akanimax/T2F

Combines LSTM text encoding with progressive GAN training to synthesize realistic facial images from natural language descriptions, using the Face2Text dataset of 400 images with captions. The architecture feeds LSTM-encoded text embeddings through a conditioning augmentation layer into ProGAN's generator, while employing layer-by-fade-in training for stable multi-resolution generation. Implemented in PyTorch with modular components for data processing, network definitions, and configurable hyperparameters for progressive depth expansion from 64×64 to higher resolutions.

547 stars. No commits in the last 6 months.