AI Toolkit

Ostris built AI Toolkit as an all-purpose trainer for diffusion models, and it has become a standard choice for FLUX LoRA training in the open model community. It supports LoRA and LoKr fine-tuning across more than 25 image architectures including FLUX, SDXL, and SD 1.5, plus video models like Wan and LTX and audio generation with ACE-Step, with fine-grained layer targeting through only_if_contains and ignore_if_contains rules so training can be restricted to specific transformer blocks. Jobs are defined in YAML configs or managed from a bundled web interface served at localhost:8675, which handles datasets, job queues, sample generation, and checkpoint resumption after interruptions. Requirements are Python 3.10 or newer, an NVIDIA GPU, and a virtual environment install from source on Linux or Windows, with macOS support experimental; example configs target 24 GB class cards, and low VRAM options extend reach downward. Everything is MIT licensed with roughly 11,500 GitHub stars, and hosted training services have built their FLUX LoRA pipelines on top of it.