Code repository for our IEEE IES Generative AI Challenge 2026 submission on lightweight defect synthesis.
Team NanoGen | IEEE IES Generative AI Challenge 2026
This repository contains the code and experimental notebooks for EdgeDiff-Industrial, a resource-efficient generative AI pipeline designed to synthesize high-fidelity industrial defect data. This project was developed as a submission for the IEEE IES Generative AI Challenge 2026 and has been recommended for submission to the IRAI 2026 conference.
The core objective of this project is to augment industrial datasets (like the MVTec AD dataset) to improve downstream anomaly detection, while drastically reducing the computational overhead typically required by diffusion models.
- Structural Conditioning: Utilizes Canny edge extraction and ControlNet to ensure synthesized defects maintain precise physical geometry.
- Hardware Optimization: Implements INT8 quantization and Latent Consistency Models (LCM), reducing denoising steps from 50 to 4.
- Resource Efficiency: Achieved a 52% reduction in memory footprint (from 5.2 GB to 2.5 GB), making the model highly viable for resource-constrained edge server environments.
- Performance Boost: Synthesized data successfully improved the downstream classification AUROC from 0.80 to 0.98 for specific manufactured textures (e.g., Bottles).
The project is divided into three main Jupyter Notebooks:
01_Data_Preprocessing.ipynb: Handles the ingestion of the MVTec anomaly detection dataset, image resizing, and Canny edge map extraction.02_Model_Training_and_Quantization.ipynb: Contains the core generative pipeline, including Stable Diffusion 1.5, ControlNet integration, LCM scheduler setup, and INT8 quantization implementation.03_Evaluation_Metrics.ipynb: Runs the downstream anomaly detection classifier (ResNet/EfficientNet) on both baseline and augmented data to calculate Accuracy, F1-Scores, and AUROC.
| Model Architecture | Precision | Denoising Steps | Memory Size (MB) |
|---|---|---|---|
| Baseline (SD 1.5) | FP16 | 50 | 5230.62 |
| EdgeDiff (Ours) | INT8 | 4 | 2506.61 |
Note: While our augmented data significantly improved detection accuracy for manufactured items (Bottles, Metal Nuts), we observed performance regressions in highly complex organic textures (Hazelnuts), highlighting a known domain-shift challenge in generative models.
- Clone this repository:
git clone [https://github.com/YourUsername/EdgeDiff-IEEE-GenAI-Hackathon-2026.git](https://github.com/YourUsername/EdgeDiff-IEEE-GenAI-Hackathon-2026.git)