ZUMA: Training-Free Zero-Shot Unified Multimodal Anomaly Detection.

Abstract

Multimodal anomaly detection (MAD) aims to exploit both texture and spatial attributes to identify deviations from normal patterns in complex scenarios. However, zero-shot (ZS) settings arising from privacy concerns or confidentiality constraints present significant challenges to existing MAD methods. To address this issue, we introduce ZUMA, a training-free, Zero-shot Unified Multimodal Anomaly detection framework that unleashes CLIP's cross-modal potential to perform ZS MAD. To mitigate the domain gap between CLIP's pretraining space and point clouds, we propose cross-domain calibration (CDC), which efficiently bridges the manifold misalignment through source-domain semantic transfer and establishes a hybrid semantic space, enabling a joint embedding of 2D and 3D representations. Subsequently, ZUMA performs dynamic semantic interaction (DSI) to enable structural decoupling of anomaly regions in the high-dimensional embedding space constructed by CDC, where natural languages serve as semantic anchors to help DSI establish discriminative hyperplanes within hybrid modality representations. Within this framework, ZUMA enables plug-and-play detection of 2D, 3D or multimodal anomalies, without training or fine-tuning even for cross-dataset or incomplete-modality scenarios. Additionally, to further investigate the potential of the training-free ZUMA within the training-based paradigm, we develop ZUMA-FT, a fine-tuned variant that achieves notable improvements with minimal parameter trade-off. Extensive experiments are conducted on two MAD benchmarks, MVTec 3D-AD and Eyecandies. Notably, the training-free ZUMA achieves state-of-the-art (SOTA) performance on both datasets, outperforming existing ZS MAD methods, including training-based approaches. Moreover, ZUMA-FT further extends the performance boundary of ZUMA with only 6.75 M learnable parameters.