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

Recent advances in text-to-video (T2V) diffusion models have demonstrated remarkable generative capabilities, yet their reliance on loosely curated training data raises pressing safety and copyright concerns. Concept erasure offers a principled remedy by removing unwanted semantics from pretrained models while preserving remaining concepts. However, existing approaches typically operate at a coarse granularity misaligned with the fine-grained, distributed nature of concept representations, leading to incomplete removal or degraded generation quality. We argue that surgical erasure fundamentally requires intervention at the level of monosemantic features, where each unit encodes a single interpretable concept. To this end, we propose EraseSAE, a novel framework that leverages sparse autoencoders to achieve surgical concept erasure in DiT-based T2V diffusion models via a principled decompose-attribute-erase pipeline. We first introduce the Partitioned Convolutional Sparse Autoencoder, which decomposes dense spatiotemporal activations into disentangled, interpretable sparse features while preserving spatiotemporal coherence. A contrastive attribution mechanism then contrasts activations from paired prompts to isolate concept-specific feature kernels. At inference, timestep-resolved spatiotemporal masks derived from the identified kernels confine erasure to regions where the target concept is active, leaving unrelated content intact. Extensive experiments across diverse diffusion models and concept erasure tasks demonstrate that EraseSAE achieves precise and robust concept removal with minimal quality degradation, substantially outperforming state-of-the-art methods. The code is available at https://github.com/HiDream-ai/EraseSAE.

Text-to-video diffusion models generate temporally coherent content from natural language, yet when a prompt describes an early scene that persists while a new event emerges on top of it---such as "a tall sandcastle standing on a beach where a wave rushes in and washes it away"---generation frequently fails to realize the late-segment event in the corresponding frames. We identify this failure as Temporal Prior Suppression (TPS): the dominant prior of the early segment captures the cross-attention trajectory across the temporal axis and suppresses the guidance signal needed for late-segment realization, a competing tendency existing guidance mechanisms do not model. We introduce Temporal Prior Decoupling (TPD), a training-free framework that restores suppressed late-segment signals during diffusion sampling. TPD constructs a temporal counterfactual by conditioning on the early segment alone, and defines the discrepancy between the full-prompt and counterfactual trajectories as a suppressed signal direction. Rather than removing this direction as in prior subtractive projection methods, TPD restores it through a frame-selective lower-bound constraint resolved jointly over diffusion timestep and video frame, realizing the suppressed event in the late frames without disrupting early-segment coherence: where prior work enforces upper-bound feasibility to remove unwanted semantics, TPD enforces lower-bound feasibility to guarantee suppressed-signal contribution. TPD runs entirely within standard diffusion sampling without retraining, and is defined purely in classifier-free guidance space, making it backbone-agnostic by construction. Experiments show that TPD significantly improves late-concept realization while preserving temporal coherence and visual fidelity, and that the targeted suppression recurs across distinct text-to-video backbones.

Text-to-Video (T2V) models have demonstrated remarkable capability in generating temporally coherent videos from natural language prompts, yet they also risk producing unsafe content such as violence or explicit material. Existing prompt-level defenses are largely inherited from text-to-image safety and operate on the lexical surface of the input, making them vulnerable to jailbreak attacks that disguise harmful intent through rephrasing or adversarial prompting. Moreover, T2V generation introduces a distinctive challenge overlooked by prior work: temporally emergent risk, where a seemingly benign prompt leads to unsafe content through the generator's temporal extrapolation toward narrative coherence. We propose \method{}, a training-free, inference-time defense framework that reformulates T2V safety as a causal intervention in a temporally structured semantic space. TrajShield handles explicit unsafe prompts, jailbreak attacks, and temporally emergent risks in a unified manner by simulating the implied trajectory of a prompt, localizing the causal origin of potential risk, and applying a minimally invasive rewrite that neutralizes the risk while preserving safety-irrelevant semantics. Experiments on T2VSafetyBench across 14 safety categories and multiple T2V backends demonstrate that TrajShield achieves state-of-the-art defenseive performance while maintaining high semantic fidelity, substantially outperforming existing defenses, with an average ASR reduction of 52.44%.