Research map / Image generation / Text-to-image diffusion
EraseSAE: Surgical Concept Erasure in Text-to-Video Diffusion Models via Sparse Autoencoders
Xinghao Wang, Dong Li, Wei Yu, Yingwei Pan, Tao Gong, Qi Chu, Nenghai Yu, Ting Yao
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 L_cpt and L_leak to L_ctx (ablation) | PConvSAE training · HunyuanVideo nudity erasure | nudity rate: 24.5 -> 8.09 | Improved |
| add temporal consistency loss L_temp (ablation) | PConvSAE training · HunyuanVideo nudity erasure | SSIM: 47.37 -> 65.69 | Improved |
| add L_temp (ablation) | PConvSAE training · HunyuanVideo nudity erasure | nudity rate: 8.09 -> 7.13 | Improved |
| use convolutional single-branch PConvSAE instead of linear SAE (ablation) | EraseSAE · HunyuanVideo nudity erasure | nudity rate: 12.76 -> 9.06 | Improved |
| use convolutional single-branch PConvSAE instead of linear SAE (ablation) | EraseSAE · HunyuanVideo nudity erasure | SSIM: 56.97 -> 61.37 | Improved |
| add dual-branch context/concept partition (ablation) | PConvSAE · HunyuanVideo nudity erasure | nudity rate: 9.06 -> 7.13 | Improved |
| add dual-branch context/concept partition (ablation) | PConvSAE · HunyuanVideo nudity erasure | SSIM: 61.37 -> 65.69 | Improved |
| use spatially-modulated CFG instead of feature subtraction (ablation) | EraseSAE · HunyuanVideo nudity erasure | nudity rate: 51.30 -> 7.13 | Improved |
| use spatially-modulated CFG instead of zeroing masked activations (ablation) | EraseSAE · HunyuanVideo nudity erasure | SSIM: 53.51 -> 65.69 | Improved |
| use spatially-modulated CFG instead of zeroing masked activations (ablation) | EraseSAE · HunyuanVideo nudity erasure | nudity rate: 6.56 -> 7.13 | Mixed |
| use PConvSAE instead of linear SAE (ablation) | kernel purity · 5 celebrity identities | purity: 0.37 -> 0.76 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 4 of 4 shared results
- EraseAnything — EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025) · better on 1 of 1 shared result
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
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.93)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.92)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.90)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.90)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.89)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 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.
