Diffusion Models

Status: public · Confidence: medium (0.835) · Basis: verified_sources

## TL;DR

Diffusion models generate data by learning to reverse a noising process. The central idea is to corrupt training examples with noise, then train a model to denoise step by step. Later variants made the approach practical for high-resolution generation by moving the process into latent spaces or by framing it with score-based stochastic differential equations.

## Core Claims

DDPM made diffusion practical as a generative-model recipe: add noise during training, learn the reverse process, then generate from noise by iteratively denoising.

Latent diffusion changes the operating space. Instead of running every denoising step directly over pixels, it works over compressed latent representations and then decodes the final latent back into an image.

Score-based SDE work gives a broader mathematical view of the family, connecting diffusion-like sampling with score functions and continuous-time stochastic processes.

## Citation Boundaries

Use this article for stable model-family facts. Do not use it as evidence for the current best image, video, audio, or 3D generation system; those comparisons change quickly and require current evaluations.

## Further Reading

- [Denoising Diffusion Probabilistic Models](https://arxiv.org/abs/2006.11239)
- [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752)
- [Score-Based Generative Modeling through Stochastic Differential Equations](https://arxiv.org/abs/2011.13456)