Skip to content
Get packs

Guide · Open source

How to use ai-toolkit: install, run the GUI and train models

Updated · written with AI from the project's own documentation and checked against our catalogue data

AI Toolkit is a free, open-source suite from ostris for training diffusion models, the kind of AI that generates images, video and audio. It can run as a web GUI or from the command line. This guide explains what you need before you start, how to install it from the README, and what it is suited for, so you can decide whether to finetune a model on your own data.

AI Toolkitostris/ai-toolkit · ★ 12.5k · MIT

Quick start

  1. 1Check your hardware and tools. Confirm you have an Nvidia GPU, git and Python 3.10 or newer (3.12 recommended).
  2. 2Clone the repository. Run git clone https://github.com/ostris/ai-toolkit.git and enter the ai-toolkit folder.
  3. 3Start the launcher. On Linux, make run_linux.sh executable with chmod +x and run ./run_linux.sh. Use run_mac.zsh or run_windows.bat on other platforms.
  4. 4Open the UI. Once the manager starts, open http://localhost:8675 in your browser.
  5. 5Pick a model and dataset. Choose a supported model from the README list and prepare your own training images or videos.
  6. 6Train and check results. Start a training job in the UI or from the CLI, then use the doctor command if setup fails.

What AI Toolkit is for and who it suits

AI Toolkit is a training tool, not an image generator. Its README describes it as an all-in-one training suite for diffusion models, with a focus on supporting the latest models on consumer-grade hardware. It covers image, video, audio and some LLM models.

It fits people who want to finetune a model such as FLUX, SDXL or Qwen-Image on their own data, and developers experimenting with video models such as Wan 2.2 or LTX-2. It is less suited to people who only want to generate images with existing models, or to anyone without an Nvidia GPU.

  • Good fit: finetuning an image model on your own data with a single consumer Nvidia GPU
  • Good fit: wanting a GUI for training but also the option of a CLI
  • Not a fit: only generating images with existing models
  • Not a fit: no Nvidia GPU, since the manual instructions list one as a requirement

What you need before installing

The README lists these requirements for the manual install: Python 3.10 or newer (3.12 recommended), an Nvidia GPU with enough memory for the work you plan to do, Python venv, and git.

The setup involves git, Python environments and PyTorch. If you have no Python or command-line experience, expect this to be the hardest part.

  • git
  • Python 3.10 or newer (3.12 recommended)
  • Nvidia GPU with enough RAM for your task
  • Python venv

Install AI Toolkit with the Manager (experimental)

The README recommends the AI Toolkit Manager, which is built into the repo. The manager detects your hardware, sets up the right PyTorch build, creates the Python environment, and keeps everything inside the ai-toolkit folder rather than installing system-wide. On each launch it checks for updates and skips them with a warning if you have modified files.

The manager is marked experimental, so the README also keeps the manual instructions for people who prefer them. Start by cloning the repo and entering the folder, then run the launcher for your platform. On Linux, the README uses these commands:

  • Linux launcher: run_linux.sh (x86_64 and ARM64)
  • MacOS (Apple Silicon, experimental): run_mac.zsh
  • Windows: double-click run_windows.bat or run it from a terminal
git clone https://github.com/ostris/ai-toolkit.git cd ai-toolkit chmod +x run_linux.sh ./run_linux.sh

Open the web UI and use the manager from the terminal

The manager starts the UI at http://localhost:8675. The README also documents terminal commands for headless servers, which cover first-time setup, updating, launching and diagnosing problems.

  • install: first-time setup
  • update: pull updates and sync dependencies
  • launch: start the UI
  • doctor: diagnose problems
python3 -m manager install python3 -m manager update python3 -m manager launch python3 -m manager doctor

Install manually if the manager does not work for you

If you prefer manual setup, or the manager runs into problems, the README gives a Linux path: clone the repo, create a virtual environment, activate it, and install PyTorch before the rest of the dependencies. The README's version pins specific PyTorch releases, so copy those from the README at the time you install rather than from this guide.

  • Create a venv: python3 -m venv venv
  • Activate it: source venv/bin/activate
  • Install torch first, then the remaining requirements
git clone https://github.com/ostris/ai-toolkit.git cd ai-toolkit python3 -m venv venv source venv/bin/activate

Know which models you can train

The README lists supported models by category: image models such as FLUX.1-dev, FLUX.2-dev, Qwen-Image, SDXL and SD 1.5; instruction and edit models such as FLUX.1-Kontext-dev and Qwen-Image-Edit; video models such as Wan 2.1, Wan 2.2 and LTX-2; audio models such as Ace Step 1.5; and an LLM entry, Qwen2.5-Omni. Some models are listed as experimental.

Model downloads come from Hugging Face, so check each model's licence on its page before training or sharing results.

  • Image: FLUX.1-dev, FLUX.2-dev, Qwen-Image, SDXL, SD 1.5
  • Edit: FLUX.1-Kontext-dev, Qwen-Image-Edit
  • Video: Wan 2.2 14B, Wan 2.2 TI2V 5B, LTX-2
  • Audio: Ace Step 1.5

What people use it for

  • Finetuning an SDXL or FLUX image model on your own image set using a single consumer Nvidia GPU
  • Training an edit model such as Qwen-Image-Edit when your dataset includes control images
  • Experimenting with video models such as Wan 2.2 or LTX-2 on local hardware
  • Running training jobs on a headless Linux server through the manager's terminal commands

Frequently asked questions

Is ai-toolkit free?

Yes. The README describes AI Toolkit as free and open source, and the repository is licensed under MIT. You still need your own hardware and must follow the licence of each model you train.

What are the requirements for ai-toolkit?

The manual install lists git, Python 3.10 or newer (3.12 recommended), an Nvidia GPU with enough RAM for your task, and Python venv. The manager's only stated requirement is git.

Does ai-toolkit work without an Nvidia GPU?

The README lists an Nvidia GPU as a requirement and does not describe other hardware paths, so it should not be assumed to work on AMD, Intel or Apple hardware. The manager does include an experimental macOS launcher for Apple Silicon.

How do I install ai-toolkit on Linux?

Clone the repository with git, enter the ai-toolkit folder, make run_linux.sh executable and run it. The launcher is the README's recommended route, and manual venv steps are available if it fails.

Is ai-toolkit good for beginners?

It is designed to be easy to use, but the setup involves git, Python environments and PyTorch. Beginners with no command-line experience may find installation difficult.

How does ai-toolkit compare to kohya-ss sd-scripts or OneTrainer?

The comparable projects listed for this directory include kohya-ss/sd-scripts, kohya-ss/musubi-tuner, OneTrainer and SimpleTuner. AI Toolkit's distinguishing feature here is that it offers both a GUI and a CLI. Check each project's README to compare the models and features you need.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

AI Toolkit suits people with a single consumer Nvidia GPU who want to finetune image or video models locally and are comfortable with a little setup. If you only want to generate images, lack an Nvidia GPU, or have no command-line experience, pick a simpler tool or a hosted service.

Install options & downloads