KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Domain Translation
Process of converting images from one visual domain to another while preserving the original semantic content, such as transforming photos into paintings or vice versa.
Cycle Consistency
Fundamental constraint of CycleGAN ensuring that an image translated and then re-translated back to its original domain should be identical to the initial image, thus guaranteeing content preservation.
Unpaired Data
Dataset where source and target domain images have no direct correspondences, allowing model training on much larger and more easily accessible corpora.
Dual Discriminator
CycleGAN architecture using two separate discriminators, each specialized in detecting authenticity for a specific domain, improving bidirectional translation quality.
Forward-Backward Cycle
CycleGAN mechanism where the image undergoes two successive translations: from domain A to B, then from B to A, forming a closed cycle to evaluate content preservation.
Unsupervised Domain Adaptation
Learning technique where the model learns to map between domains without labels or correspondences, relying solely on the inherent structure of the data.
Content-Preserving Style Transfer
CycleGAN's ability to modify the visual appearance (texture, color) of an image while maintaining its fundamental semantic structure, avoiding content distortions.
Bijective Mapping
Mathematical transformation implemented by CycleGAN establishing a reversible correspondence between the data distributions of the two domains, ensuring faithful translation in both directions.
Conditional Image Synthesis
Generation of new images based on conditional input, where CycleGAN uses the source image as a condition to produce an image in the target domain.