Kairos: Text-to-Image DCGAN

Kairos is a PyTorch research prototype that turns short text prompts into 64 × 64 images. It pairs GLOVE language embeddings with a DCGAN trained on Microsoft COCO, creating an end-to-end baseline for studying text conditioning, image generation, and model evaluation.

A complete training and inference pipeline for generating RGB images from captions. The project includes dataset loading, model checkpoints, sample generation, loss plots, training GIFs, and side-by-side comparisons of real and generated images.

Each caption is averaged into a 300-dimensional GLOVE representation and projected to 1,024 dimensions. The generator combines that text signal with a 100-dimensional noise vector; the discriminator evaluates the resulting image alongside a spatial text embedding.

The model trained for 70 epochs with batches of 512, Adam at a 0.0002 learning rate, and resumable checkpoints every 10 epochs. The full run took approximately 24 hours on GPU hardware.

The recorded run produced an FID of 322.30, an Inception Score of 0.53, a text-image matching score of -0.15 ± 0.40, and a CLIP score of 0.70. These results establish a reproducible baseline and make the model's visual quality and prompt alignment measurable rather than anecdotal.

The clearest path forward is a stronger text encoder such as CLIP or BERT, followed by higher-resolution generation and comparison with diffusion or StyleGAN-based approaches. Human evaluation would add a useful qualitative counterpoint to the automated scores.