---

title: Pipeline Overview description: How the CIA pipeline works : from real images to augmented training data through generation, evaluation, and filtering. keywords: pipeline, workflow, steps

---

Pipeline Overview

The CIA pipeline follows a linear flow from real images to augmented training data. Each step can be run independently or as part of the full pipeline.

┌─────────────┐    ┌─────────────┐    ┌──────────────────┐    ┌─────────────┐
│ Real Images  │───►│  Extract     │───►│  SD + ControlNet │───►│  Generated   │
│              │    │  Condition   │    │  Generation      │    │  Images      │
└─────────────┘    └─────────────┘    └──────────────────┘    └──────┬──────┘
                                                                         │
┌─────────────┐    ┌─────────────┐    ┌──────────────────┐             │
│ Real Images  │───►│  Feature     │───►│  Quality Metrics │◄────────────┘
│              │    │  Extraction  │    │  (FID, IS, MLD)  │
└─────────────┘    └─────────────┘    └────────┬─────────┘
                                                  │
                                           ┌──────▼──────┐
                                           │  Filtering   │
                                           │  (top-k, etc)│
                                           └──────┬──────┘
                                                  │
┌─────────────┐    ┌─────────────┐              │
│ Real Images  │───►│    Mix       │◄─────────────┘
│              │    │  Real+Synth  │
└─────────────┘    └──────┬──────┘
                          │
                   ┌──────▼──────┐
                   │    Train     │
                   │  Downstream  │
                   │   Model      │
                   └─────────────┘

Step 1: Generate

Extract a control condition from each real image using one of the available extractors, then generate synthetic variations using Stable Diffusion + ControlNet.

Key decisions:

• Which extractor to use (canny, openpose, segmentation, mediapipe_face)
• Which Stable Diffusion and ControlNet models from HuggingFace
• How many synthetic images per real image
• What prompts to use (fixed, from captions, or vocabulary-modified)

Output: Generated images in data/generated/{dataset}/{controlnet}/

Step 2: Evaluate (DTD + PTD)

Compute quality metrics comparing the real and synthetic distributions.

• DTD (Distribution-To-Distribution): FID, Inception Score : measure overall distribution similarity
• PTD (Point-To-Distribution): Mahalanobis distance : scores each individual synthetic image

Output: Metric scores saved to metadata.yaml alongside generated images

Step 3: Filter

Select the best synthetic images based on PTD scores.

• top-k: Keep the k images with smallest distances
• top-p: Keep the top proportion (0 ≤ p ≤ 1) of images
• threshold: Keep images with distance below a threshold

Output: Filtering results appended to metadata.yaml

Step 4: Mix

Combine real and filtered synthetic data into a training-ready dataset (YOLO format or CSV format).

Step 5: Train

Train downstream models (YOLOv8 for object detection, InceptionV3 for classification) using the mixed dataset.

---

Steps 1–3 are the core library functionality. Steps 4–5 are provided as example scripts in the examples/ directory.