Style-Transfer-Pipeline

Neural style transfer with quantitative evaluation — built for production-readiness.

Applies the artistic style of any reference image to a content image using a VGG-19 perceptual loss network. Includes a full evaluation suite (SSIM, PSNR) and a clean CLI for batch processing.

---

Quick start

git clone https://github.com/YOUR_USERNAME/style-transfer-pipeline.git
cd style-transfer-pipeline
pip install -r requirements.txt

python src/train.py \
  --content data/inputs/photo.jpg \
  --style   data/styles/vangogh.jpg \
  --output  data/outputs/result.jpg \
  --steps   500 \
  --size    512

Runs on CPU or GPU automatically.

---

Evaluate output quality

python evaluation/metrics.py \
  --content   data/inputs/photo.jpg \
  --style     data/styles/vangogh.jpg \
  --generated data/outputs/result.jpg

{
  "ssim_vs_content": 0.7812,
  "psnr_vs_content": 24.31
}

---

Architecture

Content image ──┐
                ├──► VGG-19 feature extractor ──► Perceptual loss ──► Adam optimizer ──► Generated image
Style image  ───┘         (frozen weights)         content + style

Loss function:

L_total = α × L_content + β × L_style

L_content = MSE(F_generated, F_content)         # conv_4 features
L_style   = Σ MSE(G_generated_l, G_style_l)     # Gram matrices at conv_1–5

where G denotes the Gram matrix.

---

Project structure

style-transfer-pipeline/
├── src/
│   ├── model.py        # VGG-19 feature extractor + loss functions
│   ├── train.py        # Optimisation loop + CLI
│   └── utils.py        # Image I/O and preprocessing
├── evaluation/
│   └── metrics.py      # SSIM, PSNR evaluation
├── data/
│   ├── inputs/         # Content images
│   ├── styles/         # Style reference images
│   └── outputs/        # Generated results (gitignored)
├── requirements.txt
└── README.md

---

Key parameters

Flag	Default	Effect
--steps	500	More steps = stronger stylisation
--style-weight	1e6	Higher = more artistic, less content
--content-weight	1.0	Higher = more faithful to original
--size	512	Output resolution (px)
--lr	0.01	Adam learning rate

---

Skills demonstrated

ML competency	Implementation
Deep learning (CV)	VGG-19 perceptual loss network
Model evaluation	SSIM + PSNR quantitative metrics
PyTorch proficiency	Custom forward pass, Gram matrices
Production mindset	CLI, GPU/CPU auto-detect, logging
Generative modelling	Iterative image synthesis via gradient descent

---

Roadmap

• Fast style transfer (feed-forward network, ~1000× faster)
• Multi-style blending
• REST API endpoint (FastAPI)
• Batch processing script

---

References

• Gatys et al., A Neural Algorithm of Artistic Style (2015)
• Johnson et al., Perceptual Losses for Real-Time Style Transfer (2016)

---

License

MIT