AI Toolkit (ostris)
An open-source diffusion model training suite supports image and video fine-tuning, with a self-hosted web interface and an MIT license.
AI Toolkit (ostris) is a training suite for people fine-tuning image and video generation models on their own hardware or rented servers. It provides a browser interface and command-line access to training jobs.
- Image and editing models: Train models including FLUX, SDXL and Qwen-Image, plus editing models such as FLUX.1-Kontext-dev and Qwen-Image-Edit.
- Video training: Work with Wan, LTX and MiniMax-H3 models, including MiniMax-H3 training conditioned on a reference image.
- Audio and multimodal models: Train ACE-Step audio models and Qwen2.5-Omni alongside the image and video models.
- Job management: Start, stop and monitor training through the web interface; jobs continue when the interface isn't running, and saved checkpoints allow training to resume.
The Python project is free and open source under MIT and targets consumer hardware. You can self-host the interface, with optional token authentication to restrict access. Training can also run on rented GPUs through Ostris Cloud, RunPod or Modal. With Modal, datasets go to cloud hardware and model outputs reside in a cloud storage volume.
Logotyper
