Inference-Time Concept Suppression and Video-Centric Evaluation for Text-to-Video Models

2026-07-15

AI Summary

This work addresses the challenge of precisely removing specific concepts from text-to-video generation while preserving non-target content and temporal coherence. The authors propose SIRUS, a training-free inference-stage framework that achieves video-level concept erasure by leveraging textual aliases to identify target concept evidence in prompts and suppressing its expression during sampling. A comprehensive evaluation protocol is introduced, integrating frame-level residual statistics, paired preservation analysis, and VBench quality diagnostics. On CogVideoX, SIRUS attains an average forgetting success rate of 70.4%, a frame hit rate of 25.7%, and incurs only a marginal VBench quality drop of 0.016, significantly outperforming baseline methods. Cross-model generalization is further demonstrated on Wan2.2.

Abstract

Text-to-video (T2V) generators can synthesize realistic and temporally coherent videos, but controllably removing a target concept from a generator remains difficult. Unlike 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. We propose \textbf{SIRUS}, a training-free inference-time framework for concept-level T2V unlearning. Given 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. We 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. Across 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. Transfer experiments on Wan2.2 further suggest that SIRUS generalizes across modern T2V backbones.

Problem

• text-to-video
• concept suppression
• inference-time unlearning
• video generation
• concept erasure

Innovation

• inference-time unlearning
• concept suppression
• text-to-video generation
• video-centric evaluation
• training-free framework