ForgeGAN

Industrial Defect Synthesis & Generative Anomaly Detection

ForgeGAN is an open, reproducible research and engineering project for studying whether controlled synthetic defect generation can improve industrial visual anomaly detection.

Status: industrial dataset pipeline. VisA and MVTec AD read-only adapters and validation commands are available. The original 2023 Keras notebooks are preserved under src/ as historical experiments; the modern implementation is being built independently with PyTorch.

Why ForgeGAN?

Industrial inspection datasets are often imbalanced: normal samples are abundant, while defect examples can be scarce and diverse. ForgeGAN explores a measurable workflow:

1. Train generative models on industrial imagery.
2. Generate controlled synthetic defect examples.
3. Evaluate image quality and diversity.
4. Train anomaly-detection baselines with real-only and augmented data.
5. Compare results under reproducible protocols.

Synthetic data is treated as an experimental intervention—not an assumed improvement. Every benchmark must include real-only baselines and disclose the exact data split.

Planned capabilities

• Modular PyTorch training engine (initial DCGAN baseline implemented)
• DCGAN, conditional GAN, ACGAN and WGAN-GP baselines
• Mask-conditioned defect synthesis
• Dataset adapters for MVTec AD and VisA (dataset access and licensing notes in DATASET_STRATEGY.md)
• Real-only vs. synthetic-augmented anomaly-detection experiments
• Quantitative evaluation, experiment configs and reproducible reports
• Optional diffusion baseline and Gradio demo

Repository status

The repository is currently in its foundation stage. Planned features are not presented as implemented. See ROADMAP.md and CONTRIBUTING.md.

Quick start

Requirements: Python 3.10–3.12.

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\\Scripts\\activate
python -m pip install -e ".[dev]"
# Add PyTorch when installing the training components:
python -m pip install -e ".[train]"
forgegan --help
forgegan audit-dataset --root "src/Dataset" --output reports/legacy-dataset-audit.json
forgegan validate-visa --root datasets/visa --split-file datasets/visa/split_csv/1cls.csv
forgegan validate-mvtec --root datasets/mvtec_ad --category bottle
# After obtaining the dataset archive:
forgegan train-dcgan --config configs/experiments/mvtec-bottle-dcgan.yaml
# Category-conditioned generation needs the complete multi-category dataset:
forgegan train-cgan --config configs/experiments/mvtec-all-categories-cgan.yaml
# Procedural defect generation uses normal train images only:
forgegan synthesize-defects --dataset mvtec --root datasets/mvtec_ad \
  --category bottle --output-dir runs/procedural-bottle --kind spot --count 32
# Train-only threshold calibration vs normal-only threshold on held-out test:
forgegan evaluate-synthetic --dataset mvtec --root datasets/mvtec_ad \
  --category bottle --synthetic-manifest runs/procedural-bottle/manifest.json \
  --output reports/bottle-threshold-comparison.json
pytest

The CLI provides dataset audits and validation. The shared Python dataset registry, spatial transforms and optional PyTorch DataLoader are documented in DATA_PIPELINE.md. Dataset inspection and versioned experiment configs are covered in EXPERIMENT_TOOLING.md. The adapters do not download or modify source datasets. Unconditional DCGAN and normal-image category-conditional GAN training are available; a procedural defect-synthesis baseline is available; learned mask-conditioned defect synthesis and downstream evaluation remain planned. See exploratory downstream evaluation and DCGAN training guide and conditional GAN guide, and procedural synthesis guide.

Project principles

• Reproducibility: explicit seeds, configs, versioned splits and saved metadata.
• Honest evaluation: no synthetic samples may leak across evaluation splits.
• Modular design: datasets, generators, trainers and metrics remain replaceable.
• Responsible data use: follow each dataset's license and terms.
• Evidence over claims: report negative and neutral results as well as gains.

Legacy experiments

The original DCGAN, ACGAN and Wasserstein GAN notebooks are retained in src/ for provenance. They are not yet part of the new training API and may require legacy dependencies and local-path cleanup.

License

MIT for the project code. Dataset licenses and terms remain separate and must be respected.