ostris/ai-toolkit

What is Ostris AI Toolkit?

Ostris AI Toolkit is an open‑source suite that lets you train and run diffusion models (image, video, audio, and even LLMs) on a single machine or cloud GPU. It provides:

A GUI web UI (served on http://localhost:8675) for starting, stopping and monitoring training jobs.
A CLI for the same tasks, useful on headless servers.
An experimental “AI Toolkit Manager” that auto‑detects your hardware, installs the right PyTorch build, creates a virtual environment, and bundles Node.js and FFmpeg so nothing is installed system‑wide.
Support for a huge list of pre‑trained diffusion models from Hugging Face (FLUX, SDXL, Qwen‑Image, etc.) covering image generation, instruction‑based editing, video generation, audio generation and even a multimodal LLM.
Configuration files (YAML) that let you train LoRA or LoKr adapters, target specific layers, and customise hyper‑parameters.
Integration scripts for cloud platforms such as RunPod and Modal, plus a link to the author’s own Ostris Cloud service.

The toolkit is aimed at hobbyists and researchers who want a “one‑click” experience on consumer‑grade GPUs while still exposing all the knobs a power user might need.

Key Features

Feature	Details
Model catalogue	Over 40 diffusion models (image, video, audio) and one LLM are listed, each linked to their Hugging Face repo.
Training UI	Web UI built with Node.js; token‑based auth for remote access.
CLI / manager	`python -m manager install
LoRA / LoKr support	Config‑driven adapters; layer‑targeting via `only_if_contains` / `ignore_if_contains`.
Cross‑platform	Scripts for Linux, macOS (Apple Silicon, experimental) and Windows.
Cloud ready	Ready‑made RunPod pod template; Modal deployment guide; optional Ostris Cloud GPU rentals.
Dataset handling	Simple folder of `.jpg/png` + matching `.txt` captions; automatic bucketing and resizing.
Security	UI token (`AI_TOOLKIT_AUTH`) to protect exposed servers.

Who Might Use It?

Artists & creators who want to fine‑tune diffusion models on their own hardware without writing boiler‑plate code.
Researchers needing a quick way to experiment with LoRA/LoKr on the latest FLUX or SDXL checkpoints.
Developers looking for a self‑contained environment that bundles PyTorch, Node, FFmpeg, and a UI.
Students who want a hands‑on playground for diffusion training and want to see results in a browser.

Getting Started (Brief)

Clone the repo.
Run the platform‑specific run_*.sh/.bat script – the manager will create a virtual env, install the correct CUDA‑enabled PyTorch, and launch the UI.
Pick a config from config/examples/ (e.g., train_lora_flux_24gb.yaml), edit paths/captions, then run:
```
python run.py config/your_config.yml
```
Monitor progress in the web UI or via the terminal.

Limitations / Things to Note

The AI Toolkit Manager is marked experimental; you can always fall back to the manual install steps.
Only JPEG/PNG images are supported; WebP is currently broken.
The UI is optional – training can be fully scripted.
Some models (e.g., FLUX.1‑dev) require a Hugging Face access token.

License & Community

The project is free and open‑source (license not specified in the README). The author encourages users to join the Discord for support and lists sponsorship options for those who wish to fund continued development.

In short: Ostris AI Toolkit is a genuine, feature‑rich training suite for diffusion models, offering both a convenient web UI and full CLI control, with extensive model support and cloud‑deployment guides.

All information above is taken directly from the repository’s README; no external assumptions have been added.