Pinned ONNX mirrors of third-party models used for on-device inference in
Arca. Nothing here is trained by us — every file is a
byte-for-byte copy of an upstream release, re-hosted so that:
Downloads are immutable. Arca pins each file by commit SHA and verifies a
SHA-256 before use. Fetching upstream main gave us a moving target: two of
our source URLs 404'd after upstream repos were restructured, silently
breaking on-device features with nothing able to detect the drift.
Upstream repos aren't used as a CDN. Pulling a 261 MB file from a
github.com/.../raw/main/ URL on every client install is not what that
endpoint is for.
If you want these models, please prefer the original sources below — they are
the authoritative, maintained copies.
Contents
Every file here is live in Arca. Nothing is staged "for later" — an unused model
in a mirror is just a file nobody can explain a year from now.
A .mlpackage is a directory bundle, not a single file, so it is zipped
(ditto -c -k --sequesterRsrc --keepParent) for single-URL distribution and
single-SHA-256 pinning, matching every other file in this repo. The installing
client (FaceModelInstaller.swift) unzips it before handing it to
MLModel.compileModel(at:) — it never compiles the zip directly.
Converted from glintr100.onnx (from the same fal/AuraFace-v1 repo as its
sibling scrfd_10g_bnkps.onnx/2d106det.onnx/1k3d68.onnx/genderage.onnx —
only glintr100.onnx is mirrored here. The rest of that repo is stock
InsightFace detection/auxiliary models with non-commercial upstream terms;
glintr100.onnx is the one file fal/AuraFace-v1's own LICENSE.md (stock
Apache-2.0, verified in full, no field-of-use rider) actually covers, and it is
the only one Arca needs — detection and landmarks come from the OS's own face
API on every platform, never from this mirror. Bridge: ONNX → PyTorch via
onnx2torch (coremltools dropped direct ONNX conversion) → traced →
coremltools.convert(..., convert_to="mlprogram"). Verified against the
original ONNX output on 5 real photos: max abs diff ~5e-6, cosine similarity
1.0 — floating-point noise, not a converter defect.
Note on the CLIP pair
clip-vit-b32-image.onnx and clip-vit-b32-text.onnx come from the same
export family (both PyTorch 2.3.0 exports of clip-ViT-B-32) and therefore
share one 512-d embedding space. Mixing an image encoder from one port with a
text encoder from another puts the vectors in different spaces and makes cosine
similarity meaningless. Treat the two as a single atomic unit.
Previously here, removed
Both were removed from main by a normal commit, so earlier pinned
resolve/<sha>/ URLs still resolve — nothing that referenced them breaks.
yunet.onnx (face detection, MIT) — Arca uses the OS
Windows.Media.FaceAnalysis.FaceDetector for detection, so this was never
loaded. arcface.onnx (the 512-d embedder) is the only ONNX face model.
yolox_tiny.onnx (object detection, Apache-2.0) — mirrored in
anticipation of an object-detection feature that doesn't exist yet, then
removed rather than left sitting here unused. If it comes back, it will be
re-uploaded alongside the code that actually calls it.
Licensing
Each file keeps its upstream license; see the table above and LICENSES.md.
The license: other / mixed-permissive marker reflects that this repo
aggregates MIT and Apache-2.0 files rather than being under one single license.
No copyleft-licensed weights are hosted here — see LICENSES.md for why
Ultralytics YOLO in particular can never be.
