Research map / Image generation / Fast diffusion sampling
Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching 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 Differential Vector Erasure (DVE), a training-free method that removes concepts from flow matching T2I models (FLUX) by projecting velocity fields onto a concept-anchor differential vector and correcting only when the alignment falls below a threshold.
Problem
Concept erasure for flow matching text-to-image models, where prior methods are designed for DDPM diffusion or require costly fine-tuning, leaving NSFW content, objects and artistic styles generable.
Approach
Builds a differential vector field between anchor-concept and erasure-concept velocities and applies a thresholded projection-based correction at inference, so only trajectories aligned with the erasure concept are pushed toward the anchor; no weights are updated.
Key results
- On I2P NSFW erasure, DVE reaches 146 total exposed body parts versus 605 for FLUX.1-dev and 199 for EraseAnything, with FID 21.70 (best).
- On adversarial NSFW attacks, DVE averages 4.0% ASR versus 18.1% for FLUX.1-dev and 9.6% for EraseAnything.
- On 12-object ImageNet erasure, UA drops to 3.3% from 88.3% for FLUX.1-dev while IRA is 86.3%.
Surprising findings
- Preprocessed differential vectors cut latency to 5.49s (near baseline) but degrade FID to 25.27 and raise I2P exposures to 226.
- Without the projection-based selective correction (psc-off), CLIP score drops to 24.83 and FID rises to 41.54, indicating over-erasure.
Limitations
- Full DVE roughly triples inference latency (16.37s vs 5.45s) because of extra velocity evaluations per step.
- Preprocessed differential vectors trade image quality (FID 25.27) and erasure (226 exposures) for speed.
- The semantic subspace confinement argument rests on an assumption that the differential vector lies in the Jacobian range, which is not empirically verified.
Experimental setup
Datasets: I2P MSCOCO ImageNet Concept-prune
Benchmarks: I2P Ring-A-Bell MMA-Diffusion Prompt4Debugging UnlearnDiff MSCOCO ImageNet object classes UnlearnDiffAtk style classifier
Metrics: NudeNet exposed body part count ASR UA IRA Top-1/3/5 accuracy FID CLIP score latency
Baselines: ESD UCE CA (model-based) CA (noise-based) MACE EAP Meta-Unlearning EraseAnything
Builds on: flow matching rectified flow classifier-free guidance FlowEdit
Closest papers by meaning
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.91)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.91)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.90)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.90)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · 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.