CleanVideo: Adaptive Concept Erasure for Text-to-Video Diffusion Models
Organizations: Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) · University of Electronic Science and Technology of China (UESTC) · The Hong Kong University of Science and Technology (HKUST)
Abstract
Concept erasure aims to selectively eliminate undesired visual semantics from pre-trained generative models without compromising their general utility. Extending concept erasure from images to video is nontrivial. Target concepts emerge gradually and vary across frames and denoising steps. As a result, fixed interventions may miss the target or introduce blurring, jitter, and content distortion. We propose CleanVideo, a selective erasure framework that performs low-dimensional subspace intervention controlled by a tri-modal gating mechanism. By jointly processing spatiotemporal visual features, timestep signals, and textual semantics, CleanVideo determines where, when, and whether to intervene, steering erased content toward natural surrogate concepts when such surrogates can be clearly defined while preserving non-target content. Experiments on three video diffusion models show that CleanVideo effectively erases target concepts while maintaining visual fidelity and temporal coherence, outperforming existing baselines under frame-level and video-level evaluations and under concept-recovery attacks when the protected pipeline remains intact.
Figures & tables
| Method | SafeSora | Ring-A-Bell | ||||
| CogX-2B | CogX-5B | Hunyuan | CogX-2B | CogX-5B | Hunyuan | |
| Vanilla | 20.3 | 20.4 | 26.8 | 32.7 | 39.9 | 61.6 |
| NegPrompt | 22.7 | 19.6 | 24.6 | 29.7 | 39.7 | 58.4 |
| SAFREE | 17.1 | 6.8 | 15.2 | 14.8 | 15.8 | 55.3 |
| T2VUnlearning | 7.5 | 3.6 | 17.8 | 5.8 | 9.2 | 37.3 |
| VideoEraser | 10.8 | 7.7 | 14.4 | 9.6 | 13.9 | 34.7 |
| Metric | Vanilla | CleanVideo | Metric | Vanilla | CleanVideo |
| Subject Cons. | 0.8961 | 0.8872 | Temporal Flicker | 0.9817 | 0.9816 |
| Background Cons. | 0.9032 | 0.8952 | Motion Smooth. | 0.8724 | 0.8682 |
| Aesthetic Qual. | 0.5750 | 0.5401 | Dynamic Degree | 1.0000 | 1.0000 |
| Imaging Qual. | 0.6121 | 0.5782 | Object Class | 0.8877 | 0.8794 |
| Appearance Style | 0.2410 | 0.2439 | Multiple Objects | 0.4961 | 0.4792 |
| Scene | 0.3867 | 0.3594 | Color | 0.9078 | 0.9148 |
| Method | Hunyuan | CogX-2B | CogX-5B | Avg. | ||||
| Acc | Acc | Acc | Acc | Acc | Acc | Acc | Acc | |
| Vanilla | 100.0 | 85.3 | 96.8 | 79.0 | 99.0 | 83.1 | 98.6 | 82.5 |
| NegPrompt | 82.6 | 64.2 | 85.5 | 55.9 | 86.0 | 61.3 | 84.7 | 60.5 |
| SAFREE | 64.0 | 56.7 | 68.8 | 48.8 | 66.9 | 54.7 | 66.6 | 53.4 |
| VideoEraser | 37.7 | 73.9 | 46.7 | 65.6 | 46.8 | 71.1 | 43.7 | 70.2 |
| T2VUnlearning | 30.7 | 78.2 | 42.6 | 69.7 | 40.1 | 75.5 | 37.8 | 74.5 |
| Variant | Nudity | Obj. Class | Subj. Cons. | Motion | Bg. Cons. |
| w/o Tri-Modal Gate | 14.2 | 95.8 | 96.1 | 95.2 | 97.5 |
| w/o Preservation Loss | 10.5 | 88.5 | 95.8 | 95.1 | 90.5 |
| w/o Diversity Loss | 11.5 | 95.2 | 96.0 | 95.5 | 97.0 |
| CleanVideo | 7.5 | 96.2 | 96.9 | 95.7 | 97.8 |
Appendix figures & tables32 assets
Supplementary material from the paper’s appendix.
Appendix
| Metric | CogX-2B | CogX-5B | Hunyuan | |||
| Vanilla | CleanVideo | Vanilla | CleanVideo | Vanilla | CleanVideo | |
| Subject Consistency | 0.8961 | 0.8872 | 0.9310 | 0.9280 | 0.9690 | 0.9670 |
| Background Consistency | 0.9032 | 0.8952 | 0.9380 | 0.9350 | 0.9800 | 0.9780 |
| Temporal Flickering | 0.9817 | 0.9816 | 0.9850 | 0.9860 | 0.9880 | 0.9890 |
| Motion Smoothness | 0.8724 | 0.8682 | 0.9140 | 0.9110 | 0.9570 | 0.9550 |
| Dynamic Degree | 1.0000 | 1.0000 | 1.0000 | 0.9535 | 1.0000 | 1.0000 |
| CogVideoX-2B | CogVideoX-5B | HunyuanVideo | |||||||||||
| Task | Method | Obj. | Subj. | Mot. | Bg. | Obj. | Subj. | Mot. | Bg. | Obj. | Subj. | Mot. | Bg. |
| Vanilla | 88.5 | 89.2 | 87.4 | 90.1 | 92.4 | 93.5 | 91.8 | 94.2 | 96.8 | 97.5 | 96.2 | 98.4 | |
| NegPrompt | 88.1 | 88.8 | 87.1 | 89.8 | 91.9 | 93.1 | 91.4 | 93.8 | 96.4 | 97.1 | 95.8 | 98.0 | |
| SAFREE | 82.4 | 80.5 | 83.2 | 85.6 | 87.5 | 85.2 | 88.4 | 89.5 | 92.1 | 90.4 | 93.5 | 94.8 | |
| T2VUnlearning | 79.8 | 78.2 | 81.5 | 82.3 | 85.2 | 82.6 | 86.1 | 87.4 | 90.5 | 88.2 | 91.2 | 92.6 | |
| Nudity | VideoEraser | 81.2 | 79.6 | 82.1 | 83.9 | 86.8 | 84.1 | 87.5 | 88.6 | 91.4 | 89.6 | 92.4 | 93.9 |
| Setting | Vanilla | NegPrompt | SAFREE | VideoEraser | T2VUnlearning | CleanVideo |
| Style / Hunyuan | 0.97 | 0.92 | 0.58 | 0.43 | 0.26 | 0.14 |
| Style / CogX-2B | 0.92 | 0.94 | 0.64 | 0.37 | 0.15 | 0.06 |
| Style / CogX-5B | 0.93 | 0.88 | 0.70 | 0.53 | 0.33 | 0.13 |
| Object / Hunyuan | 0.76 | 0.76 | 0.68 | 0.62 | 0.36 | 0.08 |
| Object / CogX-2B | 0.73 | 0.53 | 0.72 | 0.59 | 0.31 | 0.10 |
| Object / CogX-5B | 0.82 | 0.84 | 0.76 | 0.71 | 0.28 | 0.17 |
| Attack | Vanilla | NegPrompt | SAFREE | VideoEraser | T2VUnlearning | CleanVideo |
| UD | 47 | 48 | 38 | 31 | 27 | 22 |
| CCE | 61 | 59 | 47 | 42 | 33 | 28 |
| Vanilla | NegPrompt | T2VUnlearning | VideoEraser | SAFREE | Ours |
| 49 ±0.47 | 52 ±0.58 | 57 ±0.41 | 64 ±0.73 | 47 ±0.29 |
| Object Erasure (Concept Class) | ||||||||||||
| Model | Metric | Method | Cassette Player | Chain Saw | Church | English Springer | French Horn | Garbage Truck | Gas Pump | Golf Ball | Para- chute | Tench |
| CogX-2B | Vanilla | 65.2 | 78.4 | 71.2 | 93.4 | 99.4 | 72.0 | 85.3 | 99.9 | 100.0 | 73.7 | |
| NegPrompt | 45.8 | 52.1 | 21.2 | 37.7 | 11.5 | 38.5 | 28.8 | 42.7 | 52.5 | 35.3 | ||
| SAFREE | 38.4 | 45.6 | 18.5 | 68.2 | 10.4 | 51.5 | 41.1 | 82.3 | 95.2 | 67.9 | ||
| VideoEraser | 25.1 | 30.5 | 4.7 | 5.3 | 0.9 | 3.2 | 4.0 | 2.0 | 11.7 | 2.2 | ||
| T2VUnlearning | 12.4 | 15.2 | 2.2 | 3.4 | 0.6 | 1.9 | 1.5 | 0.9 | 5.4 | 0.9 | ||
| HunyuanVideo | CogVideoX-2B | CogVideoX-5B | ||||||||||||||
| Pablo | Van | Rem- | Andy | Cara- | Pablo | Van | Rem- | Andy | Cara- | Pablo | Van | Rem- | Andy | Cara- | ||
| Picasso | Gogh | brandt | Warhol | vaggio | Picasso | Gogh | brandt | Warhol | vaggio | Picasso | Gogh | brandt | Warhol | vaggio | ||
| Acc | Vanilla | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 98.5 | 100.0 | 100.0 | 85.7 | 100.0 | 100.0 | 100.0 | 100.0 | 95.0 |
| NegPrompt | 92.0 | 78.6 | 85.9 | 88.0 | 68.3 | 97.5 | 73.7 | 89.4 | 96.8 | 70.1 | 85.2 | 86.3 | 94.5 | 90.5 | 73.6 | |
| SAFREE | 76.8 | 45.7 | 74.3 | 78.5 | 44.8 | 86.4 | 66.8 | 77.5 | 73.7 | 39.8 | 77.4 | 64.4 | 72.8 | 65.4 | 54.6 | |
| VideoEraser | 55.4 | 32.1 | 48.2 | 37.4 | 15.2 | 62.1 | 25.3 | 65.4 | 58.2 | 22.4 | 45.3 | 52.8 | 61.2 | 39.4 | 35.2 | |
| Model | Removed Style | Vanilla Style Leak | CleanVideo Style Leak | Vanilla Obj. Rec. | CleanVideo Obj. Rec. | Retention |
| CogX-2B | Picasso | 0.92 | 0.04 | 0.55 | 0.52 | 0.95 |
| CogX-2B | Van Gogh | 0.95 | 0.06 | 0.62 | 0.59 | 0.95 |
| CogX-5B | Picasso | 0.85 | 0.16 | 0.72 | 0.72 | 1.00 |
| CogX-5B | Van Gogh | 0.98 | 0.13 | 0.64 | 0.67 | 1.05 |
| Hunyuan | Picasso | 0.98 | 0.14 | 0.67 | 0.64 | 0.96 |
| Hunyuan | Van Gogh | 0.97 | 0.09 | 0.58 | 0.61 | 1.05 |
| Rank | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| Nudity Rate (%) | 13.5 | 7.7 | 4.2 | 1.7 | 3.6 | 1.9 | 2.8 | 2.6 | 3.4 |
| Variant | Nudity | Obj. Class | Subj. Cons. | Motion | Bg. Cons. |
| w/o | 9.8 | 96.4 | 92.5 | 96.0 | 94.2 |
| w/o | 8.4 | 94.5 | 88.6 | 93.5 | 96.8 |
| w/o | 7.8 | 94.2 | 95.5 | 89.4 | 97.2 |
| Full (CleanVideo) | 7.5 | 96.2 | 96.9 | 95.7 | 97.8 |
| Task | Variant | Acc | Acc |
| Style (Van Gogh) | Full | 25.5 | 73.2 |
| Style (Van Gogh) | w/o | 36.5 | 66.2 |
| Style (Van Gogh) | w/o | 32.4 | 70.8 |
| Style (Van Gogh) | w/o | 35.6 | 72.0 |
| Object (Chainsaw) | Full | 6.8 | 71.0 |
| Object (Chainsaw) | w/o | 20.8 | 68.4 |
| Setting | Definition | Vanilla | NegPrompt | SAFREE | VideoEraser | T2VUnlearning | CleanVideo |
| Blood / 24 prompts | Visible blood, excluding gore or action-level violence | 0.84 | 0.79 | 0.67 | 0.43 | 0.56 | 0.14 |
| Gore / 12 prompts | Narrow static gore, excluding action-level violence | 0.91 | 0.86 | 0.77 | 0.70 | 0.48 | 0.35 |
| Strategy | Nudity | Van Gogh Style | Object (Chainsaw) |
| Lexical | naked, nude, bare, unclothed | Van Gogh, Starry Night style | chainsaw, power saw, logging saw |
| Scene | beach, shower, bedroom, pool | portrait, landscape, cityscape | forest, workshop, horror scene |
| Implicit | “in their birthday suit” | “swirling brushstrokes” | “loud cutting machine” |
| Concept | Original Prompt | Surrogate Prompt |
| Nudity | A nude woman swimming gracefully in an outdoor pool at sunset. | A woman in a swimsuit swimming gracefully in an outdoor pool at sunset. |
| A man in his birthday suit stepping out of a steamy shower. | A man wrapped in a towel stepping out of a steamy shower. | |
| Van Gogh | A wheat field under a turbulent sky in Van Gogh’s style. | A wheat field under a turbulent sky, photorealistic. |
| Portrait with swirling brushstrokes and vibrant yellows. | Portrait with natural lighting, realistic style. | |
| Chainsaw | A man cutting firewood with a chainsaw in a snowy backyard. | A man standing near firewood in a snowy backyard. |
| Close-up of a chainsaw blade spinning rapidly. | Close-up of wood grain texture. |
| Strategy | Example Surrogate | |
| Deletion | “A man in the forest.” | 6.8 |
| Replacement | “A man holding an axe in the forest.” | 19.1 |
| Negation | “A man in the forest without chainsaw.” | 28.3 |
Concept Erasure
Concept erasure removes specified knowledge or attributes from models while preserving their ability to perform unrelated tasks.
See all 70 papers