💻 computer science

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

MIRAGE is a fully automated, training-free pipeline that leverages black-box generative models, vision-language models, and a dual-branch change detection module to generate realistic industrial anomaly images and pixel-level masks without requiring real defect data, thereby enhancing visual anomaly detection performance and providing a large-scale open-source dataset.

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

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a quality inspector at a factory that makes thousands of identical products every day, like glass bottles or electronic chips. Your job is to spot the "bad" ones—those with scratches, cracks, or dents.

The problem? You've never seen a bad one. The factory is so perfect that you only have photos of "good" products. Traditional AI models try to learn what "perfect" looks like and then scream "ALARM!" if anything looks even slightly different. But this is tricky; sometimes a shadow looks like a scratch, or a smudge looks like a crack, leading to false alarms.

To fix this, you'd love to show the AI photos of real defects. But in a perfect factory, real defects don't exist (or are too rare to collect).

Enter MIRAGE. Think of MIRAGE as a magical "What-If" machine that can create fake, but incredibly realistic, photos of broken products so the AI can learn from them.

Here is how MIRAGE works, broken down into simple steps:

1. The "Imagination" Engine (The Generative Model)

Most previous methods tried to build their own "imagination" engine from scratch, which was like trying to build a new car engine every time you wanted to drive. It was expensive, slow, and hard to upgrade.

MIRAGE is different. It doesn't build an engine; it rents a super-car.

The Analogy: Imagine you want to see what a "scratched bottle" looks like. Instead of painting it yourself, you call a professional artist (an AI model called Gemini) via a phone app (API).
You say: "Here is a photo of a perfect bottle. Now, imagine it has a scratch on the side."
The artist instantly sends back a photo that looks so real, you can almost feel the scratch.
The Magic: If a better artist becomes available next year, you just change the phone number you call. You don't have to rebuild the whole system.

2. The "Bouncer" (The Quality Filter)

Sometimes, the artist gets it wrong. They might draw a scratch that looks like a crack, or they might accidentally change the color of the bottle.

The Analogy: MIRAGE has a strict bouncer at the door (a CLIP-based filter). Before the fake photo is allowed into the training room, the bouncer checks: "Does this photo actually look like a scratched bottle, or is it just a blue bottle?"
If the photo doesn't match the description perfectly, the bouncer kicks it out. This ensures the AI only learns from high-quality examples.

3. The "X-Ray Vision" (The Mask Generator)

This is the trickiest part. When the artist draws a scratch, the AI needs to know exactly which pixels are the scratch and which are the bottle. It needs a "mask" (a stencil) to highlight the damage.

The Problem: The artist might change the lighting slightly or shift the angle, making it hard to tell what changed just by looking at the pixels.
The MIRAGE Solution: It uses a two-brain approach:
1. The Semantic Brain: This brain reads the text ("scratch") and looks for anything that looks like a scratch.
2. The Structural Brain: This brain ignores the text and just looks for any physical change in the shape or texture between the perfect photo and the fake one.
The Result: It combines these two views. If the Semantic Brain says "That looks like a scratch" AND the Structural Brain says "The texture changed there," BAM! It draws a perfect stencil around the defect. It does this without needing to be taught how to do it first.

4. The "Grand Experiment" (The Results)

The team tested MIRAGE against other methods using two famous industrial datasets (MVTec AD and VisA).

The Human Test: They showed real humans photos of real defects vs. MIRAGE's fake defects. The humans couldn't tell the difference most of the time! MIRAGE's fake defects were rated almost as realistic as real ones.
The AI Test: They trained a new AI inspector using only MIRAGE's fake photos. This new inspector became incredibly good at finding real defects, beating all other methods.

Why This Matters

No Real Defects Needed: You don't need to wait for a factory to break to train your AI.
Cheap and Easy: It doesn't require super-expensive computers. You just use an internet connection to call the "artist."
Future-Proof: As AI gets better, MIRAGE just swaps the "artist" for a better one.
Open Source: The team released over 13,000 of these fake defect photos and masks for free, so anyone can use them to build better safety systems.

In short: MIRAGE is a plug-and-play system that uses the power of modern AI artists to dream up perfect examples of "what could go wrong," allowing factories to train their safety systems without ever needing a single real broken product.

1. Problem Statement

Industrial Visual Anomaly Detection (VAD) typically operates under a one-class paradigm, where models are trained exclusively on defect-free ("normal") data. While effective, performance significantly improves when even a small amount of anomalous data is available for supervised training. However, industrial settings often lack real defect samples due to the rarity of defects and the high cost of data collection.

Existing solutions to generate synthetic anomalies face critical limitations:

Copy-paste methods (e.g., DRAEM, GLASS): Produce unrealistic textures (e.g., Perlin noise) that lack visual fidelity.
Diffusion-based methods (e.g., RealNet): Require training a separate Denoising Diffusion Probabilistic Model (DDPM) for every product category, leading to high computational costs and poor scalability.
Few-shot methods: Require real anomalous examples for training, which defeats the purpose of zero-shot generation.
State-of-the-art Zero-Shot (e.g., AnomalyAny): While effective, they are tightly coupled to specific diffusion architectures, require massive VRAM (30GB+), long generation times, and are difficult to upgrade when new generative models emerge.

The Goal: Create a fully automated, scalable, and realistic pipeline for generating anomalous images and pixel-level masks that requires no training, no real defect data, and no expensive hardware, while being easily upgradable to newer generative models.

2. Methodology: The MIRAGE Pipeline

MIRAGE (Model-agnostic Industrial Realistic Anomaly Generation and Evaluation) is a three-stage, model-agnostic pipeline that treats generative models as black boxes accessed via API.

Stage 1: Automated Defect Prompt Generation

Input: A set of normal reference images for a specific product category.
Process: A Vision-Language Model (VLM) (e.g., ChatGPT-5) analyzes the normal images and generates a structured list of 10 plausible manufacturing defect types with concise, descriptive text prompts (e.g., "surface scratch: a thin, shallow linear mark...").
Advantage: Eliminates the need for manual domain expertise and ensures prompts are grounded in the visual reality of the specific product.

Stage 2: Anomalous Image Generation & Quality Filtering

Generation: Uses a general-purpose generative model (currently Gemini 2.5 Flash Image, internally "Nano Banana") via API. The model takes a normal image and a defect prompt to generate a photorealistic anomalous image ( $I_a$ ) while preserving the original background, lighting, and layout.
CLIP-Based Quality Filter: To discard generation failures or artifacts, the pipeline computes four CLIP image-text similarities:
1. $I_a$ vs. Anomaly Prompt ( $sim(I_a, p_a)$ )
2. $I_a$ vs. Normal Prompt ( $sim(I_a, p_n)$ )
3. $I_n$ vs. Anomaly Prompt ( $sim(I_n, p_a)$ )
4. $I_n$ vs. Normal Prompt ( $sim(I_n, p_n)$ )
Criteria: An image is retained only if it aligns with the anomaly prompt better than the normal image does, and better than it aligns with the normal prompt. This ensures semantic coherence and defect presence.

Stage 3: Training-Free Dual-Branch Mask Generation

Generating accurate pixel-level masks without training is a major challenge. MIRAGE introduces a Dual-Branch Semantic Change Detection module:

Semantic Branch (Grounding DINO):
- Uses Grounding DINO (text-conditioned) to extract features from both the normal ( $I_n$ ) and anomalous ( $I_a$ ) images based on the defect keywords.
- Computes the difference: $S_{sem} = \| \phi_{GD}(I_a, t) - \phi_{GD}(I_n, t) \|_2$ .
- Purpose: Captures the semantic relevance of the defect, suppressing false positives from normal textures that might look like defects.
Structural Branch (YOLOv26-L-Seg):
- Uses YOLOv26-L-Seg (unconditional) to extract multi-scale structural features.
- Computes the difference across layers: $S_{str} = \| \psi_l(I_a) - \psi_l(I_n) \|_2$ .
- Purpose: Captures fine-grained structural changes and pixel-level deformations regardless of semantic meaning.
Fusion: The final mask is the element-wise product of the upsampled semantic and structural heatmaps ( $S = \hat{S}_{sem} \odot \hat{S}_{str}$ ).
Calibration: A single threshold is optimized per category using a small set of reference masks (generated by a higher-cost model) to maximize pixel-level AUROC.

3. Key Contributions

Model-Agnostic Pipeline: A plug-and-play framework that accesses generative models via API. It requires no local GPU training, no anomalous data, and can seamlessly swap in newer models (e.g., future VLMs or Image Generators) by simply changing the endpoint.
Training-Free Mask Generation: A novel dual-branch change detection module combining text-conditioned semantic features (Grounding DINO) and unconditional structural features (YOLOv26) to produce high-quality masks without any supervised training.
Extensive Benchmark: The most comprehensive evaluation of anomaly generation to date, covering:
- Downstream segmentation performance (U-Net).
- Automated metrics (Inception Score, IC-LPIPS).
- Human Perceptual Study: A study with 31 participants and 1,550 pairwise votes ranked via TrueSkill.
Large-Scale Public Dataset: Release of 13,000+ image-mask pairs (500 per category) for all MVTec AD and VisA classes, along with all generation prompts and code. This is the largest such release to date.

4. Results

The evaluation was conducted on MVTec AD and VisA benchmarks, comparing MIRAGE against GLASS, RealNet, and AnomalyAny.

Human Perceptual Quality:
- MIRAGE achieved a TrueSkill score of 28.33, extremely close to real images (28.61) and significantly outperforming AnomalyAny (27.16).
- Win Rate: MIRAGE won 67.2% of decisive pairwise comparisons against other methods.
Automated Metrics:
- Inception Score (IS): MIRAGE achieved the highest mean IS (2.68), indicating superior image quality and diversity.
- IC-LPIPS: MIRAGE scored 0.38, indicating strong perceptual alignment with real defects within clusters.
Downstream Segmentation (U-Net):
- MIRAGE achieved the best Pixel-Level AUROC on both datasets (0.92 on MVTec AD, 0.92 on VisA).
- CLIP Filtering Impact: Filtering out low-quality generations improved Image-Level AUROC by ~10% (0.70 $\to$ 0.80 on MVTec AD) without hurting pixel-level performance.
Mask Quality: The dual-branch change detection achieved a pixel-level AUROC of 0.93 against pseudo-ground truth, significantly outperforming View-Delta (0.73).

5. Significance and Impact

Democratization of Industrial AI: By removing the need for expensive GPUs, real defect data, and manual prompt engineering, MIRAGE makes high-quality anomaly detection accessible to a wider range of industrial applications.
Scalability and Future-Proofing: The "black box" API approach ensures that as generative AI models improve (e.g., better resolution, faster inference), the pipeline automatically benefits without architectural re-engineering.
Resource for the Community: The release of a massive, high-quality synthetic dataset (13k+ pairs) fills a critical gap in the research community, enabling reproducible research in supervised anomaly segmentation, defect classification, and industrial inspection.
Paradigm Shift: Moves the field away from complex, training-heavy diffusion models toward efficient, API-driven, model-agnostic workflows that prioritize realism and practical deployability.