KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Classifier-Free Guidance
Guidance technique that eliminates the need for an external classifier by using the model itself to follow or ignore a conditioning signal, thus improving prompt fidelity.
IP-Adapter (Image Prompt Adapter)
Module that allows using a reference image as a prompt, encoding its visual characteristics to guide the generation process without modifying the diffusion model's weights.
GLIGEN (Grounded Language-to-Image Generation)
Framework that anchors bounding boxes to text concepts, enabling precise spatial control over the position and size of generated objects in an image.
Cross-Attention Guidance
Mechanism that leverages cross-attention maps between text and image to strengthen or weaken the influence of specific parts of the prompt on areas of the generated image.
LoRA (Low-Rank Adaptation)
Efficient fine-tuning technique that injects small low-rank matrices into the model's attention layers, allowing learning of new styles or concepts with minimal parameters.
T2I-Adapter (Text-to-Image Adapter)
Lightweight module that guides a pre-trained diffusion model using additional reference conditions (such as sketches or segmentation maps) without requiring full retraining.
UNet Guidance
Guidance process that operates directly on the denoising predictions of the UNet architecture, modifying the gradient direction to align generation with desired conditioning.
Attention Refocusing
Method that modifies attention maps to redirect the importance of a prompt token to another region of the image, allowing reorganization of generated elements' composition.
Semantic Guidance
Approach that uses an external semantic model (e.g., CLIP) to compute a guidance loss, ensuring that the generated image remains aligned with the prompt's meaning at a conceptual level.
Multi-Modal Guidance
Strategy that combines multiple types of guidance (text, image, mask, depth map) simultaneously for hierarchical and multi-faceted control over the generation process.
Null-Text Inversion
Image editing technique that inverts the generation process to find an optimized 'null prompt', enabling precise modifications on real images without artifacts.
SDEdit (Stochastic Differential Equation Edit)
Method that adds noise to an existing image then denoises it with guidance, allowing editing of real images while preserving their underlying structure.