Definition
Generative models that recover data structure from noise by learning a stepwise denoising or reverse-diffusion process.
Research in 2015 applied forward diffusion and a learnable reverse process to generative modeling. The 2020 paper on denoising diffusion probabilistic models demonstrated high-quality image synthesis.
This entry focuses on diffusion-based generative methods; not every denoising algorithm belongs to this technology. Language models and diffusion models follow different modeling approaches, and they can also be combined within larger systems.
Principles and impact
Problems solved
- Modeled and generated complex data distributions through a trainable, stepwise process.
Principles
- Construct a forward perturbation process and learn a reverse process that recovers data.
Improvements
Not recorded
Significance
- Established an important approach within modern generative modeling.
History and applications
People
Jascha Sohl-Dickstein、Jonathan Ho、Pieter Abbeel
Organizations
Not recorded
Events
Not recorded
Applications
Image generation、Generative modeling
Limitations
- Generation costs depend on the model and the number of sampling steps; data bias and content consistency still require evaluation.
Upstream and downstream relations
Upstream
Sources: Deep Unsupervised Learning using Nonequilibrium ThermodynamicsDenoising Diffusion Probabilistic Models
Sources: Deep Unsupervised Learning using Nonequilibrium ThermodynamicsDenoising Diffusion Probabilistic Models
Downstream
No direct downstream relations recorded