MIRAGE: Model-agnostic Industrial Realistic Anomaly Generation and Evaluation for Visual Anomaly Detection

Jinwei Hu; Francesco Borsatti; Arianna Stropeni; Davide Dalle Pezze; Manuel Barusco; Gian Antonio Susto

cs.CVMar 13, 2026

MIRAGE: Model-agnostic Industrial Realistic Anomaly Generation and Evaluation for Visual Anomaly Detection

Authors: Jinwei Hu, Francesco Borsatti, Arianna Stropeni, Davide Dalle Pezze, Manuel Barusco, Gian Antonio Susto

Organizations: University of Padova, Italy

Abstract

Industrial visual anomaly detection (VAD) methods are typically trained on normal samples only, yet performance improves substantially when even limited anomalous data is available. Existing anomaly generation approaches either require real anomalous examples, demand expensive hardware, or produce synthetic defects that lack realism. We present MIRAGE (Model-agnostic Industrial Realistic Anomaly Generation and Evaluation), a fully automated pipeline for realistic anomalous image generation and pixel-level mask creation that requires no training and no anomalous images. Our pipeline accesses any generative model as a black box via API calls, uses a VLM for automatic defect prompt generation, and includes a CLIP-based quality filter to retain only well-aligned generated images. For mask generation at scale, we introduce a lightweight, training-free dual-branch semantic change detection module combining text-conditioned Grounding DINO features with fine-grained YOLOv26-Seg structural features. We benchmark four generation methods using Gemini 2.5 Flash Image (Nano Banana) as the generative backbone, evaluating performance on MVTec AD and VisA across two distinct tasks: (i) downstream anomaly segmentation and (ii) visual quality of the generated images, assessed via standard metrics (IS, IC-LPIPS) and a human perceptual study involving 31 participants and 1,550 pairwise votes. The results demonstrate that MIRAGE offers a scalable, accessible foundation for anomaly-aware industrial inspection that requires no real defect data. As a final contribution, we publicly release a large-scale dataset comprising 500 image-mask pairs per category for every MVTec AD and VisA class, over 13,000 pairs in total, alongside all generation prompts and pipeline code.

Figures & tables

Figure 1: Overview of MIRAGE. Top: A normal image and an anomaly text prompt are fed to Gemini 2.5 Flash Image, which generates a photorealistic anomalous image. Bottom (Mask Creation Pipeline): The normal image, anomalous image, and text prompt are processed through two branches. The semantic branch (left) computes text-conditioned feature differences using Grounding DINO, yielding a coarse semantic anomaly map. The structural branch (right) computes unconditional feature differences using YOLOv26-L-Seg, yielding a fine-grained structural change map. The element-wise product of both maps produces the final anomaly mask.

Figure 2: Examples of synthetic anomalies generated. From left to right: (1) original input image, (2) GLASS, (3) RealNet, (4) AnomalyAny, and (5) MIRAGE. The rows represent three object classes: bottle (top row), hazelnuts (mid row) and electronic sensor modules (bottom row). The relative generated anomaly map or mask is overlaid on the bottom right of each image.

Method	TrueSkill ( $\mu\pm\sigma$ ) $\uparrow$	Win% $\uparrow$	App.
Real images	$28.61\pm 0.80$	73.8	644
MIRAGE (Ours)	$28.33\pm 0.80$	67.2	622
AnomalyAny [ 23 ]	$27.16\pm 0.79$	59.2	613
RealNet [ 26 ]	$23.54\pm 0.79$	33.7	642
GLASS [ 4 ]	$20.26\pm 0.84$	14.4	605

Table 1: Human perceptual realism evaluation. TrueSkill ranking [ 9 ] (

\mu\pm\sigma

, higher is better), win rate (%, excluding ties), and number of appearances across all pairwise comparisons (31 participants, 1,550 total votes). Real images are included as an upper bound.

	AnomalyAny		RealNet		GLASS		MIRAGE
Category	IS	IC-L	IS	IC-L	IS	IC-L	IS	IC-L
Bottle	1.73	0.17	1.58	0.16	2.08	0.26	2.31	0.23
Cable	2.06	0.41	1.65	0.18	1.81	0.42	2.63	0.53
Capsule	2.16	0.23	1.62	0.41	2.50	0.33	2.27	0.20
Carpet	1.10	0.34	1.03	0.25	1.29	0.27	2.09	0.33
Grid	2.31	0.38	2.24	0.37	2.46	0.41	2.80	0.56

Table 2: Visual quality metrics on MVTec AD. Inception Score (IS, higher is better) and Intra-Cluster LPIPS (IC-LPIPS, denoted IC-L) for generated anomalous images from each method, compared against real anomalous test images.

(a)	AnomAny		RealNet		GLASS		Ours
MVTec AD	I	P	I	P	I	P	I	P
Bottle	0.78	0.85	0.92	0.91	0.91	0.83	0.93	0.92
Cable	0.49	0.76	0.41	0.68	0.74	0.80	0.71	0.86
Capsule	0.57	0.89	0.49	0.88	0.43	0.87	0.45	0.93
Carpet	0.39	0.90	0.45	0.86	0.56	0.98	0.82	0.99
Grid	0.57	0.74	0.82	0.63	0.93	0.98	0.54	0.90

Table 3: Downstream anomaly segmentation. Image-level AUROC (I) and pixel-level AUROC (P) for a U-Net trained on synthetic data from each generation method. Best results in bold . (a) MVTec AD. (b) VisA.

	MVTec AD		VisA
Configuration	I	P	I	P
Without CLIP filter	0.70	0.92	0.71	0.91
With CLIP filter	0.80	0.92	0.74	0.92

Table 4: Effect of CLIP-based quality filtering on downstream anomaly segmentation. Mean image-level AUROC (I) and pixel-level AUROC (P) on MVTec AD and VisA.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Pixel-level AUC
bottle	0.9650
cable	0.8866
capsule	0.9413
carpet	0.9263
grid	0.8487
hazelnut	0.9718

Appendix

Table 5: MVTec AD pixel-level AUC on 1000 ground-truth masks generated by Gemini Image V3. Results are reported for the g-dino + yolo pipeline.

Category	Pixel-level AUC
candle	0.8967
capsules	0.8836
cashew	0.9459
chewinggum	0.8483
fryum	0.9553
macaroni1	0.9326

Appendix

Table 6: VisA pixel-level AUC on 720 ground-truth masks generated by Gemini Image V3. Results are reported for the g-dino + yolo pipeline.

Figure 3: Different types of synthetic anomalous images and corresponding heatmaps generated by MIRAGE.

[Class]	[Defect Type]	GPT-Generated Detailed Descriptions
Wood $\star$	Vertical Split	A deep crack runs parallel to the wood grain, showing a physical separation where the material has dried out and pulled apart, creating a dark void.
	Water Stain	An irregular, dark ring or splotch discolors the wood surface, creating a distinct boundary where moisture has penetrated and altered the grain’s appearance.
	Surface Burn Mark	A dark, charred patch with blackened, carbonized wood fibers disrupts the natural color, indicating contact with a high-heat source.
	Splintered Surface	The surface fibers are raised and peeling away in jagged, needle-like strips, destroying the smooth finish and creating a rough, hazardous texture.
	Deep Scratch	A distinct, light-colored linear gouge cuts across the vertical grain, interrupting the smooth brown pattern where the surface fibers have been severed.
Zipper $\star$	Punctures or Tears	Visible holes or rips in the woven fabric tape, distinct from the edges.

Appendix

Table 7: Corresponding descriptions for anomaly generation of Fig. 3 .

\star

indicates MVTec AD dataset, while

\star\star

indicates VisA classes.

Explore similar work

Aug 8, 2026

UniScale: Arbitrary-Scale Industrial Anomaly Generation

Industrial anomaly inspection faces a major challenge due to the lack of real-world anomaly samples. While generative models are used to create anomaly data, existing methods still struggle when handling small-scale anomalies. This failure occurs because extreme downsampling in diffusion models causes the information of small anomalies to be lost in the latent space. To address this, we introduce UniScale, a unified training and inference framework for high-fidelity industrial anomaly generation across arbitrary scales. During training, we introduce an Error-Suppressed Multi-Scale Training (EMT) strategy, which enables the model to learn the rich location-aware textures of anomalies, while suppressing upsampling-induced interpolation errors in texture acquisition, ensuring the model is capable of learning small-scale anomalies, while remaining effective for regular scale anomalies. For inference, we propose Generation-then-Fusion Denoising. It decouples anomaly generation from background integration, preventing small anomalies from being overwhelmed. Extensive experiments demonstrate that our method outperforms state-of-the-art competitors in both anomaly generation quality and downstream detection performance. It achieves a relative IS(a) improvement of 45.86% (from 1.81 to 2.64) on VisA and 37.70% (from 1.22 to 1.68) on MVTec AD 2, while also improving the downstream pixel-level IoU by 4.22% on VisA and AUROC by 6.55% on MVTec AD 2. Code is available at https://github.com/HUST-SLOW/UniScale.

Shilei Zeng, Linxin Guan, Xurui Li +2

School of Electronic Information and Communications, Huazhong University of Science and Technology

Industrial Anomaly Detection Synthetic Anomaly Generation Diffusion Models

Ask PDF

Apr 9, 2026

AnomalyAgent: Agentic Industrial Anomaly Synthesis via Tool-Augmented Reinforcement Learning

Industrial anomaly generation is a crucial method for alleviating the data scarcity problem in anomaly detection tasks. Most existing anomaly synthesis methods rely on single-step generation mechanisms, lacking complex reasoning and iterative optimization capabilities, making it difficult to generate anomaly samples with high semantic realism. We propose AnomalyAgent, an anomaly synthesis agent with self-reflection, knowledge retrieval, and iterative refinement capabilities, aiming to generate realistic and diverse anomalies. Specifically, AnomalyAgent is equipped with five tools: Prompt Generation (PG), Image Generation (IG), Quality Evaluation (QE), Knowledge Retrieval (KR), and Mask Generation (MG), enabling closed-loop optimization. To improve decision-making and self-reflection, we construct structured trajectories from real anomaly images and design a two-stage training framework: supervised fine-tuning followed by reinforcement learning. This process is driven by a three-part reward mechanism: (1) task rewards to supervise the quality and location rationality of generated anomalies; (2) reflection rewards to train the model's ability to improve anomaly synthesis prompt; (3) behavioral rewards to ensure adherence to the trajectory. On the MVTec-AD dataset, AnomalyAgent achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at the image/pixel level using a simple UNet, surpassing all zero-shot SOTA methods. The code is publicly available at https://github.com/sssjmmm/AnomalyAgent.

Jiaming Su, Tengchao Yang, Ruikang Zhang +3

Shanghai Jiao Tong University Shanghai, China · Tongji University Shanghai, China · Fudan University Shanghai, China

Synthetic Anomaly Generation Tool-Using Agents Industrial Anomaly Detection RL for Generative Models

Ask PDF

Sep 29, 2026

FLASH: A "Generate Once, Synthesize Many" Framework for Synthetic Anomaly Generation in Industrial Anomaly Detection

Synthetic anomaly generation helps expand industrial anomaly datasets when real defects are scarce or unavailable. Existing approaches lie at two extremes: procedural approaches are fast but struggle to represent complex anomalies, while generative approaches produce diverse defects but require costly per-sample generation. We present FLASH, a framework that decouples defect generation from anomaly synthesis under a ``generate once, synthesize many'' paradigm. Given only normal images, FLASH uses Vision-Language Model (VLM) guidance and an image-generation model to produce a small set of defect images, from which it extracts, validates, and banks reusable defect patches. For synthesis of anomalous images, Object Boundary Suppression (OBS) first identifies the probable foreground object-aware region of the host image, while Multi-Resolution Spectral Pyramid (MRSP) noise generates diverse, size-controllable masks that determine the defect location and spatial extent. It then composes a large and diverse synthetic anomalous image set by localizing the defect region, sampling size-controllable placement masks and seamlessly blending retrieved defects onto new defect-free images without further need for image generation. Experiments on the MVTec AD 2 dataset show that FLASH-generated anomalies nearly close the calibration gap on real defects, reaching 78.1% image-level F1 against an 83.6% real-anomaly upper bound and providing the most consistent calibration transfer across detectors among procedural and generative alternatives. Moreover, FLASH synthesizes anomalies more than 11.95x faster than per-sample generative approaches.

Abhay Kumar Das, Rajesh Gangireddy, Ashwin Vaidya +1

Silicon University, Bhubaneswar, India · Intel

Synthetic Data Augmentation Industrial Anomaly Detection Synthetic Anomaly Generation

Ask PDF