AI-ordlista
Den kompletta ordlistan över AI
Noise Scheduling
Strategy defining how noise is progressively added to data across different time steps in diffusion models. Scheduling controls the variance of the noise added at each timestep, influencing the quality and stability of the generation process.
Score Network
Neural network trained to predict the gradient of the data log-density as a function of noise level and current state. The architecture must be time-conditional to capture the evolving dynamics of the diffusion process.
Diffusion Time Steps
Discrete number of steps used in the discretization of the continuous diffusion process for training and inference. The choice of the number of timesteps influences the trade-off between generation quality and computational cost.
Gaussian Noise Perturbation
Addition of noise following a normal distribution to progressively transform data into a simple distribution. This controlled perturbation ensures that the process remains mathematically tractable and reversible.
Variance Schedule
Sequence of variances defining the intensity of noise added at each step of the forward diffusion process. The schedule can be linear, cosine, or other forms to optimize noise progression.
Noise Conditional Score Network
Network architecture taking as input both the noisy data and the noise level (timestamp) to predict the score. This conditionality is essential for managing different noise scales in the diffusion process.