Research map / Image generation / Fast diffusion sampling
Erasing Undesirable Influence 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 EraseDiff, a first-order constrained-optimization (value-function) unlearning method that erases forget-set influence from diffusion models while preserving utility on retained data.
Problem
Removing undesirable influence (forgotten classes or NSFW concepts) from diffusion models without full retraining, while balancing erasure efficacy against preservation of generation quality.
Approach
Reformulates diffusion unlearning as a single-constrained problem via the value function (forgetting loss minus its inner-optimized minimum as constraint), yielding a closed-form first-order update that minimizes retain loss while keeping forgetting loss decreasing, avoiding nested bi-level optimization.
Key results
- On CIFAR-10 DDPM forgetting 'airplane', EraseDiff_rl reaches FID 7.61 (vs 8.19 for SA) with forgotten-class probability 0.07 vs SalUn 0.06, and 11x faster than SA (12.70 vs 140 min).
- On nudity erasure with SD v1.4, EraseDiff achieves FID 17.01 and CLIP 30.58 versus SalUn FID 25.06 and CLIP 28.91 on 30K COCO images.
- On Imagenette class forgetting, EraseDiff averages FID 1.11 and UA 99.91% over ten classes, slightly better than SalUn (1.22, 99.82%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EraseDiff instead of SalUn for class-wise forgetting | Stable Diffusion on Imagenette · Imagenette (10-class average) | FID: 1.22 -> 1.11 | Helped |
| Use EraseDiff instead of SalUn for class-wise forgetting | Stable Diffusion on Imagenette · Imagenette (10-class average) | Unlearning accuracy (UA): 99.82 -> 99.91 | Helped |
| Use EraseDiff instead of SalUn for nudity erasure | Stable Diffusion v1.4 · I2P nudity prompts | FID: 25.06 -> 17.01 | Helped |
| Use EraseDiff instead of SalUn | DDPM on CIFAR-10 · CIFAR-10 (forget airplane) | FID: 9.16 -> 8.66 | Helped |
| Use EraseDiff instead of SA | DDPM on CIFAR-10 · CIFAR-10 (forget airplane) | P(y = forgotten class | generated image): 0.06 -> 0.24 | Hurt |
| Replace the forgetting target with uniform noise instead of random-label predictions (ablation) | EraseDiff on DDPM · CIFAR-10 (forget airplane) | FID: 8.66 -> 7.61 | Helped |
| Replace the forgetting target with uniform noise instead of random-label predictions (ablation) | EraseDiff on DDPM · CIFAR-10 (forget airplane) | Recall: 0.77 -> 0.72 | Hurt |
| Increase the step size ς (ablation) | EraseDiff · CIFAR-10 | FID | Hurt |
| Increase the inner iteration count K (ablation) | EraseDiff · CIFAR-10 | FID | Hurt |
| Use EraseDiff instead of SA | DDPM on CIFAR-10 · CIFAR-10 (forget airplane) | Time (min.): 140.00 -> 12.70 | Helped |
| Use EraseDiff instead of SalUn | DDPM on CIFAR-10 · CIFAR-10 (forget airplane) | Time (min.): 28.17 -> 12.70 | Helped |
Surprising findings
- SalUn achieves stronger forgetting than EraseDiff on CIFAR-10 and nudity erasure, so the method wins mainly on utility rather than erasure strength.
Limitations
- Erasure can be incomplete: generated images conditioned on the forgotten class may retain similar characteristics, depending on the replacement label choice.
- Trails SalUn in erasure strength on several metrics; the authors note SalUn is a very potent baseline.
- Scrubbed models may be misused for censorship or to falsely assure users that NSFW content is removed.
Experimental setup
Datasets: CIFAR-10 Imagenette I2P
Benchmarks: CIFAR-10 Imagenette I2P COCO
Metrics: FID CLIP score classification probability of forgotten class precision recall UA accuracy NudeNet exposed-part counts runtime and memory
Baselines: Unscrubbed Finetune NegGrad BlindSpot ESD FMN Selective Amnesia SalUn
Builds on: SalUn Selective Amnesia ESD multi-objective optimization DDPM
Closest papers by meaning
- Hiding Images in Diffusion Models by Editing Learned Score Functions (CVPR 2025 · similarity 0.89)
- EVODiff: Entropy-aware Variance Optimized Diffusion Inference (NeurIPS 2025 · similarity 0.89)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.89)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)
- Training-Free Safe Denoisers for Safe Use of Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Is Noise Conditioning Necessary for Denoising Generative Models? (ICML 2025 · similarity 0.88)
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.