Transform your personal photos into customized art posters using Neural Style Transfer (NST). This project blends content and style images with tunable artistic intensity, offering a cost-effective alternative to generic wall art.
This system applies NST using a pre-trained VGG-19 model to combine the content of one image with the artistic style of another. Users can upload content and style images, control the style intensity via an alpha parameter (style_weight), and download the final stylized result.
Key Features:
- Upload content and style images (from URL or local path)
- Control style intensity (1e5 – 5e7)
- Choose optimizer: Adam or LBFGS
- Track loss and visualize intermediate results
- Save final stylized image in JPG format
The project is modularized into three main classes:
- ImageLoader: Preprocesses input images
- VGGFeatures: Extracts content and style features from VGG-19 layers
- StyleTransfer: Implements NST training logic and visualizations
- Model: Pre-trained VGG-19 on ImageNet
- Framework: PyTorch
- Content Layer:
conv4_2 - Style Layers:
conv1_1,conv2_1,conv3_1,conv4_1,conv5_1 - Loss Functions:
- Content Loss (MSE)
- Style Loss (Gram matrix + MSE)
- Optimizers:
- Adam (faster, slightly noisier)
- LBFGS (slower, smoother)
Also, our model successfully swaps content and style roles to create flexible transformations
We compared performance across optimizers and different alpha values:
| Optimizer | Style Weight | Total Loss | Content Loss | Style Loss |
|---|---|---|---|---|
| Adam | 1e5 | 0.9546 | 0.4408 | 5.14e-6 |
| Adam | 5e7 | 25.7103 | 4.4712 | 4.2e-7 |
| LBFGS | 1e5 | 0.9546 | 0.4408 | 5.14e-6 |
| LBFGS | 5e7 | 25.7103 | 4.4712 | 4.2e-7 |
git clone https://github.com/yourusername/neural-style-transfer.git
cd neural-style-transfer
pip install -r requirements.txt