Text-Conditioned Image Synthesis with Latent Diffusion Models (LDM)

This project demonstrates how to train a text-conditioned Latent Diffusion Model (LDM) on the CelebA-Dialog dataset. It uses pretrained encoders and a custom diffusion model to generate face images conditioned on natural language descriptions.

For a detailed explanation of the methodology, training setup, and results, check out the full report:
📄 Read the Report (PDF)

---

Results

Latents and Reconstructions

Latent Representation	Reconstructed Image

---

Text-Conditioned Generation

The model generates images by denoising in the latent space based on a text prompt and decoding through a pretrained VAE.

---

Model Components

• Latent Encoder (VAE): Pretrained CompVis/stable-diffusion-v1-4
• Text Encoder: Pretrained CLIP (openai/clip-vit-large-patch14)
• Diffusion Model: Trained from scratch on CelebA-Dialog latents and captions

---

📦 Preprocessed Data

To speed up training and evaluation, you can use our preprocessed datasets:

• Aligned and captioned images from CelebA-Dialog
  👉 Kaggle Dataset

• Precomputed latents and tokenized captions
  👉 Hugging Face Dataset