Research map / Image generation / Text-to-image diffusion
Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters
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 AdaVD, a training-free concept erasure method that projects value vectors in diffusion cross-attention layers onto the orthogonal complement of target-concept values, with an adaptive cosine-based shift to better preserve non-target content.
Problem
Erasing unwanted concepts (copyrighted, offensive, unsafe) from text-to-image diffusion models precisely and cheaply, balancing erasure efficacy against preservation of non-target content (prior preservation).
Approach
Performs orthogonal complement projection of prompt value vectors against target-concept value vectors inside each cross-attention layer, with a token-wise sigmoid shift on cosine similarity so that weakly related tokens are erased less, supporting single and multi-concept erasure without any fine-tuning.
Key results
- On instance erasure of Snoopy, AdaVD reports CS 20.28 and FID 5.72 for Snoopy-erasure-protected concept, with FID for non-target concepts like Dog 2.32 versus 19.61 for the next best (SPM), as reported in Table 1.
- Authors claim 2 to 10 times improvement in prior preservation (FID) over the second best method across single and multi-concept erasure.
- Erasing 10 concepts and generating 10 images takes 22 s on one A40 GPU for AdaVD versus 14 s for SLD and over 10000 s for ConAbl.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AdaVD for single-concept Snoopy erasure instead of the strongest baseline | SD v1.4 · Snoopy | CLIP score: 20.78 -> 20.28 | Helped |
| Use AdaVD to preserve non-target concepts during Snoopy erasure instead of SPM | SD v1.4 · Mickey (non-target; erase Snoopy) | FID: 33.06 -> 5.72 | Helped |
| Erase Snoopy and Mickey together with AdaVD instead of MACE | SD v1.4 · Snoopy | CLIP score: 20.74 -> 20.29 | Helped |
| Use AdaVD to preserve non-target concepts during joint Snoopy and Mickey erasure instead of SPM | SD v1.4 · Spongebob (non-target; erase Snoopy and Mickey) | FID: 41.58 -> 9.34 | Helped |
| Erase Snoopy, Mickey, and Spongebob together with AdaVD instead of MACE | SD v1.4 · Snoopy | CLIP score: 20.48 -> 19.39 | Helped |
| Use AdaVD to preserve non-target concepts during Van Gogh erasure instead of SPM | SD v1.4 · Picasso (non-target; erase Van Gogh) | FID: 62.25 -> 6.82 | Helped |
| Use AdaVD instead of SLD for 10-concept erasure and generation of 10 images | SD v1.4 · 10-concept erasure and generation of 10 images | total time (s): 14 -> 22 | Hurt |
| Apply AdaVD rather than competing methods for prior preservation | AdaVD · Various non-target concepts | FID: 2 to 10 times | Helped |
Limitations
- Theoretical understanding of the accumulated effect of layer-wise orthogonal projection is left open
- Erasing relies on a fixed set of hyperparameters (threshold, scale, rate) whose choice is analyzed only empirically
- Evaluation on prior preservation relies on FID and CLIP score rather than human judgment
Experimental setup
Datasets: I2P benchmark SPM evaluation templates (80 instance, 30 art style, 25 celebrity)
Benchmarks: SPM evaluation templates I2P
Metrics: CLIP score FID time consumption
Baselines: ConAbl ESD SPM MACE NP SLD SuppressEOT SAFREE
Builds on: Safe Latent Diffusion Negative Prompt classifier-free guidance latent diffusion model
Closest papers by meaning
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.92)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.91)
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.91)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.91)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.90)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 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.