Pretrained models

Released estimator

Property Value
Model ID vdc-denoiser-m64-v1
Architecture Single-pass denoiser, 64 × 64 grid
Training step 190,000
Model host Hugging Face
SHA-256 8a1b71ad6d2012ed344b22a743a96e667d32a2ab54ade4444686e375e91f5c0f
Manifest vdc_denoiser_m64_v1.json

The model ID versions the weights separately from the Python package. A library update does not silently replace this checkpoint.

Retrained linear-grid estimators

Model ID Paper name Grid SHA-256 Status
vdc-denoiser-m64-v2 VDC-64 64 × 64 73097772a18eb4b02fb0608b4bfa87c0205b71667d97af6a39483fd44d0df0fc released
vdc-denoiser-m128-v2 VDC-128 128 × 128 f0a753d50b3cfb089e6d7dec999cfe741c6342c188454ec83183a9b3514d4653 released

These estimators of the paper's exploratory-retraining appendix are registered manifests (m64, m128) that record their hashes and evaluated inference settings. They complement, and do not replace, vdc-denoiser-m64-v1, which produced every main result of the paper. The manifests pin the Hugging Face revision of each upload; load_pretrained_model(<model id>) downloads and verifies the checkpoint, and VineCopulaModel.from_pretrained(bundle) applies the evaluated vine settings. The pilot from which both were fine-tuned is published as hsafaai/vdc-linear-grid-pilot (SHA-256 6e48f8c65012f4170ee9003c4aca28600abdb602ef624e47c3c3f92c0f5ace20) for exact retraining. See linear-grid estimators.

Download and cache

from vdc import load_pretrained_model
bundle = load_pretrained_model("vdc-denoiser-m64-v1", device="cpu")

Or download without loading the network:

vdc resolve-model --model-id vdc-denoiser-m64-v1

The cache defaults to ~/.cache/vdc; override it with VDC_PRETRAINED_CACHE or cache_dir=. The loader checks the checkpoint's SHA-256. An incorrect cached file is an error, not an accepted replacement. A public download needs no write token or personal cluster account.

Local and offline models

from vdc import load_checkpoint_bundle
bundle = load_checkpoint_bundle("/path/to/trusted-model.pt", device="cpu")

Only load trusted checkpoints. This reconstructs the model from its embedded configuration; it does not certify a custom checkpoint against the public model hash. The released denoiser and iterative diffusion models have different input and inference conventions, handled by the loader.

Publishing a separate model

Maintainers can use scripts/package_pretrained_model.py and scripts/upload_pretrained_to_hf.py; see each command's --help and the release guide. Give new weights a new model ID, publish their training/evaluation configuration, pin the host revision and record their hash. Never overwrite the weights of an existing released model ID.