KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Forward Process
Training phase where Gaussian noise is iteratively added to an original image over multiple time steps, until the image becomes pure noise, serving as the basis for model learning.
Reverse Process (or Denoising Process)
Generation phase where the model learns to reverse the forward process by removing noise step by step, thus reconstructing a clear image from a random noise distribution.
Diffusion Schedule
Strategy defining the variance of the noise added at each time step of the forward process, influencing the quality and speed of generation during the reverse process.
Latent Resampling
Accelerated inference method that skips certain time steps of the denoising process, using interpolation or resampling techniques to reduce the total number of generation steps.
Conditioning (Text-to-Image)
Integration of external information, such as a text description, into the denoising process to guide image generation towards specific content, often via cross-attention mechanisms.
Latent Space (Latent Diffusion)
Approach where the diffusion process is not applied directly to image pixels, but to a compressed representation of the image in a lower-dimensional latent space, thereby reducing computational costs.
Classifier Guidance
Technique that uses a pre-trained classifier to estimate the gradient of a target class (e.g., 'cat'), and adds it to the diffusion model's gradient to steer generation towards that class.
Classifier-Free Guidance
More flexible conditioning method that does not require an external classifier, using a single model trained on both conditioned and unconditioned data to compute a guidance gradient.
Text Embedding
Vector representation of text, generated by a Transformer model (e.g., CLIP), used to condition the diffusion model and translate text semantics into visual signals.
Negative Prompting
Technique involving providing the model with a list of concepts or objects to avoid during generation, modifying the guidance process to minimize the probability of these unwanted elements appearing.
Image-to-Image (Img2Img)
Application of diffusion models where the generation process is initialized not with pure noise, but with an existing image, allowing it to be transformed or stylized while preserving its basic structure.