Skip to content

Downloading weights

imagehub never downloads anything on its own. pipeline() only builds an object, and loading reads local files. Fetching weights is a separate command you run on purpose.

python -m imagehub.pretrained <model-id> <empty-directory>
python -m imagehub.pretrained flux1-schnell checkpoints/flux1-schnell

or from Python:

from imagehub.pretrained import download_pretrained

download_pretrained("sd-v1-5", "checkpoints/sd-v1-5")

What the downloader guarantees

  • The revision is pinned to the commit that passed the parity tests (see the catalog). Asking for another revision works but is recorded as unverified.
  • Every file is checked against its size. Weight files are also checked against the sha256 the Hub publishes; other files against their git blob hash.
  • The destination must be empty, and an imagehub_snapshot.json manifest records exactly what was downloaded.
  • Only HTTPS is accepted, and credentials are dropped on cross-host redirects.

Gated repositories

Some repositories require you to accept terms first. Accept them on the model page on Hugging Face, create a read token, and pass it through the environment:

export HF_TOKEN=...
python -m imagehub.pretrained krea2-turbo checkpoints/krea2-turbo

Disk space

Checkpoints are large, often many gigabytes. Keep only the models you use; a snapshot is an ordinary folder, so deleting it is all it takes to free the space.

Licenses

imagehub is Apache-2.0, but model weights are not. Each has its own license, listed per model on the catalog and the model pages. Some are non-commercial, and some restrict regions or uses. The license binds whoever downloads and uses the weights.