Facepoints: 620 dense face landmarks on the GNM Head mesh

Finds 620 points on a face in a photo. Every point is a fixed location on Google's open GNM Head mesh, so the output fits that model to a person directly — for building or animating a 3D avatar. The same forward pass also gives the standard 68-point (iBUG) and 98-point (WFLW) landmark sets and an error estimate for every point. To our knowledge (2026-10-23) it is the first publicly released detector trained on points defined on the GNM mesh.

Input	one face, cropped from a photo or video frame; a face detector supplies the box
Output	620 points, each (x, y) plus its own error estimate, in pixels; the 68- and 98-point sets
Points	416 landmark and 150 skin points, 36 eyelid-edge, 16 front teeth (7 upper, 9 lower), 2 iris centres (points_gnm.json maps each to the GNM Head mesh, v3.0)
Model	HRNet-W48, 256-px input, 102.7 M parameters (+1.2 M for the 68/98 heads)
Weights	model/model.safetensors + model/config.json; SHA-256 of the weights b37075caa3eaf0d00f23f32925b16ec3a8e2e45c42aa9c00956f94f2574e6526

Use

git clone https://huggingface.co/myned-ai/facepoints && cd facepoints   # git-lfs for the weights
pip install -e .    # torch, timm, safetensors, opencv-python-headless, numpy

import cv2
from facepoints import Detector, square_box, gnm_positions

det = Detector("./model")                       # the folder with model.safetensors + config.json
img = cv2.imread("face.jpg")                    # BGR
x, y, w, h = my_face_detector(img)              # any face box, e.g. OpenCV's YuNet
out = det.detect(img, square_box(x, y, w, h))   # the box, then two passes on the model's own extent

out["points"], out["std"]   # (620, 2) image px, and one predicted std per point (px)
out["300w"], out["wflw"]    # (68, 2) and (98, 2)
gnm_positions(mesh_xyz)     # (620, 3): the same points on any posed GNM Head mesh (weights x vertices)

Fitting the GNM Head to a face = solving for the mesh whose projected gnm_positions match out["points"], weighted by 1 / std. One face at a time; the output moves with the box, which is why detect re-frames on its own points. A box with no face in it yields points with a very large std — threshold it. A complete example with OpenCV's face finder: python run.py photo.jpg out.jpg (the model from ./model, or --model DIR); tests: pytest tests.

Results

Against registered 3D meshes of 9 studio people none of the models trained on, 14 cameras, their widest mouth shapes. The score is how well the points follow the face: median distance to the paired truth vertex, in mm, after removing each point's constant offset at rest (protocol in the report).

	lips	cheeks	chin	rest of face
ours	1.78	2.00	2.12	0.81
Pixel3DMM	1.81	2.19	1.67	0.78
3DDFA-V3	2.40	2.61	2.84	1.41

Level with Pixel3DMM except on the chin, ahead of 3DDFA-V3 on every region.

On the 2D benchmarks (the 68/98-point outputs; full test sets; NME = mean point error as % of the eye-corner distance):

	300W	300VW	WFLW
ours, 68/98-point heads	3.33	3.52	4.66
ours, 620 points through a linear map	3.31	3.56	4.59
STAR (official checkpoints)	3.06	3.45	4.47
SPIGA	3.02	3.37	4.06
RwSA (ICCV 2025)	2.96	3.38	4.23
FAN	3.69	4.11	7.27
MediaPipe (478 points, mapped)	3.87	4.00	7.72

Behind the three models built for these benchmarks by 0.2–0.6, level with STAR on video, ahead of FAN and MediaPipe everywhere. Profile faces are the weak spot.

The error estimate (std) comes from a Student-t likelihood (ν = 10). Error rises with it, but the intervals are too narrow: multiply std by 1.34 to calibrate on average. Measured on studio faces only.

Limitations

• Hidden points (teeth behind closed lips, the far side of a turned face) are estimated, not seen; sunglasses give guessed eyelids and irises; closed eyes come out about 1 mm open.
• The 620 points have no hand-labelled truth on wild photos; their wild accuracy is known only through the 68/98 outputs.
• On 24 held-out studio people, error correlated with darker skin (Spearman −0.44, p = 0.03), confounded with exposure; Ava-256 is biased to lighter-skinned people under 35.

Training

The 620 points: Ava-256 (Meta), 230 people, 16 cameras; labels are projections of each person's registered mesh, whose vertices were mapped onto the GNM template. The 68/98 heads: the 300W and WFLW training photos; on those photos the 620 points carry pseudo-labels. HRNet-W48, three stages, one run. To replicate the model, follow train/README.md: data preparation, the three commands and the export. The scripts and the shipped sample lists, split and pseudo-label map: prep/; every key: train/config.py. Protocol, data and analysis in full: technical report (in preparation).

Licence

Copyright 2026 Myned AI (opensource@myned.ai).

	licence	source
code	Apache-2.0	LICENSE
model/model.safetensors, points_gnm.json, prep/*.npz	CC BY-NC 4.0	LICENSE-WEIGHTS
Ava-256 · Multiface (images)	CC BY-NC 4.0	LICENSE files
300W · WFLW (WIDER FACE images) · 300VW	research only · CC BY-NC-ND · none stated	dataset pages
ImageNet start (timm hrnet_w48.ms_in1k)	weights MIT; ImageNet terms non-commercial	model card
GNM Head	Apache-2.0, attribution required	LICENSE file

Figures show Multiface people and WFLW test faces as the benchmark publishes them.