Akegarasu/lora-scripts
Lora Scripts
lora-scripts es un toolkit de ajuste fino (fine-tuning) diseñado para adaptar modelos de difusión base a estilos o sujetos específicos. Proporciona un conjunto especializado de scripts y herramientas para ejecutar trabajos de entrenamiento de adaptación de bajo rango (LoRA) y Dreambooth.
El proyecto cuenta con una interfaz gráfica basada en web que gestiona el flujo de trabajo de entrenamiento, permitiendo a los usuarios configurar y ejecutar trabajos sin editar scripts manualmente. Esta interfaz mapea las entradas del usuario a hiperparámetros y proporciona un panel en tiempo real para monitorear métricas de entrenamiento y curvas de pérdida para rastrear la convergencia del modelo.
El sistema incluye un gestor de etiquetado de datasets para organizar y editar etiquetas de imágenes. Para garantizar una ejecución consistente en diferentes hosts de hardware, el entorno de entrenamiento se proporciona como una configuración contenedorizada preconfigurada.
Features
- Custom Stable Diffusion Training - Provides a complete workflow for training custom Stable Diffusion models to recognize specific entities or styles.
- LoRA Fine-Tuning Tools - Ships a comprehensive set of scripts and tools for low-rank adaptation (LoRA) of diffusion models.
- Diffusion Model Training - Implements specialized fine-tuning logic for diffusion models via standalone Python scripts.
- Diffusion Model LoRA Fine-Tuning - Provides a web interface for configuring and executing LoRA and Dreambooth fine-tuning jobs for diffusion models.
- Model Training Management Interfaces - Features a web-based dashboard for configuring hyperparameters and triggering training runs without manual scripting.
- Web-Based Command Interfaces - Translates user inputs from a graphical dashboard into command-line arguments for training scripts.
- Automated Image Tagging - Automates the labeling of image datasets using specialized vision models and editing tools.
- Automated Processing Pipelines - Provides an automated pipeline to process raw images and generate structured metadata labels for model training.
- Dataset Tagging Managers - Includes a dedicated tagging manager for organizing and editing image labels used in training.
- Image Classification Datasets - Organizes and structures image datasets with labels for training image recognition and generation models.
- Model Training Monitoring - Tracks and visualizes model convergence and loss curves in real time to monitor training quality.
- Dataset Labeling Interfaces - Provides integrated tools for tagging and editing dataset labels to improve training accuracy.
- Training Progress Monitors - Provides a real-time dashboard for visualizing training metrics and loss curves to track model convergence.
- Item Tagging and Organization - Includes built-in editors and automated processes for organizing and managing image tags in datasets.
- Real-Time Metric Visualization - Features a web dashboard that renders live convergence curves by polling training logs and loss data.
