What it is
AI Toolkit by Ostris is a training suite for diffusion models. It is mostly used to train LoRAs, small add-on weights that teach an image or video model a style, person or concept, and it now covers audio models such as ACE-Step too.
It supports more than 50 open image, edit and video models, from FLUX.1 and Qwen-Image to Wan 2.2 and LTX-2.5, driven by YAML configs or a web UI. It runs locally on Nvidia GPUs, on RunPod or Modal, or on Ostris Cloud, a GPU rental service run by the author to fund the project.
It has no tagged releases; commits land almost daily, the latest on 23 October 2026.
Our review
AI Toolkit is the fastest route to training a LoRA on a newly released open model, because its author adds them within days. The README now lists more than 50 image, edit, video and audio models, and recent commits added Bria Fibo 2.0 on 23 October 2026. The repo has about 12,400 stars, which understates its reach among people who train LoRAs.
The web UI is what separates it from script collections. Datasets, captions, sample prompts and job control sit in one interface at port 8675, and the UI can be closed while jobs run. A new AI Toolkit Manager, marked experimental, sets up PyTorch, Node.js and FFmpeg inside the folder and checks for updates on every launch.
That update behaviour is also the main risk. There are no releases, so every launch can pull new code; the manager skips updates if you have edited files, but a working setup can still change under you. The UI is exposed on the network with no password unless you set AI_TOOLKIT_AUTH, which the README strongly advises on any cloud host. Pressing Ctrl+C during a checkpoint save will likely corrupt it.
Hardware sets the real price. The reference FLUX LoRA config targets a 24 GB card, and the Modal example defaults to a 40 GB A100 with a two-hour timeout. Ostris Cloud charges $1.43 an hour for an RTX 5090, so a typical LoRA run costs a few dollars of GPU time. Diffusers' training scripts are better documented and easier to modify; AI Toolkit trades that for breadth and speed of model support.
Where it earns its place
A freelance illustrator with an RTX 4090
Train a style LoRA on 30 of their own paintings
Load the images and captions in the web UI, start from the 24 GB FLUX config and watch sample images as it trains. No software cost; the run takes hours of a card they already own.
A creative studio producing product shots for e-commerce clients
Keep a client's product consistent across generated images
Rent an RTX 5090 on Ostris Cloud at $1.43 an hour, train a LoRA per product and hand the file to the image team. A few dollars of GPU per product; the base model's own licence decides whether the output can be used commercially.
An ML engineer evaluating a newly released video model
Test whether fine-tuning Wan 2.2 or LTX-2.5 suits the team's footage
Clone on a cloud GPU box, set AI_TOOLKIT_AUTH, and train a short LoRA from a YAML config in the CLI. Quicker than writing a training loop; pin a commit once a config works.
Before you commit
- Setup
- Clone and run `./run_linux.sh` (or the Windows and macOS scripts) to let the manager install everything and open the UI at localhost:8675; the manual path needs Python 3.10+, a CUDA build of PyTorch and Node.js 20+.
- Integrations
- Hugging Face for model downloads (gated models need a read token); official RunPod template; Modal script; Ostris Cloud; YAML configs for CLI runs.
- Data and privacy
- Runs locally; contacts Hugging Face for weights and GitHub for manager updates on launch. The UI listens on port 8675 without authentication unless AI_TOOLKIT_AUTH is set. MIT code; each base model has its own licence.
- Cost at scale
- No paid edition. Self-hosting needs an Nvidia card, 24 GB for the reference FLUX config; Ostris Cloud rents GPUs from $0.71/hour (A40) to $2.99/hour (RTX PRO 6000), billed per second.
- Limits
- No releases or changelog; checkpoints corrupt if interrupted mid-save; macOS support is experimental; WebP datasets have issues; documentation is mainly the README and example configs.
What's good
- New open models supported within days of release
- Web UI for datasets, captions, samples and jobs
- More than 50 image, edit, video and audio models
- Runs locally, on RunPod or Modal, or on Ostris Cloud
What's not
- No releases; the manager auto-updates on launch
- UI has no password unless AI_TOOLKIT_AUTH is set
- Many models need a 24 GB or larger GPU
- Thinner documentation than Diffusers
Best for
Creators training LoRAs for open image and video models.
Alternatives
AI Toolkit: common questions
- Is AI Toolkit worth it?
- AI Toolkit is where support for a new open image or video model usually lands first, and it expects you to bring the GPU and the patience. We rate it 4.2 out of 5.
- How much does AI Toolkit cost?
- Free and open source under MIT. Ostris Cloud rents preconfigured GPUs by the second, from $0.71/hour (A40 48 GB) to $2.99/hour (RTX PRO 6000 96 GB).
- What are the downsides of AI Toolkit?
- No releases; the manager auto-updates on launch; uI has no password unless AI_TOOLKIT_AUTH is set; many models need a 24 GB or larger GPU; thinner documentation than Diffusers.
- Who is AI Toolkit best for?
- Creators training LoRAs for open image and video models.
- How much does AI Toolkit cost for a team or in production?
- No paid edition. Self-hosting needs an Nvidia card, 24 GB for the reference FLUX config; Ostris Cloud rents GPUs from $0.71/hour (A40) to $2.99/hour (RTX PRO 6000), billed per second.
- Is AI Toolkit safe for company data?
- Runs locally; contacts Hugging Face for weights and GitHub for manager updates on launch. The UI listens on port 8675 without authentication unless AI_TOOLKIT_AUTH is set. MIT code; each base model has its own licence.
- What are the best AI Toolkit alternatives?
- The closest alternatives we have reviewed are Hugging Face Diffusers and ComfyUI.
Answers checked Oct 23, 2026.
Make AI Toolkit? Show this rating on your site
