reported for two training-based baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CURE for Van Gogh style removal | Stable Diffusion v1.4 · Van Gogh removal | LPIPS_e: 0.42 -> 0.44 | Helped |
| apply CURE for artist style removal | Stable Diffusion v1.4 · COCO-30k | FID: 1.81 -> 1.44 | Helped |
| apply CURE for nudity removal | Stable Diffusion v1.4 · I2P | nude body parts (Total): 59 -> 11 | Helped |
| apply CURE against adversarial prompts | Stable Diffusion v1.4 · Ring-A-Bell red-teaming | ASR: 0.076 -> 0.013 | Helped |
| apply CURE against adversarial prompts | Stable Diffusion v1.4 · UnlearnDiffAtk red-teaming | ASR: 0.176 -> 0.281 | Hurt |
| apply CURE for object removal | Stable Diffusion v1.4 · Imagenette object removal | Accuracy of Erased Class (Average): 0.3 -> 1.3 | Hurt |
| set alpha from 1 to 2 (ablation) | CURE · Cat and feline removal | Acc_e: 0.47 -> 0.08 | Helped |
| set alpha from 1 to 2 (ablation) | CURE · Cat and feline removal | LPIPS_u: 0.17 -> 0.19 | Hurt |
| set alpha from 2 to 1000 (ablation) | CURE · Cat and feline removal | Acc_e: 0.08 -> 0.00 | Helped |
| set alpha from 2 to 1000 (ablation) | CURE · Cat and feline removal | LPIPS_u: 0.19 -> 0.31 | Hurt |
| set alpha from 2 to infinity (ablation) | CURE · Cat and feline removal | Acc_u: 0.94 -> 0.52 | Hurt |
Surprising findings
- Adding more erased concepts (up to 50 sequential erasures) causes outputs to differ by LPIPS while CLIP scores remain stable.
- On the Ring-A-Bell/MMA-Diffusion/UnlearnDiffAtk red-teaming set, CURE is not best on UnlearnDiffAtk (ASR 0.281 vs 0.176 for MACE).
Limitations
- Highly adversarial prompts may still trigger partial leakage of erased concepts.
- The spectral expansion parameter alpha adds a tuning knob; higher alpha degrades unrelated concepts.
- No formal guarantees on global optimality or leakage bounds under prompt perturbations.
Experimental setup
Datasets: I2P COCO-30k Imagenette P4D Ring-A-Bell MMA-Diffusion UnlearnDiffAtk
Benchmarks: I2P COCO-30k P4D Ring-A-Bell MMA-Diffusion UnlearnDiffAtk Imagenette GIPHY Celebrity Detector
Metrics: LPIPS GPT-4o style accuracy NudeNet detections FID CLIP score attack success rate top-1 classification accuracy
Baselines: SLD SAFREE CA ESD RECE UCE SD-NP MACE SA SDID
Builds on: UCE Tikhonov regularization ESD RECE
Closest papers by meaning
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.92)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.91)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.91)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.91)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.90)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.