AI-ordlista
Den kompletta ordlistan över AI
Classifier Guidance
Technique using a pre-trained classifier to guide the diffusion process toward specific attributes or classes. Guidance is applied via the classifier's gradient during denoising to precisely control generation.
Conditional Diffusion
Extension of diffusion models that incorporates conditions or labels to generate samples meeting specific criteria. Conditioning can be applied at each step of the diffusion or reverse process.
Diffusion Scheduler
Strategy defining the evolution of noise variance through the forward diffusion steps. The scheduler influences the quality and convergence speed of the reverse generation process.
Energy-Based Diffusion
Formulation of diffusion models as energy-based models where training minimizes an energy function. Allows natural integration with gradient guidance techniques.
Iterative Denoising
Progressive refinement process where each iteration applies a slight correction to reduce residual noise. Diffusion models typically use hundreds to thousands of iterations for high-quality generation.
Guidance Interpolation
Technique allowing the combination of multiple guidance vectors to generate smooth transitions between different concepts or attributes. Useful for creating controlled variations in the generation space.
Diffusion Architecture
Specific neural structure designed for the diffusion process, typically based on U-Nets with attention mechanisms. The architecture must efficiently handle spatial and temporal dependencies of noise.