MediaPipe Face Landmarker

Multi-stage MediaPipe facial geometry bundle (face_landmarker.task, 3.76 MB, FP16 weights) combining BlazeFace detection, 478-point 3D FaceMesh V2 landmarks (including 10 iris landmarks), and 52 Blendshape V2 expression coefficients for the MediaPipe Face Landmarker (FaceLandmarker) task.

Documentation: MediaPipe Face Landmarker Docs

Samples: Android Β· iOS Β· Python Colab Β· Web Β· Flutter

Try live: Web Demo

face_landmarker.task bundles three neural networks plus canonical landmark geometry metadata (geometry_pipeline_metadata_landmarks.binarypb):

  1. face_detector.tflite: BlazeFace short-range detector ([1, 128, 128, 3] RGB input).
  2. face_landmarks_detector.tflite: FaceMesh V2 model ([-1, 256, 256, 3] RGB input) outputting 478 3D face/iris landmarks ([-1, 1, 1, 1434]), face presence probability, and tongue_out score.
  3. face_blendshapes.tflite: MLP-Mixer network consuming 146 2D landmarks ([1, 146, 2]) to output 52 facial expression blendshape coefficients in [0, 1].

Mirroring note: *Left and *Right blendshape scores refer to the subject's face, not the viewer (mirrored selfie frames will invert left/right blendshapes).

Files

File Precision Size SHA-256
face_landmarker.task float16 3.76 MB (3,758,596 B) 64184e229b263107bc2b804c6625db1341ff2bb731874b0bcc2fe6544e0bc9ff

Companion Files in Repo Root

File Role Size SHA-256
MediaPipe BlazeFace Model Card (Short Range).pdf Upstream Google model card (PDF) 434.4 KB (434,425 B) cd335c06fc0de7807cd815a0777a697932598bcdb28fa98adaaabf847485f758
Model Card Blendshape V2.pdf Upstream Google model card (PDF) 2.32 MB (2,324,099 B) c8e9cf60a39998f4b341740623917590e050d1c97004e2de4568d84e026445ae
Model Card MediaPipe Face Mesh V2.pdf Upstream Google model card (PDF) 3.48 MB (3,476,420 B) c6add060f4ebfb37b2690136b6c711c7e5fcb7038baa2649ae3338b83979565a

Download

pip install -U huggingface_hub
hf download litert-community/MediaPipe-FaceLandmarker face_landmarker.task --local-dir .

Inputs, Outputs, and Model Specifications

Bundle Contents (face_landmarker.task)

Bundled Member Uncompressed Size SHA-256 (prefix)
face_detector.tflite 229.7 KB (229,746 B) b4578f35940bf5a1…
face_landmarks_detector.tflite 2.55 MB (2,553,590 B) c7d54204ce044847…
geometry_pipeline_metadata_landmarks.binarypb 19.4 KB (19,376 B) bdbcda96dfcb7da8…
face_blendshapes.tflite 955.3 KB (955,312 B) 4f36dded049db18d…

Sub-model face_detector.tflite

  • Embedded TFLITE_METADATA: name Short Range Face Detection Β· version 1 Β· author MediaPipe
Direction Tensor Shape Type Notes
Input input [1, 128, 128, 3] float32 image β€” Input image to be detected β€” Normalization: mean=[127.5], std=[127.5]
Output regressors [1, 896, 16] float32 raw boxes/keypoints β€” Undecoded face bboxes location and keypoints
Output classificators [1, 896, 1] float32 scores β€” Scores of the detected bboxes.

Sub-model face_landmarks_detector.tflite

  • Embedded TFLITE_METADATA: name Face mesh detection model v2 Β· version 1 Β· author MediaPipe
Direction Tensor Shape Type Notes
Input input_12 [-1, 256, 256, 3] float32 image β€” Input image to be detected β€” Normalization: mean=[0.0], std=[255.0]
Output Identity [-1, 1, 1, 1434] float32 face_landmarks β€” Undecoded face landmarks. Each landmark contains x, y, and z coordinates
Output Identity_1 [-1, 1, 1, 1] float32 presence β€” Prersence score of the face.
Output Identity_2 [-1, 1] float32 tongue_out β€” Score of whether tongue is out of mouth.

Sub-model face_blendshapes.tflite

Direction Tensor Shape Type Notes
Input serving_default_input_points:0 [1, 146, 2] float32 β€”
Output StatefulPartitionedCall:0 [52] float32 β€”

Use with MediaPipe Tasks

Python Quickstart

pip install -U mediapipe huggingface_hub
import mediapipe as mp
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="litert-community/MediaPipe-FaceLandmarker",
    filename="face_landmarker.task",
)

BaseOptions = mp.tasks.BaseOptions
FaceLandmarker = mp.tasks.vision.FaceLandmarker
FaceLandmarkerOptions = mp.tasks.vision.FaceLandmarkerOptions
VisionRunningMode = mp.tasks.vision.RunningMode

options = FaceLandmarkerOptions(
    base_options=BaseOptions(model_asset_path=model_path),
    running_mode=VisionRunningMode.IMAGE,
    num_faces=1,
    output_face_blendshapes=True,
)

with FaceLandmarker.create_from_options(options) as runner:
    mp_image = mp.Image.create_from_file("input.jpg")
    result = runner.detect(mp_image)
    if result.face_landmarks:
        print(f"Landmarks: {len(result.face_landmarks[0])}")
    if result.face_blendshapes:
        top = sorted(result.face_blendshapes[0], key=lambda c: c.score, reverse=True)[:3]
        print("Top blendshapes:", [(c.category_name, round(c.score, 3)) for c in top])

Cross-Platform Guides & Samples

Latency Benchmarks

The MediaPipe Face Landmarker documentation does not publish a Pixel 6 latency table. Bundled model cards evaluate BlazeFace (Short Range) (2.94 ms Pixel 6 CPU / 7.41 ms GPU in Face Detector docs), Face Mesh V2 accuracy, and Blendshape V2 consistency.

Intended Use, Limitations, and Fairness

Summarized from the three bundled upstream PDF model cards:

  1. MediaPipe BlazeFace Model Card (Short Range).pdf (June 9, 2021) β€” short-range (β‰ˆ2 m) face detection. Across 17 UN subregions: recall 99.1% / precision 95.1%; gender: recall 98.4% / precision 99.9%; 5 Fitzpatrick groups: recall 98.1% / precision 99.7%.
  2. Model Card MediaPipe Face Mesh V2.pdf (September 15, 2022) β€” 478 3D surface & iris landmarks evaluated on 1,700 selfie images across 17 UN regions (100/region) via 2D IOD MAE (vs. 2.56% human inter-annotator discrepancy): overall 2.62% tracking / 3.24% reacquisition; across genders 2.61%–2.62% tracking; across Fitzpatrick 1–6 2.49%–2.90% tracking.
  3. Model Card Blendshape V2.pdf (November 11, 2022) β€” 52 AR facial blendshapes evaluated on 511 multi-view lab subjects (18,199 high-activation expression items): overall Mean Absolute Deviation (MAD) 0.199 (STDEV 0.237); gender MAD 0.196–0.203; Monk skin tones 1–10 MAD 0.192–0.210.
  • Out-of-scope uses:
    • Any form of surveillance or identity recognition is explicitly out of scope and not enabled by this technology.
    • Predicted face landmarks and facial blendshapes do not provide facial recognition or identification and do not store any unique face representation.
    • Human life-critical decisions, crowd counting, or faces > 80Β° rotated / < 50% visible.

Provenance

Every model binary in this repository is byte-identical (unmodified weights, metadata, and .task bundle archive bytes) to the official artifact served by Google from storage.googleapis.com/mediapipe-models.

Repo File Upstream URL GCS Generation Upstream Last-Modified Pinned Bucket Version
face_landmarker.task face_landmarker.task 1683136941468629 Wed, 03 May 2023 18:02:21 GMT /1/

License and Citation

  • License: apache-2.0 β€” Stated as Apache License, Version 2.0 in the bundled upstream Google model-card PDFs.

Citation

@inproceedings{bazarevsky2019blazeface,
  title     = {BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs},
  author    = {Bazarevsky, Valentin and Kartynnik, Yury and Vakunov, Andrey and Raveendran, Karthik and Grundmann, Matthias},
  booktitle = {CVPR Workshop on Computer Vision for Augmented and Virtual Reality},
  year      = {2019},
  url       = {https://arxiv.org/abs/1907.05047}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including litert-community/MediaPipe-FaceLandmarker

Paper for litert-community/MediaPipe-FaceLandmarker