Checking session...

Overview

About Anomalia

Anomalia is an AI-powered image analysis application that demonstrates how modern vision-language models (VLMs) can detect anomalies in images — often without task-specific training.

Originally developed as part of applied research collaborations at Mila, the project was inspired by real-world challenges in infrastructure inspection, where anomalies are rare, diverse, and difficult to model using traditional supervised approaches.

From research to demo

In collaboration with Hydro-Québec, Mila researchers explored how to detect defects in power line infrastructure using AI. One key insight emerged:

Recent vision-language models can identify anomalies out of the box — even in domains they were not trained on.

These models were able to:

  • Detect subtle defects (scratches, broken strands, contamination)
  • Generalize across different cable types and conditions
  • Perform competitively with little to no labeled training data

Anomalia was created to make this capability tangible — allowing users to upload their own images and explore how far these models can go in real-world scenarios.

What Anomalia demonstrates

Rather than being a production system, Anomalia is a hands-on exploration tool designed to showcase key capabilities of modern AI:

  • Zero-shot anomaly detection — identify unusual patterns without prior domain-specific training
  • Natural language + vision reasoning — models interpret images using general world knowledge
  • Bounding box localization — highlight and describe potential defects or regions of interest
  • Baseline comparison workflows — compare against "normal" images to improve signal detection

Why it matters

Many industrial and operational contexts share the same constraints:

  • Anomalies are rare
  • Data is limited or expensive to label
  • Conditions vary widely

Anomalia explores a different paradigm — where general-purpose AI models reduce the need for specialized datasets, enabling faster prototyping and broader applicability.

A MilaHub project

Anomalia is part of MilaHub, a platform for showcasing applied AI systems and experiments. It is designed to:

  • Demonstrate real capabilities (not mockups)
  • Encourage exploration and experimentation
  • Bridge research insights with practical applications

Media

Anomalia analysis results: a broken strand on a cable highlighted with a bounding box and label.
Sample project (hydroelectrical cables): vision-language model output with defect localization and description.

Tags

Anomaly DetectionComputer VisionGenerative AI
Sign in to launchBack to home