ai-toolkit is an all-in-one training suite for fine-tuning diffusion models on consumer-grade hardware. It supports image, video, and audio diffusion models and can be operated through a web GUI or command-line interface.
Project overview
It targets fine-tuning the latest diffusion model architectures on consumer-grade hardware, and its dual web GUI and CLI modes make training jobs accessible across different workflows.
Deployment
Refer to project documentation
License
MIT
Best for
Creators, AI engineers, and developers who want to fine-tune diffusion models on consumer-grade hardware.
Key capabilities
A web interface to start, stop, and monitor jobs and train models with a few clicks.
Run training jobs via python run.py using YAML config files.
Train LoRA models with specific layer targeting.
Train LoKr models using specific network configurations.
Automatically downscales and places images in buckets for batching, handling varying aspect ratios without manual cropping.
An experimental manager that detects hardware, sets up PyTorch, creates the python environment, and grabs local copies of Node.js and FFmpeg. It is still experimental and may have issues.
Limitations and risks
Currently, the only supported dataset formats are jpg, jpeg, and png. Webp currently has issues.
Pressing Ctrl+C while the toolkit is saving a checkpoint will likely corrupt that checkpoint.
The AI Toolkit Manager is still experimental and may have issues.
Getting started
Clone the repository, create and activate a Python venv, install torch and requirements, then run a training config or start the UI. The experimental AI Toolkit Manager can alternatively detect hardware and set up the environment.