tomas-gajarsky/facetorch-detector-retinaface

RetinaFace ResNet-50 face detector

RetinaFace with a ResNet-50 backbone, exported for Facetorch face detection.

This repository contains immutable model artifacts used by Facetorch. Use the packaged Facetorch manifest to select a revision and artifact; do not treat mutable main or older unlisted files as a release contract.

Contract

Field	Value
Model ID	detector-retinaface
Architecture	RetinaFace, ResNet-50 backbone
Input	RGB image tensor converted to BGR, with spatial dimensions divisible by 32
Output	Sequence of bounding-box regressions [batch, anchors, 4], class probabilities [batch, anchors, 2], and five-point landmark regressions [batch, anchors, 10].
Dynamic shapes	Batch 1 for the release contract; height and width are dynamic multiples of 32.
Weights license	MIT

Preprocessing: Convert RGB to BGR and subtract channel means [123, 117, 104]. Facetorch performs this preprocessing.

Release artifacts

File	Format	Runtime	Devices	SHA-256
model-torch2.6.pt2	pt2	>=2.6, <2.7	cpu, cuda	36ae7d95c4ae0f4faebc87d42ea9dbc55c9a030808c75725cb567d886cd09a11
model-torch2.11.pt2	pt2	>=2.11, <2.12	cpu, cuda	9541805b9b11d1fc6e329e365f010c310f5c3509b7d5e4236eb0797dfed66d01
model.pt	torchscript	>=2.6, <2.12	cpu	05e524af9b55bbf92b752b064c298f170ae763eb0ac0a4162a92e39fff007def

Facetorch v1 supports the Torch 2.6 and 2.11 cohort files listed in its manifest. The legacy TorchScript object is CPU-only and requires the explicit legacy opt-in. Files from unsupported cohorts are not part of the v1 release contract.

Loading the manifest-selected artifact

import torch
from huggingface_hub import hf_hub_download
from facetorch.artifacts import get_model_manifest

MODEL_ID = "detector-retinaface"
device = "cuda" if torch.cuda.is_available() else "cpu"
artifact = get_model_manifest().candidates(
    MODEL_ID,
    torch_version=torch.__version__,
    device=device,
    allow_legacy_models=False,
)[0]
path = hf_hub_download(
    repo_id=artifact.repo_id,
    revision=artifact.revision,
    filename=artifact.filename,
)
model = torch.export.load(path).module().to(device).eval()
example = torch.randn(1, 3, 480, 640, device=device)
with torch.inference_mode():
    output = model(example)

The random tensor above is only a loading smoke test. Use Facetorch's documented preprocessing for meaningful inference.

Provenance

Upstream	Immutable revision	Role	License
https://github.com/biubug6/Pytorch_Retinaface	b984b4b775b2c4dced95c1eadd195a5c7d32a60b	checkpoint publisher and architecture source	MIT

Upstream checkpoint	SHA-256	Source
Resnet50_Final.pth	6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d	publisher location

Mapping method: exact_tensor_equality_after_module_prefix_removal_and_torchscript_attribute_recovery.

Result: 456 of 456 tensors matched exactly; 219 legacy TorchScript state attributes were recovered without substitution.

The repository owner approved the mapping and redistribution record on 2026-08-23. Under the recorded policy, an author-published checkpoint in a permissively licensed repository with no separate checkpoint terms uses that repository license. MIT and Apache-2.0 have not been converted or treated as interchangeable. See LICENSE, THIRD_PARTY_NOTICES.md, and Facetorch's facetorch/models/governance.json.

Papers

• RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild

Intended use

• Research and assisted face localization with application-specific evaluation.

Limitations and responsible use

• May miss, duplicate, or mislocalize faces under occlusion, extreme pose, small scale, or domain shift.
• Detection performance is not evidence of equitable performance across populations or capture conditions.
• The checkpoint license does not grant rights to upstream training datasets.
• The artifact license does not itself license training datasets, input data, or a deployment's processing of personal data.
• Do not use model output as the sole basis for consequential decisions about a person.