This project demonstrates the key idea from the paper
"Compositional Generative Modeling: A Single Model is Not All You Need"
by training:

• A Shape VAE on grayscale 2D shapes
• A Color Mapper CNN to colorize them
• A Compositional Generator that mixes shape + color

• generate_shapes_dataset.py: creates the toy dataset
• train_shape_vae.py: trains the shape generator
• train_color_mapper.py: trains the color mapper (with edge-aware loss)
• compose_generate.py: composes grayscale + color into new samples
• view_results.ipynb: visualizes progress over epochs

This reproduces the compositional generalization discussed in the paper, showing improved performance over a monolithic VAE.