A Generation Algorithm for Text to Image

This project focuses on developing a Text-to-Image generation system using the Stable Diffusion 1.5 model fine-tuned with the CUB-200 dataset. The primary objective is to generate high-quality images from textual descriptions. Stable Diffusion, a latent diffusion model, is employed in this system to create visually realistic images from textual inputs. The model leverages key components such as CLIP Text Encoder, UNet, and VAE (Variational Autoencoder) for text understanding, image generation, and latent space encoding, respectively. LoRA (Low-Rank Adaptation) is used for efficient fine-tuning of the pre-trained model, adapting it to the specific task of bird image generation from textual prompts. This system provides options for image generation on CPU, fast CPU, and GPU, allowing users to choose based on their available computational resources. The project is developed using the Flask framework, with SQLite for storing user data, and a front-end built with HTML, CSS, and JavaScript. The evaluation of the generated images is carried out using FID (Fréchet Inception Distance) and Inception Score, two standard metrics for assessing the quality of generated images. This approach offers significant advancements in generative models, specifically in the field of text-to-image synthesis for specialized datasets like CUB-200.

Keywords:

Text-to-image generation, Stable Diffusion, CLIP Text Encoder, UNet, VAE, LoRA, Flask, SQLite, Image generation, CUB-200.