Research map / Image generation / Text-to-image diffusion
Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient
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 DoCo, which unlearns concepts from text-to-image diffusion models by adversarially aligning the output domains of target and anchor concepts and uses a gradient-surgery step to preserve non-target concepts.
Problem
Concept unlearning in text-to-image diffusion models fails to generalize beyond the exact unlearned prompts and often degrades the model's utility on related concepts.
Approach
Treat unlearning as distribution-level alignment: a PatchGAN discriminator is trained adversarially on latent denoised outputs conditioned on target vs. anchor concepts, and the unlearning gradient is projected to remove components conflicting with a retention (diffusion) loss gradient.
Key results
- On erasing Snoopy, DoCo reaches CS 49.10 and CA 35.77 (lower is better for erased concept) versus ESD 50.77/38.62 and SPM 55.15/40.25, while retaining Mickey (CS 71.84) and Spongebob (CS 71.18) better than the baselines' trade-off.
- Ablation on Snoopy: without retain loss CS 52.27 with FID 146.53; adding L2 raises FID-based retention but drops unlearning (CA 60.25); gradient surgery (CP) gives CS 49.10, CA 35.77, FID 204.37 with better retention of Mickey and Spongebob.
- Celebrity unlearning (ISM): DoCo erases Cristiano Ronaldo to 0.89 ISM and keeps LeBron James-related retention at 0.69 and Robert Downey Jr. at 0.62, competitive with baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DoCo instead of ESD | Instance unlearning · Snoopy unlearning | CLIP Score: 50.77 -> 49.10 | Helped |
| Use DoCo instead of ESD | Instance unlearning · Snoopy unlearning | FID: 150.22 -> 204.37 | Helped |
| Use DoCo instead of SPM | Style unlearning · Van Gogh unlearning | CLIP Score: 51.80 -> 49.55 | Helped |
| Use DoCo instead of SPM | Style unlearning · Van Gogh unlearning | FID: 198.65 -> 215.76 | Helped |
| Add concept-preserving gradient rather than omit retaining operations (ablation) | DoCo · Snoopy unlearning | FID: 146.53 -> 204.37 | Helped |
| Add concept-preserving gradient instead of L2 retention loss (ablation) | DoCo · Snoopy unlearning | CLIP Score: 56.68 -> 49.10 | Helped |
| Add concept-preserving gradient instead of L2 retention loss (ablation) | DoCo · Snoopy unlearning | FID: 187.30 -> 204.37 | Helped |
Limitations
- Retention improvements come at the cost of some unlearning strength in certain ablations (L2 loss reduces unlearning).
- Hyperparameter lambda for gradient surgery set to default 1 without a sensitivity analysis shown in the text.
Experimental setup
Metrics: CLIP Score CLIP Accuracy FID ISM
Baselines: ESD ConAbl SPM
Builds on: ESD ConAbl SPM PCGrad PatchGAN
Closest papers by meaning
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.92)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.92)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.91)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.91)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.91)
- Concept Replacer: Replacing Sensitive Concepts in Diffusion Models via Precision Localization (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.