Research map / Video / Video generation
Video Unlearning via Low-Rank Refusal Vector
Simone Facchiano, Stefano Saravalle, Matteo Migliarini, Edoardo De Matteis, Alessio Sampieri, Andrea Pilzer, Emanuele Rodolà, Indro Spinelli, Luca Franco, Fabio Galasso
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 PCA refinement to the refusal vector (ablation) | refusal vector · T2VSafetyBench Pornography | censorship rate (%): 18.0% -> 16.9% | Improved |
| replace PCA with cPCA refinement (ablation) | refusal vector + PCA · T2VSafetyBench Pornography | censorship rate (%): 16.9% -> 13.4% | Improved |
| add refusal vector to Open-Sora (ablation) | Open-Sora · T2VSafetyBench Pornography | censorship rate (%): 44.7% -> 18.0% | Improved |
| use highly similar safe/unsafe prompt pairs instead of low similarity (ablation) | refusal vector + cPCA · T2VSafetyBench Pornography | censorship rate (%): 25.3% -> 23.4% | No clear effect |
| resample the neutral prompt set in cPCA (ablation) | refusal vector + cPCA · T2VSafetyBench Pornography | censorship rate (%): 17.5% to 19.3% | No clear effect |
| increase cPCA rank to k=100 (ablation) | cPCA subspace · T2VSafetyBench | censorship rate (%) | Mixed |
| increase suppression coefficient lambda (ablation) | refusal vector · T2VSafetyBench | censorship rate (%) | Mixed |
| compose five refusal vectors instead of one (ablation) | refusal vectors · T2VSafetyBench Pornography (100 prompts) | censorship rate (%): 26.0 -> 16.0 | Improved |
| compose five refusal vectors instead of one (ablation) | refusal vectors · T2VSafetyBench Sequential Action Risk (100 prompts) | FVD: 273.40 -> 579.37 | Hurt |
| use refusal vector + cPCA instead of NullSCE fine-tuning | Open-Sora · T2VSafetyBench average | censorship rate (%): 25.5% -> 12.6% | Improved |
| use refusal vector + cPCA instead of SAFREE filtering | ZeroScopeT2V · SafeSora average | censorship rate (%): 43.1% -> 9.9% | Improved |
| use refusal vector + cPCA instead of adapted UCE | ZeroScopeT2V · SafeSora Sexual | FVD: 103.3 -> 63.51 | Improved |
Closest papers by meaning
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.88)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.87)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.86)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.86)
- Direct Unlearning Optimization for Robust and Safe Text-to-Image Models (NeurIPS 2024 · similarity 0.86)
- Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models (NeurIPS 2026 · similarity 0.86)
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.
