You can deploy this model and start earning money today!

panopstor/
MiVOLO-mirror

Name: MiVOLO-mirror
Availability: InStock

Views

No views yet

MiVOLO weights mirror

Convenience HuggingFace mirror of the two MiVOLO checkpoints needed by the upstream WildChlamydia/MiVOLO inference pipeline. Upstream distributes these via Google Drive only; this mirror exists so automated tooling can huggingface_hub.snapshot_download them instead of relying on gdown.

This is a re-distribution. All credit and attribution goes to the original MiVOLO authors (Maksim Kuprashevich and Irina Tolstykh, SberDevices). Please cite their paper and use the upstream repo as the source of truth:

Paper: MiVOLO: Multi-input Transformer for Age and Gender Estimation
GitHub: https://github.com/WildChlamydia/MiVOLO
Model card on the upstream repo: see their README for benchmarks and citations.

Files

File	Size	What it is
`yolov8x_person_face.pt`	~131 MB	YOLOv8x detector for person + face bounding boxes. Standard `torch.load`-compatible PyTorch zip checkpoint.
`mivolo_face_body.pth.tar`	~105 MB	MiVOLO age + gender model (face+body variant; the SOTA option in the upstream repo with reported MAE ≈ 4.24 yr face+body / 6.87 yr body-only). Despite the suffix, this is not a tar archive — it's a single `torch.load`-compatible Python pickle. The dict contains `state_dict` (290 tensor params) plus metadata fields (`min_age`, `max_age`, `avg_age`, `no_gender`, `with_persons_model`). The `.pth.tar` extension is an old PyTorch-examples convention that stuck.

Both files are identical to the originals downloaded from the upstream Google Drive links cited in the MiVOLO README.

Original Drive links (for reference / diff verification):

yolov8x_person_face.pt → https://drive.google.com/file/d/1CGNCkZQNj5WkP3rLpENWAOgrBQkUWRdw/view
mivolo_face_body.pth.tar → https://drive.google.com/file/d/11i8pKctxz3wVkDBlWKvhYIh7kpVFXSZ4/view

License note

The upstream MiVOLO repo includes a top-level LICENSE file that is Apache License 2.0, but their setup.py classifier declares CC BY-SA 4.0. The two are inconsistent. The plain-text LICENSE file is typically considered the legally controlling artifact, so this mirror is tagged apache-2.0 to match. If the upstream authors clarify the discrepancy in the future, this metadata may be updated.

If you re-distribute these weights elsewhere, preserve the credit, link back to the upstream repo, and reproduce both license artifacts in your copy until the upstream authors resolve the inconsistency.

Usage from the parent project

These checkpoints are consumed by MicroMaxField's training-data safety pass (see its SAFETY.md for the design). The scripts/setup_safety.sh setup script in that repo can fetch these files via huggingface_hub instead of Google Drive.

Standalone usage (with the upstream MiVOLO repo)

python

import argparse
from huggingface_hub import snapshot_download
from mivolo.predictor import Predictor

ckpt_dir = snapshot_download("panopstor/MiVOLO-mirror")

config = argparse.Namespace(
    detector_weights=f"{ckpt_dir}/yolov8x_person_face.pt",
    checkpoint=f"{ckpt_dir}/mivolo_face_body.pth.tar",
    device="cuda",
    with_persons=True,
    disable_faces=False,
    draw=False,
)
predictor = Predictor(config, verbose=False)

(Note: at the time of this mirror, MiVOLO pinned timm==0.8.13.dev0 and ultralytics==8.1.0 in their requirements.txt. Newer versions of either break model instantiation — see the MicroMaxField SAFETY.md for a sidecar pattern that isolates the old pins from a modern training stack.)