{"slug": "inference-time-concept-suppression-and-video-centric-evaluation-for-text-to", "title": "Inference-Time Concept Suppression and Video-Centric Evaluation for Text-to-Video Models", "summary": "Researchers propose SIRUS, a training-free inference-time framework for concept-level unlearning in text-to-video generators, which suppresses target concepts across frames without updating model parameters. On CogVideoX, SIRUS achieves 70.4% average forgetting success and reduces quality drop to -0.016 VBench, outperforming baselines like VideoEraser. The method also introduces a video-oriented evaluation framework for measuring forgetting, preservation, quality, robustness, and efficiency.", "body_md": "arXiv:2607.14194v1 Announce Type: new\nAbstract: Text-to-video (T2V) generators can synthesize realistic and temporally coherent videos, but controllably removing a target concept from a generator remains difficult.\nUnlike text-to-image concept erasure, T2V unlearning must suppress a target concept that may persist across frames while preserving non-target subjects, actions, scenes, and temporal structure.\nWe propose \\textbf{SIRUS}, a training-free inference-time framework for concept-level T2V unlearning.\nGiven textual aliases of a target concept, SIRUS localizes target-related prompt evidence and suppresses target expression during sampling, without updating the text encoder or denoising network.\nWe further introduce a video-oriented evaluation framework for T2V unlearning that separately measures target forgetting, non-target preservation, video quality, jailbreak robustness, and efficiency, using video-level failure criteria, frame-level residue statistics, paired preservation analysis, VBench-based quality diagnostics, and deployment overhead measurement.\nAcross five safety, object, and style concepts on CogVideoX, SIRUS reaches 70.4\\% average forgetting success and 25.7\\% average frame hit, compared with 44.4\\% / 47.2\\% for VideoEraser, while reducing the average VBench quality drop from -0.043 to -0.016, yielding the strongest forgetting-quality trade-off among fully evaluated baselines.\nTransfer experiments on Wan2.2 further suggest that SIRUS generalizes across modern T2V backbones.", "url": "https://wpnews.pro/news/inference-time-concept-suppression-and-video-centric-evaluation-for-text-to", "canonical_source": "https://arxiv.org/abs/2607.14194", "published_at": "2026-07-17 04:00:00+00:00", "updated_at": "2026-07-17 04:07:43.481145+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "generative-ai", "ai-safety", "computer-vision"], "entities": ["SIRUS", "CogVideoX", "VideoEraser", "Wan2.2", "VBench"], "alternates": {"html": "https://wpnews.pro/news/inference-time-concept-suppression-and-video-centric-evaluation-for-text-to", "markdown": "https://wpnews.pro/news/inference-time-concept-suppression-and-video-centric-evaluation-for-text-to.md", "text": "https://wpnews.pro/news/inference-time-concept-suppression-and-video-centric-evaluation-for-text-to.txt", "jsonld": "https://wpnews.pro/news/inference-time-concept-suppression-and-video-centric-evaluation-for-text-to.jsonld"}}