NOVA: Normal-Side Modeling for Training-Free Zero-Shot Video Anomaly Detection

Wei-Chih Yin; Yun-Ching Kao; Cheng-Kuan Lin; Yu-Chee Tseng

cs.CVSep 6, 2026

NOVA: Normal-Side Modeling for Training-Free Zero-Shot Video Anomaly Detection

Authors: Wei-Chih Yin, Yun-Ching Kao, Cheng-Kuan Lin, Yu-Chee Tseng

Organizations: Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

Abstract

Training-free zero-shot video anomaly detection (ZS-VAD) leverages vision-language models (VLMs) to localize anomaly instances from a predefined anomaly vocabulary, without providing any video. Existing CLIP-based methods often emphasize anomaly-side semantics, while the competing normality side remains less carefully formulated. We identify two key limitations in existing solutions: (i) blurred decision boundary: normal prompts may contain ambiguous verbs, such as running, that are semantically close to anomalies, reducing normal and abnormal separation in the VLM embedding space; and (ii) modality gap: poor alignment between features of textual normal anchors and visual frames. We propose NOVA, a training-free ZS-VAD framework that strengthens the normal side at both linguistic and visual levels. NOVA introduces Normality-Aware Prompt Construction (NA), which excludes anomaly-adjacent verbs and biases normal descriptions toward static, low-motion scenes. To overcome the text-vision modality gap, NOVA constructs a Visual Normality Anchor (VNA), which creates a weighted visual normal anchor from the initial frames of each test video, providing a video-specific normal reference without task-specific training or annotations. NOVA achieves 89.86 percent AUC on UCF-Crime and 95.07 percent AUC and 84.82 percent AP on XD-Violence, reaching state-of-the-art performance among comparable training-free zero-shot methods.

Figures & tables

Figure 1: The NOVA framework. NOVA uses a frozen vision-language encoder and a pre-constructed prompt bank, and constructs a video-specific visual normality anchor at test time. The framework comprises four key components: Normality-Aware Prompt Construction (NA), prompt-contrastive anomaly scoring, a per-video Visual Normality Anchor (VNA), and Motion-Aware Stabilization (MAS).

	Method	UCF-Crime	XD-Violence
	Method	AUC (%) $\uparrow$	AP (%) $\uparrow$	AUC (%) $\uparrow$
Weakly supervised	RareAnom [ 21 ]	83.56	—	79.89
	VERA [ 29 ]	86.55	—	88.26
	CLIP-TSA [ 6 ]	87.58	82.19	—
	VadCLIP [ 26 ]	88.02	84.51	—
	Dong [ 4 ]	88.52	—	—

Table 1: Comparison with VAD methods under different training paradigms. Bold indicates the best result, while underline indicates the best result within a paradigm. “—” denotes results not reported in the original paper. † LAVIDA and † LaGoVAD use external segmentation data to synthesize pseudo-anomalies and are therefore not considered strictly training-free.

NA	VNA	MAS	UCF AUC (%)	XD AUC (%)	XD AP (%)
$\times$	$\times$	$\times$	$85.62\pm 1.41$	$93.51\pm 0.48$	$80.36\pm 1.51$
✓	$\times$	$\times$	$88.79\pm 0.50$	$94.29\pm 0.45$	$84.33\pm 1.60$
$\times$	✓	$\times$	$86.99\pm 1.03$	$94.08\pm 0.41$	$80.47\pm 1.45$
$\times$	$\times$	✓	$86.73\pm 1.27$	$91.54\pm 0.23$	$72.88\pm 0.64$
✓	✓	$\times$	$89.54\pm 0.45$	$94.85\pm 0.30$	$84.65\pm 1.35$
✓	$\times$	✓	$89.34\pm 0.61$	$91.07\pm 0.36$	$72.96\pm 0.66$

Table 2: Full-factorial ablation of NA, VNA, and MAS. Results are reported as mean

\pm

standard deviation over five independently generated prompt banks.

Config	UCF AUC (%)	XD AUC (%)	XD AP (%)
Template, anomaly classes only	66.85	90.99	77.94
LLM, anomaly classes only	$72.50\pm 2.09$	$91.81\pm 0.71$	$76.44\pm 2.37$
LLM, normal+anomaly (no NA)	$85.62\pm 1.41$	$93.51\pm 0.48$	$80.36\pm 1.51$
LLM, normal+anomaly (NA)	88.79 $\pm$ 0.50	94.29 $\pm$ 0.45	84.33 $\pm$ 1.60

Table 3: Decomposition of normal-side and anomaly-side contributions on UCF-Crime and XD-Violence.

Figure 2: t-SNE visualization of UCF-Crime prompt embeddings from five independently generated prompt banks, with and without NA. Stars denote the corresponding centroids; cosine similarities are computed in the original embedding space.

Figure 3: Geometric analysis of VNA on the representative UCF-Crime video Burglary017_x264 . (a) Per-video PCA projection of frame and prompt embeddings using the seed-10 prompt bank, together with the visual normality anchor $\mathbf{a}_{\mathrm{VNA}}$ and the ground-truth normal-frame centroid $c_{n}$ used only for post-hoc analysis. (b) Original-space cosine similarities to $c_{n}$ for textual normal and anomaly centroids across 13 anomaly categories and five independently generated prompt banks, together with the five VNA anchors, one per prompt bank. Large markers and error bars denote mean $\pm$ standard deviation.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Generator	Configuration	UCF AUC	XD AUC	XD AP
GPT-5.4-mini	NA-only	88.87	92.99	79.15
GPT-5.4-mini	Full NOVA	90.09	94.69	83.21
Gemini 3.5	NA-only	89.19	92.90	79.77
Flash Lite	Full NOVA	90.33	95.34	84.27
Qwen-2.5-72B-	NA-only	87.81	94.27	84.12
Instruct	Full NOVA	89.21	95.48	85.40

Appendix

Table S1 : Sensitivity to the prompt-generation LLM. Each model generates

M=20

positive and 20 negative descriptions per anomaly category. Results are percentages from one prompt-generation seed; all downstream settings are fixed.

$M$	UCF AUC	XD AUC	XD AP
10	87.53	93.86	83.69
20	88.64	94.31	84.60
40	89.07	94.50	83.76
100	89.22	94.58	84.19

Appendix

Table S2 : Sensitivity to the number

M

of positive and negative prompt descriptions per anomaly category. For each dataset, every row uses the same generation seed and draws

M

descriptions per side from the same set of 100 positive and 100 negative descriptions. NA is enabled, whereas VNA and MAS are disabled.

Figure S1 : Per-video PCA projections of six illustrative UCF-Crime videos, including the Burglary017_x264 example shown in the main paper and five additional examples. The yellow star denotes the score-weighted $\mathbf{a}_{\mathrm{VNA}}$ , and the teal cross denotes the ground-truth normal-frame centroid $c_{n}$ . Blue/red dots are normal/anomalous frames; blue/red triangles are category-specific normal/anomaly prompts; gray markers are prompts from the other categories. Annotated cosine similarities are computed in the original embedding space.

$N_{\text{warm}}$	UCF AUC	XD AUC	XD AP
8	$89.76\pm 0.40$	$94.62\pm 0.38$	$83.53\pm 0.68$
16	$89.88\pm 0.41$	$94.91\pm 0.35$	$84.28\pm 0.59$
24	$89.87\pm 0.42$	$95.01\pm 0.33$	$84.60\pm 0.58$
$32^{\dagger}$	$89.86\pm 0.43$	$95.07\pm 0.33$	$84.82\pm 0.56$
48	$89.72\pm 0.45$	$95.11\pm 0.33$	$84.98\pm 0.54$
64	$89.49\pm 0.47$	$95.19\pm 0.33$	$85.20\pm 0.53$

Appendix

Table S3 : Sensitivity analysis of

N_{\text{warm}}

. Results are the mean and standard deviation over five prompt banks. † denotes the default used in the main experiments.

Panel of Figure 11 — Figure S2 : Per-frame qualitative analysis on UCF-Crime. Pink regions denote the annotated ground-truth anomaly frames.

Explore similar work

Mar 9, 2026

VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer

Zero-shot anomaly detection (ZSAD) requires detecting and localizing anomalies without access to target-class anomaly samples. Mainstream methods rely on vision-language models (VLMs) such as CLIP: they build hand-crafted or learned prompt sets for normal and abnormal semantics, then compute image-text similarities for open-set discrimination. While effective, this paradigm depends on a text encoder and cross-modal alignment, which can lead to training instability and parameter redundancy. This work revisits the necessity of the text branch in ZSAD and presents VisualAD, a purely visual framework built on Vision Transformers. We introduce two learnable tokens within a frozen backbone to directly encode normality and abnormality. Through multi-layer self-attention, these tokens interact with patch tokens, gradually acquiring high-level notions of normality and anomaly while guiding patches to highlight anomaly-related cues. Additionally, we incorporate a Spatial-Aware Cross-Attention (SCA) module and a lightweight Self-Alignment Function (SAF): SCA injects fine-grained spatial information into the tokens, and SAF recalibrates patch features before anomaly scoring. VisualAD achieves state-of-the-art performance on 13 zero-shot anomaly detection benchmarks spanning industrial and medical domains, and adapts seamlessly to pretrained vision backbones such as the CLIP image encoder and DINOv2. Code: https://github.com/7HHHHH/VisualAD

Yanning Hou, Peiyuan Li, Zirui Liu +4

College of Intelligence Science and Technology, National University of Defense Technology, Changsha, China · School of Artificial Intelligence, Anhui University, Hefei, China · State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, Anhui University, Hefei, China

Zero-Shot Anomaly Detection Vision Foundation Model Adaptation Medical Image Anomaly Detection

Ask PDF

Oct 1, 2026

Cog-VADU: A Training-Free Cognitive Reasoning Framework for Video Anomaly Detection and Understanding

Video Anomaly Detection (VAD) aims to temporally localize abnormal events in videos. Most existing approaches rely on dataset-specific training and curated annotations, limiting generalization in open-set scenarios. Recent zero-shot methods based on Large Vision- Language Models (LVLMs) alleviate this dependency but often lack temporal continuity and structured reasoning. We propose Cog-VADU, a fully training-free framework that reformulates VAD as a sequential cognitive reasoning task. Cog-VADU introduces Chain-of- Anomaly Detection Thought Prompting (CoADTP), which unrolls an LVLM into a recurrent reasoning chain across video segments. By propagating structured rationales over time, the model maintains implicit temporal memory, enabling robust discrimination between com- plex anomalies and high-motion normal activities. To improve reliability, we further design a cross-modal re-ranking stage that aligns textual rationales with visual embeddings, enforcing semantic consistency and temporal coherence for refined and stable predictions. Extensive experiments on multiple public VAD benchmarks demonstrate that Cog-VADU achieves competitive zero-shot performance. Moreover, cross-model evaluations show that CoADTP consistently enhances reasoning-based anomaly detection in a model-agnostic manner, pro- viding interpretable and generalizable anomaly understanding for real-world applications.

Mohd Ubaid Wani, Sara Atito, Josef Kittler +1

Centre for Vision, Speech and Signal Processing (CVSSP) University of Surrey, UK · Surrey Institute for People-Centred AI (PAI) University of Surrey, UK

CoT Reasoning Interpretable Anomaly Detection Video Anomaly Detection Zero-Shot Anomaly Detection Temporal Video Understanding

Ask PDF

Feb 22, 2026

No Need For Real Anomaly: MLLM Empowered Zero-Shot Video Anomaly Detection

The collection and detection of video anomaly data has long been a challenging problem due to its rare occurrence and spatio-temporal scarcity. Existing video anomaly detection (VAD) methods under perform in open-world scenarios. Key contributing factors include limited dataset diversity, and inadequate understanding of context-dependent anomalous semantics. To address these issues, i) we propose LAVIDA, an end-to-end zero-shot video anomaly detection framework. ii) LAVIDA employs an Anomaly Exposure Sampler that transforms segmented objects into pseudo-anomalies to enhance model adaptability to unseen anomaly categories. It further integrates a Multimodal Large Language Model (MLLM) to bolster semantic comprehension capabilities. Additionally, iii) we design a token compression approach based on reverse attention to handle the spatio-temporal scarcity of anomalous patterns and decrease computational cost. The training process is conducted solely on pseudo anomalies without any VAD data. Evaluations across four benchmark VAD datasets demonstrate that LAVIDA achieves SOTA performance in both frame-level and pixel-level anomaly detection under the zero-shot setting. Our code is available in https://github.com/VitaminCreed/LAVIDA.

Zunkai Dai, Ke Li, Jiajia Liu +2

Beijing University of Posts and Telecommunications · Northwestern Polytechnical University

Video Anomaly Detection Zero-Shot Anomaly Detection Synthetic Anomaly Generation

Ask PDF