Research map / Image generation / Text-to-image diffusion
SPEED: Scalable, Precise, and Efficient Concept Erasure for Diffusion Models
Ouxiang Li, Yuan Wang, Xinting Hu, Houcheng Jiang, Tao Liang, Yanbin Hao, Guojun Ma, Fuli Feng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Invariant Equality Constraints (ablation) | SPEED null-space editing · Erase Van Gogh, non-target FID (avg 4 styles) | non-target FID: 50.43 -> 48.17 | Improved |
| add Influence-based Prior Filtering (ablation) | SPEED + IEC · Erase Van Gogh, non-target FID (avg 4 styles) | non-target FID: 48.17 -> 38.02 | Improved |
| add Directed Prior Augmentation (vs random augmentation) (ablation) | SPEED + IEC + IPF · Erase Van Gogh, non-target FID (avg 4 styles) | non-target FID: 32.62 -> 29.35 | Improved |
| add random prior augmentation (ablation) | SPEED + IEC + IPF · Erase Van Gogh, non-target FID (avg 4 styles) | non-target FID: 38.02 -> 32.62 | Improved |
| add random prior augmentation (ablation) | SPEED + IEC + IPF · Erase Van Gogh, MS-COCO | FID: 20.57 -> 20.99 | Hurt |
| null-space constrained editing vs weighted least squares (UCE) | SPEED · Erase 100 celebrities | Acc_r: 20.92 -> 85.54 | Improved |
| use SPEED instead of MACE | concept erasure · Erase 100 celebrities | runtime: 1736 s -> 5.0 s | Improved |
| use SPEED instead of MACE | concept erasure · Erase 100 celebrities | Acc_e: 4.80 -> 5.87 | Hurt |
| increase target concepts from 1 to 3 | SPEED vs UCE · Pikachu non-target | FID: SPEED +4.59 (16.81 -> 21.40) vs UCE +10.14 | Improved |
| use SPEED rather than RECE for target erasure | few-concept erasure · Erase Snoopy | target CS: 23.50 vs 18.38 | Hurt |
Closest papers by meaning
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.93)
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.92)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.92)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.92)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.92)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.92)
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.
