KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Denoising Diffusion Probabilistic Model (DDPM)
Generative architecture that learns to reverse a Gaussian diffusion process, by progressively adding noise to data and then training a network to predict the added noise to reconstruct the original.
Variance Schedule
Predefined series of variance coefficients (β_t) that control the amount of noise added at each timestep of the forward process, directly influencing the diffusion trajectory.
Timestep
Discrete integer representing a specific step in the Markov chain of the diffusion process, ranging from clean data (t=0) to pure noise (t=T).
Noise Prediction Network (U-Net)
Neural network architecture, typically a U-Net, used in DDPMs to predict the noise added to data at a given timestep, conditioned on that timestep.
Langevin Sampling
Stochastic optimization method that can be used to approximate the denoising process, using score gradients to guide generation.
Simplified Denoising Objective
DDPM loss function that simplifies training by requiring the model to directly predict the added noise, rather than the mean or covariance of the denoising distribution.
Resampling
Inference technique where multiple denoising trajectories are explored in parallel to improve the quality and diversity of generated samples.
Conditioning
Mechanism allowing to guide the generation process by providing additional information to the model, such as text, an image, or a class, often integrated via embeddings.
Guided Inference
Sampling strategy that modifies the denoising process to bias generation towards desired attributes, using an external classifier (Classifier-Free Guidance) or the score gradient.
Denoising Step
A single iteration of the reverse process where the model predicts noise and an update is applied to move from a noisy state x_t to a slightly less noisy state x_{t-1}.
Diffusion Constant
Numerical value derived from the variance schedule, used to parameterize the evolution of data through time steps, ensuring stable convergence towards a Gaussian distribution.
Stochastic Diffusion Equation
Stochastic differential equation that formally describes the continuous evolution of data under the effect of noise, of which the discrete DDPM process is a discretization.