Diffusion model
An AI that creates pictures by starting with a screen of static, like a snowy TV, and slowly refining that mess into a clear, detailed image.
A generative model that learns to reverse a diffusion process, where it is trained to iteratively remove Gaussian noise from an image until a coherent output emerges from random input.
A generative model that learns to reverse a gradual noising process, iteratively denoising random noise into a coherent sample; the basis of most modern image and video generators.
evolution
- 2015 · historyDeep Unsupervised Learning using Nonequilibrium Thermodynamics
Sohl-Dickstein et al. introduced the theoretical framework for diffusion-based generative modeling.
- 2020 · historyDenoising Diffusion Probabilistic Models (DDPM)
Ho et al. demonstrated that diffusion models could produce high-quality image synthesis, rivaling GANs.
- 2021 · historyDiffusion Models Beat GANs on Image Synthesis
Dhariwal and Nichol introduced classifier guidance, significantly improving sample quality and training stability.
- 2022 · historyLatent Diffusion Models (Stable Diffusion)
Rombach et al. enabled high-resolution image generation by performing the diffusion process in a compressed latent space.