Glosarium AI
Kamus lengkap Kecerdasan Buatan
Custom Embedding
Numerical vector file trained via Textual Inversion, capturing the semantic characteristics of a specific visual concept, which can be used in the prompt to influence image generation.
LyCORIS Weights
Generalized fine-tuning weight format extending LoRA, capable of applying low-rank modifications to any part of the model (including U-Net layers), offering superior flexibility for adaptation.
Concept Fine-tuning
Training process that adjusts the entire diffusion model on a targeted dataset so that it masters a specific subject, style, or object, at the cost of greater loss of generalization.
Semantic-Preservation Regularization
Technique used during fine-tuning (notably with DreamBooth) to prevent overfitting and loss of the model's ability to generate other concepts, by using varied regularization images.
Fine-tuning Checkpoint
Complete model file (often several gigabytes) resulting from concept fine-tuning, containing all modified weights of the diffusion network, replacing or combining with the base model.
Control Adapter (ControlNet)
Conditioning system that adds a trainable auxiliary neural network to precisely control image generation from spatial inputs such as sketches, poses, or depth maps.
Class-Guided Diffusion Fine-tuning
Variant of fine-tuning where the model is conditioned not only by text but also by class labels, enabling more granular control over the attributes of generated objects.
Custom Diffusion Model
Diffusion model that has been specifically adapted, via techniques such as DreamBooth or LoRA, to excel in generating a unique style, character, or visual universe.
Diffusion Learning on Style
Application of fine-tuning where the objective is to teach the diffusion model a particular artistic style (e.g., watercolor, cyberpunk) by training it exclusively on representative images of that style.
Fine-tuning Weight Fusion
Mathematical process of combining multiple sets of fine-tuning weights (e.g., multiple LoRAs) to create a hybrid effect, adjusting their respective influence ratios.
Reference Prompt for Fine-tuning
Descriptive text used during fine-tuning training to associate training images with a textual concept, serving as a bridge between visual data and the model's embedding space.
Weight Quantization for Fine-tuning
Technique for reducing the numerical precision of a fine-tuning model's weights (e.g., from FP32 to FP16 or INT8) to decrease file size and memory usage, often at the cost of slight quality loss.
Low-Shot Fine-tuning
Fine-tuning challenge that consists of adapting a model with a very limited number of training examples, requiring techniques like DreamBooth or Textual Inversion to be effective.
Catastrophic Forgetting Degradation
Phenomenon where a diffusion model, after intensive fine-tuning on a concept, forgets how to generate other concepts it previously mastered, reducing its versatility.