CIA: Controllable Image Augmentation

CIA is a Python library for synthetic data augmentation using Stable Diffusion + ControlNet. Generate high-quality synthetic images from real seed images, evaluate their quality, and use them to improve downstream ML models.

Features

• Synthetic image generation using Stable Diffusion controlled by Canny edges, OpenPose, Segmentation, or MediaPipe face features
• Quality metrics -- Frechet Inception Distance (FID), Inception Score (IS), Mahalanobis distance
• Quality-based filtering -- keep only the best synthetic images via top-k, top-p, or threshold filtering
• Auto-captioning -- generate image captions using OpenAI or Ollama vision models
• Multiple interfaces -- Python API, CLI, and Hydra config

📖 Citation

If you use ciagen in your research or project, please consider citing our work:

CIA: Controllable Image Augmentation Framework Based on Stable Diffusion (IEEE MIPR 2024) > M. Benkedadra et al., "CIA: Controllable Image Augmentation Framework Based on Stable Diffusion," 2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR), 2024. DOI: 10.1109/MIPR62202.2024.00102.

@inproceedings{benkedadra2024cia,
  title={CIA: Controllable Image Augmentation Framework Based on Stable Diffusion},
  author={Benkedadra, Mohamed and Rimez, Dany and Godelaine, Tiffanie and Chidambaram, Natarajan and Khosroshahi, Hamed Razavi and Tellez, Horacio and Mancas, Matei and Macq, Benoit and Mahmoudi, Sidi Ahmed},
  booktitle={2024 IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR)},
  year={2024},
  organization={IEEE},
  doi={10.1109/MIPR62202.2024.00102}
}

From Unity Simulation to Diffusion-Based Augmentation: Quantifying Dataset Balance for Robust Object Detection > M. Benkedadra, A. Saoudi, M. Gloesener, S. A. Mahmoudi, and M. Mancas, "From Unity Simulation to Diffusion-Based Augmentation: Quantifying Dataset Balance for Robust Object Detection," Proceedings of the 2025 8th International Conference on Computational Intelligence and Intelligent Systems, pp. 29–36, 2025. DOI: 10.1145/3787256.3787261.

@inproceedings{benkedadra2025unity,
  title={From Unity Simulation to Diffusion-Based Augmentation: Quantifying Dataset Balance for Robust Object Detection},
  author={Benkedadra, Mohamed and Saoudi, Aissa and Gloesener, Maxime and Mahmoudi, Sidi Ahmed and Mancas, Matei},
  booktitle={Proceedings of the 2025 8th International Conference on Computational Intelligence and Intelligent Systems},
  pages={29--36},
  year={2025},
  publisher={ACM},
  doi={10.1145/3787256.3787261}
}

Try it now

Run CIA in your browser with Google Colab: no installation required. Open the Quickstart notebook to generate, evaluate, and filter synthetic images in under 15 minutes.

Installation

pip install ciagen

With optional dependencies:

pip install ciagen[captioning]    # OpenAI/Ollama auto-captioning
pip install ciagen[training]     # YOLO/classifier training
pip install ciagen[datasets]      # COCO, Flickr30K, FER, MOCS datasets
pip install ciagen[all]           # Everything

Development

git clone https://github.com/fennecinspace/ciagen.git
cd ciagen
pip install -e ".[all]"

Docker

./run_and_build_docker_file.sh nvidia
docker exec -it ciagen zsh

Quick Start

Python API

from ciagen import generate, evaluate, filter_generated

# Generate synthetic images
result = generate(
    source="data/real/train/images/",
    output="data/generated/",
    extractor="canny",
    sd_model="fennecinspace/sd-v15",
    cn_model="lllyasviel/sd-controlnet-canny",
    num_per_image=3,
    prompt="a person walking in a park",
    seed=42,
    device="cuda",
)
print(f"Generated {result['total_generated']} images")

# Evaluate quality
scores = evaluate(
    real="data/real/train/images/",
    generated="data/generated/",
    metrics=["fid", "mld"],
    feature_extractor="vit",
)
print(f"FID: {scores['dtd']['fid']}")

# Filter to keep the best images
kept = filter_generated(
    generated="data/generated/",
    method="top-k",
    value=100,
)

CLI

# Generate images
ciagen generate \
    --source data/real/train/images/ \
    --output data/generated/ \
    --extractor canny \
    --sd-model fennecinspace/sd-v15 \
    --cn-model lllyasviel/sd-controlnet-canny \
    --num 3 \
    --prompt "a person walking"

# Evaluate quality
ciagen evaluate \
    --real data/real/train/images/ \
    --generated data/generated/ \
    --metrics fid mld

# Filter generated images
ciagen filter \
    --generated data/generated/ \
    --method top-k \
    --value 100

# Auto-caption images
ciagen caption \
    --images data/real/train/images/ \
    --output data/real/train/captions/ \
    --engine ollama \
    --model llava

Hydra (Advanced)

python run.py task=gen model.cn_use=lllyasviel_canny prompt.base="a person"
python run.py task=dtd
python run.py task=ptd
python run.py task=filtering
python run.py task=mix
python run.py task=train

See ciagen/conf/config.yaml for all configuration options.

Pipeline

The recommended workflow:

real images ──► condition extraction ──► SD + ControlNet ──► synthetic images
                                                              │
real images ──────────────────────────────────────────────► evaluate ──► filter ──► mix ──► train

1. Generate -- Extract a control condition (edges, pose, segmentation) from each real image, then generate synthetic variations using Stable Diffusion + ControlNet
2. Evaluate -- Compute distribution-level metrics (FID, IS) and per-image metrics (Mahalanobis distance)
3. Filter -- Select the best synthetic images based on quality scores
4. Mix -- Combine real and filtered synthetic data into a training dataset
5. Train -- Train your downstream model (YOLOv8 for detection, InceptionV3 for classification)

Available Extractors

Extractor	Description	Use Case
canny	Canny edge detection	General purpose, preserves structure
openpose	Human pose estimation	People, actions, body pose
segmentation	YOLOv8 semantic segmentation	Object boundaries
mediapipe_face	MediaPipe face landmarks	Facial emotion, face generation

Available Metrics

Metric	Type	Description
fid	Distribution-to-Distribution	Frechet Inception Distance -- lower is better
inception_score	Distribution-to-Distribution	Inception Score -- higher is better
mld	Point-to-Distribution	Mahalanobis distance -- per-image, lower is better

Data Structure

data/
├── real/{dataset}/
│   ├── train/{images,labels,captions}/
│   ├── val/{images,labels,captions}/
│   └── test/{images,labels,captions}/
├── generated/{dataset}/{controlnet-model}/
│   ├── metadata.yaml
│   └── *.png
└── mixed/{dataset}/

Example Datasets

python run.py task=prepare_data data.base=coco       # COCO People
python run.py task=prepare_data data.base=flickr30k   # Flickr30K Entities
python run.py task=prepare_data data.base=fer         # Facial Emotion Recognition
python run.py task=prepare_data data.base=mocs        # Construction Sites

Documentation

Full documentation is available in the docs/ directory and can be built with MkDocs:

pip install mkdocs-material mkdocstrings[python]
mkdocs serve

Contributing

See CONTRIBUTING.md for development setup, code style, and PR guidelines.

License

This project is licensed under the GNU Affero General Public License v3.

Copyright (c) 2026 Universite de Mons, Multitel, Universite Libre de Bruxelles, Universite Catholique de Louvain.