Teach a vision model with a few hundred pictures.

DinoTraining is a free desktop app for the whole loop: annotate pictures with Grounding DINO and SAM, train small heads on frozen DINOv2/DINOv3 backbones in minutes, run them, and let them annotate new data.

Looking up the latest version…

Windows

Windows 10/11, 64-bit

First release coming

The installer is not signed with a paid certificate, so Windows shows a SmartScreen warning once: More info → Run anyway. No admin rights needed.

macOS

Apple silicon (M1 and newer)

Open Terminal and paste:

bash -c "$(curl -fsSL https://raw.githubusercontent.com/JanderHungrige/DinoTraining/main/scripts/install-mac.sh)"

Installs into your Applications folder, without admin rights and without a security warning. Not in the browser: macOS blocks unsigned apps that a browser downloaded. Run it again to update. Intel Macs are not supported (PyTorch no longer builds for them).

Linux

Ubuntu / Debian, x86-64

First release coming

Then, in the download folder:

sudo apt install ./DinoTraining_*_amd64.deb

The first start

The installers are small (about 20 MB). On its first start the app downloads Python and PyTorch once, from the official sources, in exactly the versions it was tested with: about 1 GB, or about 3.5 GB with NVIDIA GPU support. After that it works offline.

Your GPU

NVIDIA cards are detected and used through CUDA, Apple silicon through MPS. Without a GPU the app runs on the CPU; you can switch later in the app.

Open source

MIT licensed. Models are downloaded in the app from Hugging Face; your pictures and datasets never leave your computer.

All releases and checksums · The code on GitHub · Uninstalling, and building from source