Latent DPO for Concept Erasure in Text-To-Video Diffusion Models

Le résumé fourni par la source

We address concept erasure in text-to-video models, targeting harmful concepts like copyrighted identities, celebrities, while preserving visual quality and temporal coherence. We cast erasure as latent preference learning: a decoding-free adaptation of Direct Preference Optimization (DPO) trains the denoiser, under matched noise and timestep, to prefer concept-free trajectories, updating parameters via LoRA. To stabilize cross-frame behavior, we introduce a spatiotemporal-semantic alignment (STSA) loss based on a key-centric attention context. We also contribute VCE-24K, a training-free paired corpus built from a frozen Wan-2.1 via early noising and CLIP filtering. On VCE-24K, our method markedly lowers concept alignment with near-baseline DOVER and VBench-2 scores. A blinded user study reports the lowest human detection rates with minimal MOS drop. To our knowledge, this is the first concept erasure for the video generation model.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.