MoveNet Lightning, Thunder and MultiPose: Inputs, Outputs and Licence

This page is about MoveNet alone. The choice between it and MediaPipe is made in MediaPipe vs MoveNet, and the licence layers are read in full in pose estimation model licenses. Everything below was read from Google's two MoveNet model cards and the tensorflow/tfjs-models repository on October 17, 2026.

The three models#

The SinglePose card describes both single-person models as a MobileNetV2 feature extractor with a Feature Pyramid Network decoder and CenterNet prediction heads, Lightning at depth multiplier 1.0 and Thunder at 1.75. The MultiPose card gives the same architecture at depth multiplier 1.5 and recommends resizing the image so its larger side is 256 pixels, keeping the aspect ratio. In the TF.js package that size is multiPoseMaxDimension: a multiple of 32, 256 by default, with 128 to 512 recommended.

What comes out#

Read the coordinate order before writing any code. Per the SinglePose card, the raw model returns a float32 tensor of shape [1, 1, 17, 3]: for each keypoint, y and x normalized to the image frame (0.0 to 1.0), then a confidence score. The MultiPose card's output is [1, 6, 56]: six instance slots, each holding 17 groups of y, x and score, then a bounding box (ymin, xmin, ymax, xmax) in normalized coordinates and an instance score.

The TF.js pose-detection package converts this for you. Its README says each keypoint "contains x, y, score and name", that x and y "represent the actual keypoint position in the image", and that poseDetection.calculators.keypointsToNormalizedKeypoints converts them to the 0 to 1 range. With MultiPose, each returned pose also carries a box, and an id when tracking is on.

The 17 keypoints follow the COCO order, from 0 nose to 16 right_ankle. The full table, and the other models that share it, are in the COCO 17 keypoint map.

Confidence thresholds#

The SinglePose card says the model "predicts 17 human keypoints of the full body even when they are occluded", gives keypoints outside the frame low confidence scores, and recommends a default threshold of 0.3 to filter them. The TF.js package has a separate, pose-level threshold: minPoseScore, which falls back to DEFAULT_MIN_POSE_SCORE, 0.25 in the MoveNet constants file, when you do not set it.

Runtimes and where the weights come from#

In the browser, the MoveNet README pairs the package with a TF.js backend: WebGL, or WebAssembly, which it calls "the typical backend for devices with lower-end or no GPUs". By default the package loads weights from tfhub.dev, and the constants file names the three model URLs:

https://tfhub.dev/google/tfjs-model/movenet/singlepose/lightning/4
https://tfhub.dev/google/tfjs-model/movenet/singlepose/thunder/4
https://tfhub.dev/google/tfjs-model/movenet/multipose/lightning/1

The modelUrl option points the package at your own copy instead; the README calls it useful "for area/countries that don't have access to the model hosted on tf.hub", and it also accepts an io.IOHandler such as tfjs-react-native's bundleResourceIO to load from an app bundle. Outside the browser, both cards name TensorFlow Lite for on-device use, and the MultiPose card gives its input as int32 for TF.js and uint8 for TF Lite.

The published speed numbers#

The MoveNet README publishes frames per second for its three models, in the order SinglePose Lightning, SinglePose Thunder, MultiPose Lightning:

These are the TF.js team's figures on its own hardware. The README adds that for MultiPose "the number of detected persons does not impact inference speed and the accuracy of detections is similar to that of SinglePose Lightning". The cards' millisecond benchmarks are set against MediaPipe's on the comparison page; measure on your own target devices before you choose.

Licence#

The tensorflow/tfjs-models LICENSE is the Apache License, Version 2.0, and the SinglePose card says "Licensed Under Apache License, Version 2.0". The text of the MultiPose card has no license line. That gap, and what Apache-2.0 asks of you when you ship, are covered in pose estimation model licenses.

Two statements that disagree#

The pose-detection package README still says "Currently, only PoseNet supports multi-pose estimation." The MoveNet README in the same repository documents MULTIPOSE_LIGHTNING, and the MoveNet constants file lists it among the valid models. We go by the MoveNet README and the code.

Which one to load#

Our reading of Google's own descriptions: SinglePose Lightning for live rep counting where latency matters most, Thunder when you can spend more compute for a cleaner pose, and MultiPose Lightning only when more than one person must be tracked. The SinglePose card is explicit that it focuses "on detecting the pose of the person who is closest to the image center" and ignores the rest. What you then compute from the keypoints is covered in how rep counting works.