Skip to contentExperiment
Generating by denoising
Distinguish corruption during learning from progressive generation.
Essentials
How it works
In DDPM, a forward process progressively adds noise to data. A network learns information for reversing this corruption. Sampling begins with noise and applies denoising steps to generate a new sample.
Takeaway
The network, objective and sampling schedule are separate choices. Data and compute also affect quality. Denoising does not necessarily recover a memorised original image.
From concept to practicePut it into practice
Scientific sources
- Ho, Jain & Abbeel · 2020Denoising Diffusion Probabilistic ModelsOriginal publication
Teaching synthesis of the cited sources. Published results remain tied to their tasks, protocols and budgets.