SIRUS suppresses unwanted concepts in text-to-video AI with 70.4% success
Training-free framework outperforms existing methods while preserving video quality.
Researchers Wenxuan Chen and Wenjie Feng have proposed SIRUS, a novel training-free inference-time framework for concept-level unlearning in text-to-video (T2V) models. Unlike prior methods that require retraining or fine-tuning, SIRUS works at inference time by localizing target-related prompt evidence and suppressing the expression of unwanted concepts across frames—while preserving non-target subjects, actions, scenes, and temporal coherence. The approach addresses a key challenge in T2V unlearning: removing a concept that persists across multiple frames without degrading overall video quality.
Tested on the CogVideoX backbone, SIRUS achieves 70.4% average forgetting success and a frame hit rate of 25.7%, significantly outperforming VideoEraser’s 44.4% and 47.2% respectively. It also minimizes quality degradation, with a VBench quality drop of only -0.016 compared to -0.043 for VideoEraser. The team also introduces a comprehensive video-centric evaluation framework that measures target forgetting, non-target preservation, video quality, jailbreak robustness, and efficiency. Transfer experiments on Wan2.2 confirm SIRUS generalizes across modern T2V architectures, making it a practical solution for safer, more controllable AI video generation.
- SIRUS achieves 70.4% average forgetting success on CogVideoX, vs 44.4% for VideoEraser
- Minimal VBench quality drop of -0.016, compared to -0.043 for the baseline
- Training-free and generalizes to other backbones like Wan2.2 without retraining
Why It Matters
Enables precise removal of unsafe or unwanted concepts from AI-generated videos without sacrificing quality or retraining.