KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Denoising Diffusion Model
Generative architecture where data is progressively corrupted by adding Gaussian noise, then restored by learning to reverse this diffusion process. These models excel at generating high-quality images by iteratively denoising random samples.
Diffusive Random Walk
Mathematical model describing the evolution of data in latent space through successive accumulation of Gaussian random increments. This formulation allows for analytical characterization of the corruption process and its statistical properties.
Iterative Denoising
Generative procedure applying successive denoising steps over a reversed time horizon to reconstruct data from initial noise. Each iteration progressively refines the quality of the generated sample by eliminating part of the residual noise.
Variational Lower Bound
Variational objective optimized during the training of diffusion models, guaranteeing a lower bound on the data log-likelihood. This formulation enables stable and efficient estimation of the reverse diffusion process parameters.
Langevin Resampling
Sampling technique based on Langevin dynamics using score estimates to explore the target data distribution. This method constitutes the theoretical basis of the denoising process in modern diffusion models.