AI Art Generation Toolkit

A PyTorch implementation of creative AI models for artistic image generation and style translation.

Overview

This repository contains implementations of three groundbreaking deep learning approaches for artistic image generation:

• Neural Style Transfer: Blend content and style from different images
• Pix2Pix: Paired image-to-image translation using conditional GANs
• CycleGAN: Unpaired image translation with cycle consistency

What's Inside

Neural Style Transfer

Combine the content of one image with the artistic style of another using a pre-trained CNN. The magic happens through three key components:

Content Preservation

• Matches high-level content features between source and output
• Uses intermediate CNN layer activations to capture object structure

Style Transfer

• Gram matrix computation captures texture and style patterns
• Multiple layers capture different brushstroke scales
• Shallower layers = finer details, deeper layers = broader patterns

Visual Smoothing

• Total variation loss reduces noise and artifacts
• Encourages spatial coherence in generated images

Pix2Pix - Paired Image Translation

For when you have matched input-output pairs, Pix2Pix uses conditional GANs to learn the mapping:

Generator Architecture

• U-Net with skip connections to preserve spatial information
• Encoder-decoder with bottleneck features
• Skip connections shuttle low-level details directly

Discriminator

• PatchGAN classifier operating on image patches
• Focuses on high-frequency structure
• More efficient and effective than full-image discrimination

Training Strategy

• Combines adversarial loss with L1 pixel-wise loss
• Instance normalization instead of batch normalization
• Adam optimizer with carefully tuned parameters

CycleGAN - Unpaired Translation

When you don't have paired training data, CycleGAN learns bidirectional translations:

Key Innovation

• Two generators: X→Y and Y→X
• Two discriminators for each domain
• Cycle consistency ensures meaningful translations

Cycle Consistency

• Forward cycle: x → G(x) → F(G(x)) ≈ x
• Backward cycle: y → F(y) → G(F(y)) ≈ y
• Prevents mode collapse and ensures semantic preservation

Architecture Details

• Generator: ResNet blocks with instance normalization
• Discriminator: 70×70 PatchGAN
• Identity loss for color preservation in some domains

Quick Start

# Style Transfer
python run_style_transfer.py --content image.jpg --style painting.jpg

# Pix2Pix Training
python train_pix2pix.py --dataset facades --direction BtoA

# CycleGAN Training  
python train_cyclegan.py --dataset horse2zebra

Technical Highlights

PyTorch Lightning Integration

• Reproducible experiments
• Mixed precision training support
• Hardware scalability (CPU/GPU/TPU)
• Clean separation of research and engineering code

Optimization Advances

• Least squares GAN loss for stable training
• Historical generated images buffer for reduced oscillations
• Multiple normalization strategies (instance/batch/layer)
• Adaptive learning rate scheduling

Results Gallery

The models excel at various creative tasks:

• Style Transfer: Photo → Van Gogh, Picasso, etc.
• Pix2Pix: Architectural labels → facades, maps → satellite
• CycleGAN: Horses ↔ zebras, photos → Monet paintings

Requirements

torch>=1.7
pytorch-lightning>=1.0
torchvision
pillow
numpy

License

MIT License - feel free to use these models for your creative projects!