AI Toolkit

Ostris's all-in-one training suite for image, video and audio diffusion models, with a web interface and a command line.

  • Fine-tuning
  • MIT · Open source
  • macOS, Windows, Linux
  • Verified

AI Toolkit is an all-in-one training suite for diffusion models, maintained by Ostris. Its author says it aims to support the latest image and video models on consumer-grade hardware, and it runs either as a web interface or from the command line. The interface at localhost port 8675 starts, stops and monitors training jobs, and it can be protected with an auth token. The jobs themselves keep running without the interface.

Its supported-model list is long. The README includes FLUX.1, FLUX.2 and FLUX.3 Image Klein, Qwen-Image 2.1, Z-Image, Krea 2, HiDream O1, Ideogram 4 in FP8, Ming-Image, MOSS-Image and FIBO 2 for images, plus instruction-edit variants, Wan 2.1 and 2.2, LTX-2.5 and MiniMax H3 for video, and ACE-Step 1.5 and YuE2 for audio. Several of those models were released within the last few months.

The recommended install is its built-in AI Toolkit Manager, which detects your hardware, sets up the matching PyTorch build and keeps everything inside the folder. The manager is still marked experimental, and the Mac script is labeled Apple Silicon, experimental. The manual instructions require an Nvidia GPU with enough memory for the model you train.

Good for

  • Training LoRAs for new open image and video models with a point-and-click interface
  • Running the same training job from a script on a headless server
  • Keeping up with models that the older trainers do not support yet

Watch out for

  • Manual installation needs an Nvidia GPU, and Mac support through the manager is experimental
  • Memory needs depend on the model and settings, so check the README's notes for each model before starting
  • Respect the license of the base model and of your training data, because the MIT license only covers the toolkit

Links