KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Diffusion Transformer
Hybrid architecture integrating multi-head attention mechanisms into the iterative diffusion process to enhance the overall coherence of generated data.
U-ViT
Variant of Vision Transformer where U-Net connections are integrated to effectively combine multi-scale features in diffusion models.
DiT (Diffusion Transformer)
Architecture replacing traditional U-Net convolutions with Transformer blocks in the diffusion process, using time embeddings for conditionality.
Latent Diffusion Transformer
Model applying Transformer mechanisms in compressed latent space, reducing computational complexity while preserving generative quality.
Cross-Attention Diffusion
Mechanism allowing diffusion models to align with external conditions via cross-attention layers between noise and conditional embeddings.
Transformer Denoiser
Transformer-based module responsible for predicting noise at each denoising step in the forward-backward diffusion process.
Patch Diffusion
Technique where data is divided into patches processed by Transformer attention mechanisms before the iterative diffusion process.
Adaptive Layer Normalization
Normalization method conditioned by time embeddings in Diffusion-Transformer architectures to stabilize training.
Self-Attention Noise Prediction
Use of self-attention to model long-distance dependencies in noise prediction during the diffusion process.
Transformer Score Matching
Application of Transformer architectures to estimate the log-density gradient (score) in score-based diffusion models.
Multi-Scale Transformer Diffusion
Hierarchical approach using Transformers at different scales to capture both fine details and global structure in generation.
Conditional Diffusion Transformer
Architecture integrating conditions (text, images, classes) through attention mechanisms in the Transformer diffusion process.
Rotary Position Embedding in Diffusion
Positional encoding technique applied to Transformer diffusion models to better capture spatial relationships in structured data.
Diffusion-Guided Transformer
Model where the diffusion process guides the Transformer's attention to improve coherence and quality of structured generations.
Sparse Transformer Diffusion
Variant using sparse attention mechanisms to reduce computational complexity in high-resolution diffusion models.
Transformer Latent Space Diffusion
Diffusion process applied in the latent space learned by a Transformer autoencoder for efficient generation of structured data.
Diffusion-Aware Self-Attention
Modified self-attention mechanism that accounts for the current noise level in the iterative diffusion process.
Hierarchical Transformer Diffusion
Multi-level architecture where Transformers progressively generate increasingly refined representations through diffusion.