Conditional Face Generation using CVAE

This project explores conditional face generation with a Conditional Variational Autoencoder (CVAE) trained on CelebA. The model learns to generate and edit face images based on a vector of binary facial attributes.

What the project does

• Generates new faces from a latent vector and attribute set
• Reconstructs CelebA faces from the encoder-decoder pipeline
• Edits selected attributes such as Male, Eyeglasses, Blond_Hair, and Smiling
• Compares different beta values to show the reconstruction vs. generation trade-off

Main files

• main.py - experiment entry point, training, generation, and reconstruction export
• src/model.py - CVAE and the alternative fully connected CVAE2 model
• src/train.py - training loop, beta warm-up, and evaluation helpers
• src/utils.py - generation, reconstruction, editing, and interpolation utilities
• projekt.ipynb - results, experiments, analysis, and figures used in the presentation

Dataset

The project uses the CelebA dataset with 40 binary attributes. Images are loaded through src/dataset.py, resized for training, and normalized to the [-1, 1] range.

Current setup

The current configuration is stored in config.yaml and controls:

• latent size
• image resolution
• beta value
• batch size
• number of epochs
• training subset size

main.py reads the config file, trains the model, saves the checkpoint, and exports generated samples and reconstructions to experiments/.

Experiment outputs

The repository already contains saved experiment folders with:

• model checkpoints
• training histories
• reconstruction figures
• generated samples

These artifacts are used by the notebook and presentation slides.

How to run

Install dependencies:

pip install -r requirements.txt

Run training and export the latest experiment outputs:

python main.py

The run will create a folder under experiments/ with the current config name and save:

• model.pth
• history.json
• generated face grids
• reconstruction plots

Notebook

projekt.ipynb is the analysis notebook. It collects:

• reconstruction examples
• attribute interpolation results
• attribute editing examples
• latent-space experiments
• beta trade-off observations

Limitations

• Some attributes are still weakly controlled
• Fine facial details are blurred
• Rare attribute combinations are harder to generate