• Built a DDPM with cosine noise scheduling and a U-Net–based noise predictor, implementing both forward noising and reverse denoising processes.
• Trained and sampled images via iterative reverse diffusion and achieving FID <160 after ~2 hours of training on a T4 GPU.