Papers · DDPM
AI PAPER

Denoising Diffusion Probabilistic Models

Introduced the modern diffusion model framework that generates images by learning to gradually remove noise, enabling high-quality image generation that eventually surpassed GANs.

Authors
Jonathan Ho, Ajay Jain, Pieter Abbeel
Institution
UC Berkeley
Published
June NaN, 2020
Citations
15,000
Category
Architecture
Impact
foundational
PAPER EXPLAINED

Plain-English Summary

Diffusion models generate images through a simple but powerful idea: start with pure random noise, then gradually clean it up step by step until a clear image emerges. The model learns this “denoising” process by studying what happens when you progressively add noise to real images, then learning to reverse each step.

Before diffusion models, the best image generators were GANs (Generative Adversarial Networks), which were notoriously difficult to train and prone to mode collapse (generating only limited varieties of images). Diffusion models proved more stable to train, produced higher quality results, and offered better diversity. Within two years, they powered every major image generation system.

Think of it like a sculptor starting with a rough block of marble and making many small refinements. Each step removes a little “noise” to reveal more of the underlying image. The model learns what refinements to make at each stage.

Key Innovation

The paper showed that a neural network trained to predict and remove noise at each of many small steps could generate high-quality images rivaling GANs. The key insight was connecting diffusion processes (from thermodynamics) with denoising score matching, providing a principled training objective that is stable and scalable.

The forward process adds Gaussian noise over T steps until the image becomes pure noise. The reverse process learns to undo each step. By parameterizing the network to predict the noise (rather than the clean image), training becomes surprisingly simple — just mean squared error between predicted and actual noise.

Impact on the Field

DDPM launched the diffusion model revolution. Within two years, diffusion models went from a research curiosity to powering DALL-E 2, Stable Diffusion, and Midjourney — products used by millions to generate images from text descriptions. The approach proved more scalable and controllable than GANs, enabling conditional generation, inpainting, and editing.

The paper also opened a floodgate of follow-up research. Hundreds of papers improved sampling speed, added conditioning mechanisms, extended to video and 3D, and connected diffusion to other generative frameworks.

Models That Built on This

DALL-E 2 (OpenAI) combined diffusion with CLIP for text-to-image generation. Stable Diffusion (Stability AI) applied diffusion in a compressed latent space for efficiency. Midjourney built a commercial product around diffusion-based generation. Imagen (Google) used large language models as text encoders for diffusion. Sora (OpenAI) extended diffusion to video generation. The technique now underpins virtually all high-quality generative media AI.