SuperADD: Training-free Class-agnostic Anomaly Segmentation -- CVPR 2026 VAND 4.0 Workshop Challenge Industrial Track

The paper presents SuperADD, a training-free, class-agnostic anomaly segmentation pipeline that leverages a DINOv3 backbone and robust preprocessing techniques to achieve state-of-the-art performance on the MVTec AD 2 dataset under challenging distribution shifts without requiring per-class hyperparameter tuning.

The Gist

Imagine you are the head of quality control in a massive factory. Your job is to spot tiny defects on products rolling down a conveyor belt. Usually, you have a team of experts who have studied thousands of perfect products. They know exactly what a "good" wall plug, a piece of fabric, or a jar of jelly should look like. If they see something that doesn't match that perfect memory, they flag it as a defect.

However, there's a catch: the lighting in the factory keeps changing. Sometimes it's bright, sometimes dim, sometimes the shadows are weird. This confuses the experts because the same perfect product looks different under different lights. They might start crying "Defect!" when it's actually just a shadow, or worse, they might miss a real crack because the light is hiding it.

This paper presents a new, super-smart system called SuperADD designed to solve this exact problem. Here is how it works, broken down into simple concepts:

1. The "No-Training" Superpower

Most AI systems are like students who need to sit in a classroom for months to learn what a defect looks like for each specific product. If you introduce a new product or change the lighting, you have to send them back to school to relearn everything.

SuperADD is different. It's like a detective who doesn't need to study the specific product beforehand. It uses a pre-trained "brain" (called DINOv3) that has already seen millions of images from the internet. It knows what "normal" textures and shapes generally look like. Because it doesn't need to be retrained for every new factory line, it can be deployed instantly. It's a "plug-and-play" solution.

2. The "Memory Bank" Strategy

Instead of trying to memorize every single perfect image, the system builds a Memory Bank.

• Imagine you take a photo of a perfect wall plug.
• The system breaks that photo into thousands of tiny puzzle pieces (patches).
• It saves the "essence" of those pieces into a giant library (the Memory Bank).
• When a new product comes down the line, the system breaks it into the same puzzle pieces and asks: "Do I have a perfect match for this piece in my library?"
• If a piece doesn't match anything in the library, it's flagged as weird (an anomaly).

3. The "Overlapping Puzzle" Trick

The original version of this system had a problem: it looked at the product in big, non-overlapping blocks. If a defect happened to sit right on the line between two blocks, the system might miss it or get confused, like trying to read a word that is cut in half by a book binding.

SuperADD fixes this by using overlapping patches. Imagine looking at the product through a window that slides around, but the window is so big it overlaps with the previous view. This ensures that no matter where a defect is, it gets seen clearly from multiple angles, making the system much more reliable.

4. The "Lighting Simulator"

To prepare for the changing factory lights, the system doesn't just look at the training photos as they are. It artificially dims and brightens the images during its setup phase. It's like practicing for a test by studying in a dark room, then a bright room, and then a room with flickering lights. This trains the system to ignore the lighting changes and focus only on the actual shape and texture of the product.

5. The "Morphological Closing" (The Glue)

Sometimes, the system spots a defect, but the result looks like a broken, dotted line instead of a solid scratch. It's like seeing a scratch on a car but only the middle part is highlighted.

To fix this, SuperADD uses a step called Morphological Closing. Think of this as a magical glue. It looks at the broken, dotted highlights and gently connects the dots to form a solid, smooth shape. It also fills in any tiny holes inside the defect area, ensuring the final report shows a complete, clean picture of the problem.

The Results

The system was tested in a tough competition (the CVPR 2026 VAND 4.0 Industrial Track) using a dataset called MVTec AD 2, which includes tricky items like shiny metal cans, transparent jars, and piles of rice.

• The Challenge: The test data had different lighting conditions than the training data, and the system had to work on all different types of objects using the same settings (no custom tuning for each object).
• The Outcome: SuperADD was submitted to the challenge; final rankings have not been announced yet. The paper reports the team's own scores on the challenge data.
  • The system correctly identified defects in Fabric about 88% of the time.
  • It correctly identified defects in Rice about 74% of the time.
  • Most importantly, the approach demonstrates that you can achieve competitive results without needing a complex, custom-trained AI for every single product, even under difficult lighting conditions.

Summary

SuperADD is a smart, flexible, and fast way to spot factory defects without needing to retrain the AI for every new product or lighting change. It uses a pre-trained brain, looks at products in overlapping pieces to avoid missing details, practices with fake lighting changes to stay tough, and uses "glue" to make sure the final defect map is clean and complete. It is a "one-size-fits-all" solution designed to perform robustly across various industrial scenarios.

Technical Summary

Technical Summary: SuperADD – Training-free Class-agnostic Anomaly Segmentation

1. Problem Statement

The paper addresses Visual Anomaly Detection (AD) in industrial inspection, specifically targeting the challenge of distribution shifts caused by varying acquisition conditions (e.g., illumination changes) between training and deployment. The work is situated within the VAND 4.0 Industrial Track, which utilizes the MVTec AD 2 dataset.

Key constraints and challenges include:

• Unsupervised Setting: Models are trained exclusively on normal (defect-free) images.
• Robustness: Models must maintain performance despite significant appearance shifts (lighting, texture variability) between training and test sets.
• Class-Agnostic Requirement: Unlike previous iterations (VAND 3.0) where class-specific architectures or hyperparameters were common, the challenge mandates a single architecture and shared hyperparameter configuration across all object classes to ensure practical deployability and minimal adaptation effort.
• Evaluation: Performance is measured by pixel-level F1 score and AU-ROC on private test splits (TESTpriv and TESTpriv,mix), where ground truth is hidden to prevent overfitting.

2. Methodology

The proposed method, SuperADD, is a training-free pipeline built upon the SuperAD framework, which itself is inspired by PatchCore. It leverages a frozen pre-trained Vision Transformer backbone to extract features and performs nearest-neighbor outlier detection without updating model weights.

2.1. Architecture and Feature Extraction

• Backbone: The authors replace the DINOv2 backbone used in SuperAD with DINOv3 (ViT-H+/16), leveraging its superior pre-trained visual representations.
• Multi-Layer Embeddings: Feature vectors are extracted from four intermediate layers (7, 15, 23, and 31) of the transformer.
• Memory Bank Construction: A memory bank of "normal" prototypes is constructed from the training data.

2.2. Key Technical Modifications

The paper introduces several specific adaptations to enhance robustness and generalization:

• Overlapping Patch-wise Processing:

  • Instead of processing the whole image or non-overlapping tiles, input images are divided into overlapping patches ($P=640$, overlap $O=128$).
  • Purpose: This reduces sensitivity to grid-position artifacts and prevents false anomalies in empty regions or at image borders. It eliminates the need for zero-padding, which can create unrealistic reference embeddings.
  • Inference: Redundant predictions in overlapping regions are discarded, and remaining embeddings are reassembled into a coherent map.

• Refined Subsampling Strategy:

  • Problem: The original SuperAD subsampled 16 images, which failed to remove near-duplicate feature vectors within an image or across similar regions.
  • Solution: The authors perform subsampling directly on feature vectors using a k-nearest-neighbor (k-NN) based approach.
  • Mechanism: For each candidate vector, the number of neighbors within a global distance threshold is calculated. Vectors with low scores (lying in sparsely populated regions of feature space) are retained. This ensures a compact, diverse memory bank that better covers the data distribution while reducing memory usage.

• Intensity-Based Augmentation:

  • During training data processing, pixel values are scaled by a random factor uniformly sampled from $[0.8, 1.2]$.
  • Purpose: To simulate varying integration times and illumination conditions, thereby improving robustness to lighting shifts between training and test data.

• Thresholding and Post-Processing:

  • Thresholding: Instead of class-specific thresholds derived from test data, a single threshold is defined as a scaled version (gain factor 1.3–1.5) of the 95th percentile of anomaly map values from the training data.
  • Morphological Closing: An iterative morphological closing step (16 iterations with line structuring elements of radius 26 pixels at various orientations) is applied to connect fragmented linear defects (e.g., scratches) and close small gaps.
  • Region Filling: A final step fills holes in the binary mask to ensure spatial consistency, particularly where anomalies cross patch boundaries.

3. Key Contributions

The authors claim the following contributions:

1. Class-Agnostic Framework: A unified pipeline using a single architecture and hyperparameters for all object classes, adhering to VAND 4.0 constraints.
2. Improved Subsampling: A feature-space subsampling method that improves data distribution coverage and computational efficiency compared to image-level selection.
3. Patch-wise Preprocessing: The introduction of overlapping patches to mitigate position-dependent artifacts and improve generalization.
4. Robust Post-Processing: The application of iterative, multi-oriented morphological closing to generate spatially consistent anomaly maps.
5. Lighting Robustness: The use of intensity scaling to simulate illumination shifts during training.
6. Backbone Upgrade: The successful integration of DINOv3 as the feature extractor.

4. Results

The method was evaluated on the MVTec AD 2 dataset across three splits: TESTpub, TESTpriv, and TESTpriv,mix.

• Performance Metrics:
  • TESTpub: Achieved a mean F1 score of 62.61% and AU-ROC0.05 of 83.93%.
  • TESTpriv: Achieved a mean F1 score of 57.42%.
  • TESTpriv,mix: Achieved a mean F1 score of 54.35%.
• Comparison:
  • SuperADD outperformed the previous state-of-the-art (ISVL from VAND 3.0), which scored 53.81% on TESTpriv and 51.43% on TESTpriv,mix.
  • It also surpassed other top methods from the previous year (RoBiS, ASEG) and standard baselines like PatchCore and EfficientAD.
• Class-Specific Performance:
  • High performance was observed on Fabric (88.47% F1 on TESTpriv) and Rice (73.83% F1).
  • Performance was lower on Can (0.00% F1 on TESTpub, 11.59% on TESTpriv), attributed to fine defects barely visible to the human eye.
  • Wallplugs showed a significant drop in performance on TESTpriv compared to TESTpub, likely due to more subtle defects and a lower tolerance for false positives in the ground truth.

5. Significance and Claims

The paper positions SuperADD as a practically deployable solution for industrial anomaly detection. Its significance lies in:

• Training-Free Efficiency: By avoiding model retraining, the method allows for rapid integration of new product classes or design changes, a critical requirement in dynamic industrial environments.
• Generalization: The approach demonstrates that a single, class-agnostic configuration can effectively handle diverse object types (bulk goods, textured, reflective, transparent) and varying lighting conditions without per-class tuning.
• Robustness to Distribution Shifts: The combination of DINOv3, intensity augmentation, and patch-wise processing successfully mitigates the performance degradation typically caused by acquisition condition shifts.

The authors acknowledge limitations, such as difficulty in detecting missing parts (e.g., broken pieces) or very thin scratches on reflective surfaces, but emphasize that the method successfully localizes small defects in categories like rice and walnuts and large-scale defects with high coherence. Future work is suggested to explore dual memory banks incorporating synthetic anomalies via diffusion models, though this remains outside the scope of the current training-free claim.