This project implements a text-to-image generation pipeline inspired by the Stable Diffusion architecture. The pipeline was built entirely from scratch in PyTorch. It integrates a Variational Autoencoder (VAE) for latent space compression, Denoising Diffusion Probabilistic Models (DDPM) for iterative denoising, and CLIP-based text embeddings for aligning text and images effectively.
• Custom Variational Autoencoder (VAE): Compresses images into latent representations for efficient generation.
• DDPM Sampling: Implements iterative denoising to generate high-quality images from noise.
• Text Embedding with CLIP: Ensures precise alignment of text and generated images.
| Resource | Description |
|---|---|
| Tokenizer | Tokenizer files for Stable Diffusion. |
| Model Repository | Main Hugging Face repository for Stable Diffusion v1.5. |
| Research Paper | Original Stable Diffusion research paper. |