This project explores the challenge of generating realistic images from textual descriptions using cutting-edge deep learning methods. By leveraging models like GANs, Diffusion Models, and Transformer-based architectures, we aim to push the boundaries of text-to-image generation, focusing on both qualitative and quantitative evaluations.
- Explore state-of-the-art text-to-image generation techniques.
- Implement and fine-tune models such as GANs, Diffusion models, and Transformer-based models.
- Evaluate and compare generated images qualitatively and quantitatively.
- Demonstrate proficiency in managing computational resources efficiently.
The project employs GAN-based architectures (e.g., DCGAN, StyleGAN) for baseline experiments, and advanced Diffusion models (e.g., Stable Diffusion) and Transformer-based models (e.g., DALL-E inspired). Fine-tuning pretrained models will be the primary approach, with domain-specific subsets to enhance image quality and reduce computational costs.
- Qualitative: Visual inspection, side-by-side comparisons, and optional human surveys.
- Quantitative: Fréchet Inception Distance (FID), Inception Score (IS), and CLIP-based metrics.
- Performance and statistical analysis will guide model improvements and comparisons.
- COCO Captions
- CUB-200 Birds
- Oxford-102 Flowers
- (Optional) A custom annotated dataset may be created for domain-specific experimentation.