KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
No Diffusion Timestep
Discrete parameter representing a specific step in the Markov chain of the diffusion process, indicating the level of noise applied to a sample.
Latent Resampling
Technique aimed at improving the quality of the latent space by reorganizing or reweighting latent points for better coverage and more faithful generation.
U-Net Noise Model
Neural network architecture, often U-shaped, specifically designed to predict the noise added at each step of the reverse diffusion process in latent space.
Diffusion Scheduler
Mechanism defining the variance of noise added at each timestep of the forward process, influencing the speed and quality of generation.
Classifier Guidance
Method for conditioning the generation of a diffusion model using the gradient of a pre-trained classifier to guide denoising towards a target class.
Classifier-Free Guidance
Conditioning technique that combines predictions from a conditional and unconditional model to control generation without requiring an external classifier.
Diffusion Model Distillation
Compression process where a large diffusion model (teacher) is used to train a smaller, faster model (student) to perform the same generation task.
Progressive Denoising
Fundamental principle of diffusion models where generation is viewed as a sequence of denoising steps, transforming noise into structured data.
Noise Space
High-dimensional space where Gaussian noise samples are drawn from, serving as the starting point for the reverse diffusion generation process.
Latent Space Interpolation
Operation involving creating smooth transitions between two points in the latent space, thus generating coherent semantic variations between corresponding samples.
Hierarchical Autoencoder
Type of autoencoder with multiple levels of latent spaces, allowing data decomposition at different scales and more controlled generation in latent diffusion models.