FaceGen v1 - 128px DCGAN

This model is a Deep Convolutional Generative Adversarial Network (DCGAN) trained to generate high-quality 128x128 images of human faces. It was trained for 250 epochs on a curated dataset of feline images, pushing the boundaries of traditional GAN architectures at this resolution.

Sample

Here's a sample after epoch 200:

Model Details

• Architecture: DCGAN (Deep Convolutional GAN)
• Resolution: 128x128 pixels (RGB)
• Parameters: ~186M (Generator)
• Training Duration: ~22 hours on NVIDIA RTX 5060 Ti 16GB
• Framework: PyTorch with Mixed Precision (AMP)

Training Hyperparameters

• Batch Size: 128
• Learning Rate: 0.0002
• Optimizer: Adam (Beta1: 0.5, Beta2: 0.999)
• Latent Vector (Z): 128 dimensions

Training details

The full training code can be found as train.py in this repo. The training data we used was from HF: stable-bias/faces

Intended Use

This model is intended for artistic and research purposes. It demonstrates how GANs can capture complex faces and even eye reflections at medium resolutions.

How to use

To use this model, clone this repository and run the provided inference script. Ensure you have matplotlib, torch and torchvision installed.

python3 inference.py

Sample output: