Image-to-Image Translation with DDPM

This repository contains implementations of Image-to-Image translation tasks using Denoising Diffusion Probabilistic Models (DDPM). We explore two approaches:

1. Dual Diffusion Implicit Bridges (DDIB)
2. Custom Latent Space Transforms

Notebooks

• ddib_implementation.ipynb: Implementation of the DDIB approach
• latent_transforms.ipynb: Our custom latent space transformation method

Results

Our latent transformation approach achieves:

• FID score: 42.559 (vs DDIB's 99.614)
• Average color difference: 0.271 (comparable to DDIB's 0.270)

Dataset

The experiments were conducted on colored MNIST digits (specifically digits '2' and '3').