CIA: Controllable Image Augmentation¶

Generate high-fidelity synthetic images using Stable Diffusion + ControlNet. A complete pipeline for data augmentation with built-in quality metrics.

Get Started View on GitHub

Why CIA?¶

Feature	Description
Controlled Generation	Use edge maps, pose estimation, segmentation, or face meshes as control signals
Quality-Aware	Evaluate with FID, Inception Score, and Mahalanobis distance
Smart Filtering	Select best samples using PTD scores (top-k, top-p, threshold)
Multiple Interfaces	Python API, CLI, or Hydra config, pick what fits your workflow

The Pipeline¶

Real Images → Extract Condition → SD + ControlNet → Generated Images
     ↓                                              ↓
  Features ←── Quality Metrics (FID, IS, MLD) ──→ Filtered Set
                                                    ↓
                                              Mixed Dataset → Train

Quick Example¶

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",
)

# Evaluate quality
scores = evaluate(
    real="data/real/train/images/",
    generated="data/generated/",
    metrics=["fid", "mld"],
)

# Filter best images
kept = filter_generated(
    generated="data/generated/",
    method="top-k",
    value=1000,
)

Installation¶

pip install -e .

Supported Extractors¶

Extractor	Use Case
`canny`	Edge detection
`openpose`	Human pose estimation
`segmentation`	Instance segmentation
`mediapipe_face`	Face mesh landmarks

Supported Metrics¶

Metric	Type	Description
`fid`	DTD	Fréchet Inception Distance
`inception_score`	DTD	Inception Score
`mld`	PTD	Mahalanobis Distance

Questions? Open an issue on GitHub.