How dhruvnet/synthetic-image-generation-for-ai-training works
Project Type and Purpose
• This is a research‐oriented, open-source synthetic image generation pipeline implemented as Jupyter notebooks and Python modules.
• Purpose: Generate diverse, high-resolution images under varying conditions (day/night, weather effects) to augment AI training datasets, with built-in safeguards against adversarial/ deep-fake threats.
File Structure Analysis
• Top-level directories:
• Algorithm/ – Python modules or helper scripts for each generative model
• COMPARISiON/ – Evaluation scripts and output comparisons between models
• FINAL OP/ – Final output artifacts (generated images, notebooks)
• Notebooks:
• day-night-image-translations-using-cyclegan.ipynb – CycleGAN pipeline
• Train_ConditionalVAE_FINAL.ipynb – CVAE training for weather simulation
• FYP_dataset_building.ipynb – Orchestrator notebook, dataset creation, UI control
• Config/build: No explicit CI/CD or deployment scripts, this is a local/ research pipeline.
README Insights
• Core Modules: CycleGAN, CVAE, Real-ESRGAN, and a User Interface module.
• Data sources: Separate datasets for CycleGAN (~30k images) and CVAE (~89k images).
• Evaluation: Reconstruction and latent loss metrics; comparison against baselines (DALL-E2, RunwayML, Artbreeder).
• Tech Stack: Python (PyTorch/TensorFlow), OpenCV, Matplotlib/Seaborn, Jupyter.
a. Main Components
1. Data Ingestion & Preprocessing
– CycleGAN dataset loader
– CVAE attribute dataset loader
– Image augmentation & normalization (OpenCV)
2. Generative Modules
– CycleGAN module (day↔night)
– CVAE module (fog/rain/snow simulation)
– Real-ESRGAN module (super-resolution)
3. Control Interface
– Notebook-based UI to set intensity, select models, tune parameters
4. Security/Guard Module (optional)
– Adversarial‐attack detection or privacy filters
5. Evaluation & Comparison
– Metrics calculator (reconstruction, total, latent loss)
– Comparison scripts vs. baselines
6. Output & Storage
– Final images (FINAL OP/)
– Logs & plots (COMPARISiON/)
b. Relationships & Data Flow
• Feedback loop from Evaluation → Control Interface for parameter tuning
c. Architectural Patterns
• Modular pipeline: each generative component is a replaceable service
• Directed acyclic data flow (no circular dependencies)
• Notebook orchestration as a “controller” layer
• Separation of concerns: data, compute, evaluation, UI
d. Key Technologies
• PyTorch / TensorFlow code modules (Algorithm/)
• OpenCV preprocessing scripts
• Real-ESRGAN library integration
• Jupyter Notebook as UI & orchestrator
• Matplotlib & Seaborn for visualizations