Research map / Image generation / Text-to-image diffusion
Where Concept Erasure Should Occur: Concept–Layer Alignment in Text-to-Video Diffusion Models
Yiwei Xie, Ping Liu, Zheng Zhang
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 separability contrastive loss Lcon to layer search (ablation) | CLEAR · Parachute, Wan2.2-5B | generative rate: 27.0% -> 18.4% | Improved |
| add separability contrastive loss Lcon to layer search (ablation) | CLEAR · Springer dog, Wan2.2-5B | generative rate: 14.0% -> 4.8% | Improved |
| add separability contrastive loss Lcon to layer search (ablation) | CLEAR · Nudity, Wan2.2-5B | generative rate: 9.4% -> 10.9% | Hurt |
| intervene at top-1 only instead of top-1 and top-2 layers (ablation) | CLEAR · Nudity, Wan2.2-5B | generative rate: 30.1% -> 11.1% | Improved |
| increase intervention strength gamma from 8 to 12 (ablation) | CLEAR · Nudity, Wan2.2-5B | generative rate: 28.4% -> 3.2% | Improved |
| increase intervention strength gamma from 10 to 12 (ablation) | CLEAR · Nudity, Wan2.2-5B | consistency: 0.1754 -> 0.0933 | Hurt |
| use sparse autoencoder features instead of dense steering vector (ablation) | CLEAR · Nudity, Wan2.2-5B | generative rate: 24.5% -> 11.1% | Improved |
| learned differentiable layer search instead of brute-force search (ablation) | CLEAR · Springer dog, Parachute, Nudity | search cost: 24x -> 1x | Improved |
| use CLEAR instead of fine-tuning (T2VUnlearning) on small model | CLEAR · Object erasure, CogVideoX-2B | imaging quality: 0.3907 -> 0.4683 | Improved |
| use CLEAR instead of T2VUnlearning | CLEAR · Object erasure, Wan2.2-5B | generative rate: 24.5% -> 12.8% | Improved |
| use CLEAR instead of VideoEraser | CLEAR · Celebrity erasure, CogVideoX-2B | identity similarity: 0.1140 -> 0.1802 | Hurt |
| apply CLEAR to evasive Ring-A-Bell prompts | CLEAR · Ring-A-Bell, Wan2.2-5B | generative rate: 69.6% -> 25.6% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.91)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.91)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.90)
- Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.90)
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.89)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.89)
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.
