Abstract
Progress in video anomaly understanding (VAU) has long been limited by inherent deficiencies of real-world anomaly videos, which are hard to collect and offer little control over their content. Synthetic anomaly approaches partially alleviate data scarcity, yet their generation remains largely controlled at the category or prompt level. They also lack component-level verification of video-text consistency and provide insufficient hard normal samples near the normal-anomaly boundary. To address this, we present AnomalyCraft-700K, a component-controllable synthetic anomaly dataset for fine-grained VAU, containing over 40K videos and over 700K task-level textual annotations. From fine-grained semantic components and a progressive three-stage pipeline, we craft anomaly events that are richly detailed, semantically controlled, and temporally structured, and additionally construct per-category hard normal samples to prompt the model to discriminate based on anomaly semantics rather than surface visual cues. Moreover, using the components as verification units, AnomalyCraft-700K further performs component-wise correction of video–text discrepancies introduced during generation, providing reliable annotations with verified cross-modal alignment for six tasks that progress from anomaly detection, through anomaly retrieval and captioning, to fine-grained anomaly reasoning. Evaluations of widely used methods under both traditional and MLLM-based protocols demonstrate that AnomalyCraft-700K serves as an effective source of supervision, from anomaly detection to fine-grained anomaly understanding.
Synthetic Video Examples
Anomaly Videos with Animate Subjects
Anomaly Videos with Inanimate Subjects
Normal Videos
Visual cues commonly associated with anomalous events:
1Sparks and smoke. 2Mud splashes and pronounced body motion.
Why AnomalyCraft?
Fine-grained VAU goes beyond assigning category labels: it requires models to accurately understand anomaly semantics from multiple perspectives, including identifying the fine-grained visual evidence that characterizes the event, determining who is involved and how entities interact, and reasoning about why the event is anomalous.
Real-world anomaly datasets face inherent limitations:
- Real-world anomaly videos are difficult to collect, with samples often unevenly distributed across anomaly categories.
- Balancing annotation cost and the coverage of fine-grained anomaly semantics remains challenging: manual annotation is costly, while generic automatic annotation often overlooks subtle but distinctive details.
Although recent video generation models provide a new route toward scalable anomaly data construction, several challenges remain unresolved:
- Existing control is mainly exercised over scenes, anomaly categories, and storyline segments, while the semantic constituents that determine how an anomaly is realized remain insufficiently parameterized.
- Generation prompts are often inherited as annotations, either directly or after event-level verification, leaving component-level semantic discrepancies largely unchecked.
- Normal samples are also generally introduced to enrich behavioral diversity rather than to form category-specific semantic boundaries with visually similar anomalies.
To address these limitations, we present AnomalyCraft-700K, a component-controllable synthetic anomaly dataset for fine-grained VAU, containing over 40K videos and over 700K task-level textual annotations.
- We propose the first synthetic anomaly dataset with controllable fine-grained anomaly semantics, achieved by composing structured semantic components through a three-stage pipeline that mitigates the limited anomaly priors of general-purpose video generators.
- We turn the original generation-time components into targeted units, checking and correcting component-level discrepancies between the intended semantic components and their visual realization item by item, pushing cross-modal alignment verification from the narrative level down to the component level and providing downstream tasks with verified fine-grained alignment annotations.
- Guided by the surface visual cues of each anomaly category, we construct boundary samples that are visually close to anomalies yet semantically normal, so as to prompt the model to discriminate based on anomaly semantics rather than surface visual cues.
- We provide unified supervision for six tasks spanning three capability dimensions, supporting a complete evaluation from anomaly detection to fine-grained semantic understanding.
Construction and Dataset Overview
AnomalyCraft-700K decomposes each anomaly event into five core components, including context, background, subject, object, and directed interaction, together with category-specific attributes. A progressive three-stage pipeline gradually relaxes non-essential constraints while preserving the defining anomaly semantics, balancing precise controllability with event diversity and temporal structure. The same semantic components are further reused as verification units, allowing component-to-video discrepancies to be identified and manually corrected rather than directly treating generation prompts as annotations. We also construct category-specific hard normal counterparts that preserve anomaly-like visual cues while remaining semantically normal, encouraging models to distinguish anomalies through event semantics rather than superficial appearance.
To support comprehensive anomaly understanding, AnomalyCraft-700K provides six tasks across three capability dimensions: anomaly detection, anomaly retrieval and captioning, and fine-grained anomaly reasoning. It contains 40,100 videos, more than 3 million frames, and over 700K task-level multimodal supervision instances, enabling evaluation from conventional anomaly detection to fine-grained semantic reasoning.
Benchmark Results
We benchmark representative methods on AnomalyCraft-700K across six tasks in three capability dimensions: Anomaly Detection (T1 Video Anomaly Detection and T3 Anomaly Judgment); Anomaly Retrieval and Captioning (T2 Video Anomaly Retrieval and T4 Structured Event Captioning); and Fine-grained Anomaly Reasoning (T5 Structured Event Reasoning and T6 Fine-Grained Question Answering).
Results on Anomaly Detection
| Method | Binary (%) | 30-way (%) | ||
|---|---|---|---|---|
| AUROC | AP | Top-1 | Macro-F1 | |
| RTFM (2021) | 92.43 | 97.12 | – | – |
| UR-DMU (2023) | 94.00 | 97.64 | – | – |
| BN-WVAD (2024b) | 94.52 | 97.96 | – | – |
| PEL4VAD (2024) | 94.01 | 97.70 | – | – |
| JDC (2017) | 92.63 | 96.81 | 58.46 | 50.84 |
| VadCLIP (2024) | 96.82 | 98.72 | 73.94 | 69.81 |
| AnomalyCLIP (2024a) | 96.43 | 98.40 | 67.73 | 60.67 |
| DSANet (2026) | 96.57 | 98.58 | 72.39 | 68.57 |
Table 2 shows that all baselines achieve strong binary detection, indicating that the generated videos contain consistent and discriminative visual evidence of anomalous events. However, category-level recognition remains substantially more challenging, showing that identifying whether an event is anomalous is considerably easier than identifying its specific category.
| Model (set.) | AUROC (%) | Top-1 (%) | Macro-F1 (%) |
|---|---|---|---|
| Qwen2.5-3B (ZS) | 42.37 | 27.36 | 6.05 |
| Qwen2.5-3B (SFT) | 95.78 | 82.97 | 80.99 |
| Qwen3-4B (ZS) | 86.33 | 28.48 | 11.91 |
| Qwen3-4B (SFT) | 96.81 | 87.13 | 85.07 |
| Qwen3-8B (ZS) | 86.78 | 25.98 | 8.54 |
| InternVL3-8B (ZS) | 30.89 | 14.17 | 9.22 |
Table 3 shows the same pattern for MLLM-oriented anomaly judgment: the zero-shot results reveal a clear discrepancy between coarse anomaly awareness and fine-grained semantic discrimination, while supervised fine-tuning on AnomalyCraft-700K substantially improves category accuracy and macro-F1. Together, T1 and T3 demonstrate that AnomalyCraft-700K provides reliable anomaly evidence while offering the fine-grained supervision required to transform coarse anomaly detection into category-level semantic understanding.
Results on Anomaly Retrieval and Captioning
| Method | T→V (%) | V→T (%) | ||
|---|---|---|---|---|
| R@1 | R@5 | R@1 | R@5 | |
| CLIP4Clip (2022) | 85.36 | 97.52 | 87.48 | 97.89 |
| X-CLIP-B/32 (2022) | 85.42 | 97.53 | 86.29 | 97.73 |
| X-CLIP-B/16 (2022) | 87.99 | 98.37 | 87.82 | 98.17 |
| TS2-Net (2022) | 84.31 | 97.54 | 87.28 | 97.92 |
| UCoFiA (2023) | 86.89 | 97.87 | 87.32 | 98.13 |
| EagleNet (2026) | 82.10 | 96.31 | 82.84 | 96.71 |
Table 4 reports consistently strong bidirectional retrieval across architectures, indicating reliable global correspondence between the generated videos and their verified textual annotations. As T2 is intended to assess supervision quality rather than retrieval difficulty alone, these results primarily validate the reliability of verified video-text alignment.
| Model (set.) | B-4 | R-L | CIDEr | Slot |
|---|---|---|---|---|
| Qwen2.5-3B (ZS) | 0.0479 | 0.2247 | 0.7825 | 0.7215 |
| Qwen2.5-3B (SFT) | 0.1190 | 0.3219 | 1.4024 | 0.7853 |
| Qwen3-4B (ZS) | 0.0515 | 0.2238 | 0.8892 | 0.7160 |
| Qwen3-4B (SFT) | 0.1384 | 0.3459 | 1.6406 | 0.7912 |
| Qwen3-8B (ZS) | 0.0519 | 0.2217 | 0.9056 | 0.7153 |
| InternVL3-8B (ZS) | 0.0485 | 0.2380 | 0.9246 | 0.7332 |
Table 5 further shows that supervised fine-tuning consistently improves structured event captioning. Supplementary slot-level results indicate improved grounding of individual event components. Together, T2 verifies reliable global video-text association, while T4 demonstrates that AnomalyCraft-700K supervision supports fine-grained grounding and description of event components rather than merely learning a template-like output.
Results on Fine-grained Anomaly Reasoning
| Model | Setup | CLS | KM | FLU | INF | FAC | Total |
|---|---|---|---|---|---|---|---|
| Qwen2.5-3B | ZS | 2.03 | 2.42 | 6.50 | 2.47 | 2.65 | 16.06 |
| Qwen2.5-3B | SFT | 8.34 | 6.82 | 8.60 | 7.03 | 6.78 | 37.57 |
| Qwen3-4B | ZS | 5.96 | 5.24 | 8.23 | 5.49 | 5.01 | 29.93 |
| Qwen3-4B | SFT | 8.77 | 7.80 | 8.93 | 7.95 | 7.77 | 41.22 |
| Qwen3-8B | ZS | 6.53 | 5.60 | 8.35 | 5.87 | 5.25 | 31.59 |
| InternVL3-8B | ZS | 3.17 | 4.40 | 7.30 | 3.99 | 4.85 | 23.72 |
Table 6 shows that supervised fine-tuning consistently yields substantial improvements across all five reasoning dimensions. The marked gains in classification, key-event matching, informativeness, and factual consistency indicate that AnomalyCraft-700K primarily enhances semantic understanding and structured reasoning rather than surface-level language generation. The zero-shot Qwen3-8B result further suggests that high-quality task supervision contributes more to fine-grained anomaly reasoning than simply increasing model scale.
| Model/set. | QA | MCQ | ||
|---|---|---|---|---|
| B-F1 | Corr. | Acc. | HN Acc. | |
| Q2.5-3B/ZS | 0.7048 | 0.2064 | 0.8480 | 0.7362 |
| Q2.5-3B/SFT | 0.7883 | 0.5816 | 0.9872 | 0.9858 |
| Q3-4B/ZS | 0.6983 | 0.3559 | 0.8060 | 0.7455 |
| Q3-4B/SFT | 0.7952 | 0.7522 | 0.9904 | 0.9881 |
Table 7 shows the same overall benefit for fine-grained question answering: QA correctness improves substantially after fine-tuning, whereas the corresponding B-F1 gains are much smaller. This discrepancy indicates that lexical similarity alone cannot adequately measure semantic correctness, highlighting the necessity of reasoning-oriented evaluation. For multiple-choice QA, the largest improvements are observed on the hard normal subset, confirming that the proposed hard normal samples provide an effective test of semantic boundary understanding for visually similar events.
Note: Table numbers follow those in the paper.
Qualitative Results
We further present representative qualitative results for structured event captioning (T4), structured event reasoning (T5), and fine-grained question answering (T6). Green and red highlights indicate correctly grounded and incorrect or unsupported model outputs, respectively. Bracketed ellipses denote omitted text.
Structured Event Captioning (T4)
Figure 3 compares the zero-shot and SFT outputs of Qwen3-VL-4B on two illustrative cases. In AnimalAttackFarm, the zero-shot model captures the overall danger but reverses the subject and object roles and weakens the direct attack to “sniffing and circling.” SFT correctly grounds the boars as the attackers and the man as the victim, while recovering the progression from pursuit to falling and continued attack. In ErraticBehavior, the zero-shot model promotes a possible medical detail to the Object slot and constructs an unsupported causal interpretation around it. SFT instead identifies that no explicit object is involved and grounds the anomaly in the boy's uncontrolled vocalizations and body movements. These cases illustrate that AnomalyCraft's six-slot annotations provide supervision beyond generic captioning, supporting fine-grained role grounding, anomaly attribution, and temporal event modeling.
Structured Event Reasoning (T5)
Figure 4 compares the zero-shot and SFT outputs of Qwen3-VL-4B on a Hard Normal example. The zero-shot model incorrectly infers a raised truck bed, spilled debris, and uncontrolled motion from the salient water spray, extends these unsupported details into an accident trajectory, and consequently judges the normal event as anomalous. The SFT model instead attributes the spray to the wet road, recognizes the truck's steady motion and the absence of collision, overturning, or damage, and reaches the correct Normal judgment. This case illustrates how AnomalyCraft's structured multi-stage analyses and Hard Normal annotations provide supervision for distinguishing visually salient yet harmless effects from genuine anomalies through evidence-grounded process and outcome reasoning.
Fine-Grained Question Answering (T6)
Figure 5 presents three selected Open-QA pairs from the same Shooting example, covering the event onset, core interaction, and final state. The zero-shot model consistently interprets the shooting as a playful or training interaction: it misidentifies the event trigger, characterizes the dog's evasive response as excitement, and infers a harmless resolution. After SFT, the model correctly identifies the man raising and firing the handgun, the dog's startled retreat, and the unresolved threat at the end of the video. Accordingly, all three answers change from rejected to accepted by the judge. These examples illustrate that AnomalyCraft's fine-grained QA annotations support evidence-grounded understanding of event progression and subject–object dynamics beyond coarse scene recognition.