Denoising Diffusion Probabilistic Models (DDPM)

Connected diffusion models with score matching by parameterizing the reverse process as a simple weighted noise-prediction objective with a U-Net backbone.

The breakthrough paper proving that diffusion models surpass GANs in sample quality, establishing the modern generative diffusion pipeline.

Assumptions

• Simplified variational bound objective ignoring weighting terms produces vastly superior visual quality compared to exact negative log-likelihood

Limitations

• Slow ancestral sampling requires hundreds to thousands of sequential forward passes per generated image