ostris/ai-toolkit

Features

• Custom Diffusion Model Training - Offers a comprehensive toolkit for fine-tuning image and video diffusion models using custom datasets.
• Low-Rank Adaptation - Optimizes model training using low-rank adaptation (LoRA) to reduce hardware requirements by updating small weight matrices.
• Diffusion Weight Optimizers - Optimizes weights for image, video, and audio diffusion models to reduce hardware requirements for training.
• Generative Model Fine-Tuning - Fine-tunes image and video diffusion models using custom datasets to achieve specific styles or subjects.
• Multimodal Model Trainers - Functions as a multimodal trainer that orchestrates GPU access and dependencies for fine-tuning generative models.
• LoRA Training - Implements LoRA and LyCORIS techniques to efficiently customize models without full parameter retraining.
• Model Fine-Tuning - Implements specialized techniques to adapt pre-trained models to specific target datasets.
• Image-Text Pair Pipelines - Provides an automated pipeline for organizing image-text pairs with resizing and aspect-ratio bucketing for model training.
• Training Management Dashboards - Provides a management dashboard to monitor and control AI model training experiments.
• Cloud Infrastructure Deployment - Supports deploying training workloads to remote GPU infrastructure using containerized environments.
• GPU Training Clusters - Manages cloud GPU resources and container orchestration for high-memory AI training workloads.
• Training Job Orchestrators - Ships a dashboard and command-line interface to monitor and control training processes on remote GPU infrastructure.
• Container Environment Orchestrators - Provides isolated container environments to manage GPU dependencies and persistent storage for training workloads.
• Containerized Deployment Orchestration - Orchestrates training environments via container definitions to manage GPU access and persistent storage.
• Job Scheduling - Offers a command-line interface and dashboard to control and monitor AI model training jobs.
• Web-Based Control Panels - Provides a web-based control panel to monitor training progress and modify hyperparameters in real-time.
• Aspect Ratio Bucketing - Implements aspect ratio bucketing to minimize padding and preserve image compositions during generative AI training.
• Dataset Preparation Tools - Automates captioning, resizing, and organization of images to curate high-quality machine learning datasets.
• Image Description Generation - Utilizes vision-language models to automatically generate descriptive text captions for image training datasets.
• Kronecker Product Optimizations - Optimizes model weights using the LyCORIS network type through low-rank Kronecker product configurations.
• Layer-Specific Training - Allows restricting training to specific weight groups or excluding layers by filtering for naming patterns.
• Image Captioning - Automatically generates descriptive text for images using multimodal models to create training pairs.
• Multimodal Captioning Tools - Includes a multimodal captioning tool that uses vision-language models to generate descriptive text for training datasets.
• Caption Dataset Utilities - Uses structured JSON files to map image paths to captions, decoupling raw assets from training labels.
• Dataset Preparation Tools - Automates image resizing, aspect ratio bucketing, and folder organization to prepare training datasets.
• Image Generation - Toolkit for training and fine-tuning image models.
• Image Training Tools - Toolkit for training and fine-tuning generative models.

Star history