From dataset to trained model.

Train image, video, and audio diffusion models in one studio.
A guided web UI when you want it. YAML and the CLI when you need them.

Get started · Models · Documentation · Report a bug

Set up a training run	Generate with your models

• Prepare your data. Organize datasets, edit captions, and use optional dataset encryption.
• Train your way. Run LoRA, LoKr, or supported full fine-tunes, locally or on remote workers.
• Review and create. Track runs, compare samples, generate images, and export jobs from one library.

AITK Studio is a maintained fork of Ostris AI Toolkit, with additional model integrations and workflows. Some behavior differs from upstream.

Quick start

For local NVIDIA setups, you'll need Python 3.12, Git, and Node.js 22.12+ (22.x), 24.x, or 26.x for the web UI. GPU memory requirements depend on the model and training settings.

The default installer uses CUDA 13.0 and requires a compatible NVIDIA driver. See installation profiles if you need CUDA 12.8 compatibility.

Expand your platform to install the runtime and launch the UI:

Linux

git clone https://github.com/BadAtCaptchas/AITK-Studio.git
cd AITK-Studio
python3.12 -m venv venv
source venv/bin/activate
python scripts/install_runtime.py
cd ui
npm run build_and_start

Windows · PowerShell

For automatic setup, updates, and launch, double-click run_windows.bat in your checkout, or run .\run_windows.bat from PowerShell. The launcher uses the current manager to provision Python, PyTorch, Node.js, and other dependencies. Local changes cause the code update to be skipped; dependency setup and launch still run.

For manual setup:

git clone https://github.com/BadAtCaptchas/AITK-Studio.git
cd AITK-Studio
py -3.12 -m venv venv
.\venv\Scripts\Activate.ps1
python scripts/install_runtime.py
cd ui
npm run build_and_start

Open localhost:8675. The launch command installs UI dependencies, builds the app, and starts the managed services.

Other setups: Apple silicon (experimental) · DGX / Spark · Docker · RunPod · Modal

Hosting on a server or connecting over your LAN? Follow the network access and authentication setup before exposing the UI.

Your first run

1. Add a dataset in Datasets and review its captions.
2. Create a training job, choose a model and dataset, then review the settings and start it.
3. Check progress and samples, then use Try this model to load a checkpoint in Generate.

Prefer the terminal? After installing the Python runtime, copy and edit a model example, then run it from the repository root:

python run.py config/your_training_config.yml

See the training reference for configuration and resume behavior.

Supported models

• Image & editing: FLUX, Qwen Image / 2.1, Z-Image, SDXL, Ideogram 4, HiDream, Krea 2, and more.
• Video: Wan 2.1 / 2.2, LTX-2 / 2.3 / 2.5, and MiniMax H3.
• Audio: Ace Step 1.5 and 1.5 XL, plus experimental YuE2; MiniMax H3 also supports joint video/audio workflows.
• Multimodal text: Qwen2.5-Omni instruction tuning with image, audio, or video inputs.

Browse the full model catalog → for model links and experimental variants. Supported training modes and memory needs vary by model.

Documentation

I want to…	Read
Install or troubleshoot the runtime	Installation
Work with datasets, captions, generation, and job exports	User guide
Configure LoKr, training phases, watermarking, or VRAM savings	Training reference
Set up remote workers, cloud training, storage, or TensorBoard	Deployment and operations
Upgrade an installation or recover interrupted work	Execution and recovery · Workspace upgrades
Develop or test changes to the project	Development utilities

Help and credits

Report a reproducible bug with steps, environment details, logs, and your version or commit. This repository's issue tracker is for bugs in this fork; setup help, usage questions, and feature requests are outside its scope. Report upstream-only bugs to the original project.

Built on Ostris AI Toolkit. Released under the MIT license; preserve the original license and attribution when redistributing.