Can One Camera Move an Avatar’s Whole Body? What the Free Software HIROMEIRO Does, From Its Manual and Source

2026年10月13日

13 October 2026

I usually write about local generative AI here, but this time it is software that reads body movement from camera video. It is called HIROMEIRO, and its official release came out free of charge on 17 September 2026.

To move an avatar with the whole body, as a VTuber does, people tend to assume you need dedicated gear such as body-worn trackers. HIROMEIRO says it moves a full-body avatar with a single webcam or phone camera. The thing reading the body’s movement is AI running on your PC or phone. Can you really move an avatar’s whole body with the PC or phone you already have?

For this article I read the official manual and the published source code. I sorted out what HIROMEIRO can do, what runs on which OS, and what camera and space you need. I also compared it in numbers, price and the number of points read, with similar software that moves the whole body from a camera alone.

This reflects information as of 28 September 2026.

What I looked into and how I checked: scope and sources

I looked at five things: what kind of software HIROMEIRO is, where the AI that reads body movement runs, which OS and cameras it works with, how it differs from similar software, and what to watch for before using it.

The main program that displays the avatar is Windows-only, and the PCs I use here run Linux and macOS, so I have not tried HIROMEIRO yet. Everything I say about how it behaves is what is written in the sources below.

Source	Contents	Treatment
Official site and manual	What it can do, operating environment, how to use it, known bugs, FAQ	Official
GitHub README and public source	How the apps that read body movement (Windows, Android and Python versions) are built	Official (source read)
List of files in the distribution	Names of files inside the official release zip (contents not opened)	Checked distribution
PR TIMES, MoguLive	The announcement of the official release and an article introducing it	Official announcement and news

What does HIROMEIRO do?

According to the manual, it is software that moves a VRM avatar you already have with the whole body, without tracking hardware. All you need is one camera, and ordinary clothes are fine. It was developed by Kushinada Tech Co., Ltd., and the official release follows a development version public since 2023.

What is VRM?

VRM is a file format for saving 3D avatars. It holds a human-shaped model together with its skeleton (bones) and facial expression information in a fixed form. Because the format is standardized, an avatar made in one program can be brought into another and moved. The README says HIROMEIRO has been tested with VRM files exported from VRoid Studio.

What it can do

Feature	Details
Full-body tracking	Maps whole-body movement onto the avatar with one camera
Upper-body-only mode	Moves by the angle of the face and shoulders and the position of both hands. For streaming while seated, or when the whole body does not fit in the frame
Loading avatars	Switch by dragging and dropping a VRM file onto the screen. A file to load at startup can also be set
Backgrounds (stages)	Three: green screen, park, indoor gym. The green screen’s color can be changed
Vertical video	Turning the camera 90 degrees in the settings file gives a vertical screen for short videos
Hiding the input video	The PC tracking app does not show your own image from the camera on screen (on from the start)
Recording and playback of motion	Saves the read motion as data lighter than video, and later plays it back at normal speed to send to the main program

Where does the AI that reads body movement run?

What reads the movement of the body in the camera image is the Holistic Landmarker, an AI model in MediaPipe, a toolkit published by Google. For a person in the video, it estimates the key points (landmarks) of body, face and both hands together with a single model. Google’s description gives 33 points for the body, 478 for the face and 42 for both hands together, 553 in total.

Landmarks are points that mark the body, such as shoulders, elbows, wrists, fingertips, and the eyes and corners of the mouth. HIROMEIRO calculates joint angles from where these points are and makes the avatar take the same posture. The conversion from points to joints is published under the name mediapipe_receiver_cpp.

Two apps: estimation and display

HIROMEIRO is split into two parts: an app that reads motion (the tracking software) and the main program that displays the avatar. The read results reach the main program over the network using UDP on port number 38013. The flow is as follows.

Step	What happens	Where
1. Camera	Webcam or phone, captured at 1280×720	Tracking software (Windows / Android / Python)
2. Estimation	Reads points of body, face and hands with the MediaPipe Holistic Landmarker	Tracking software
3. Send	Within the same PC or the same network, UDP, port 38013	Network
4. Display	Turns it into joint angles and moves the avatar (Unreal Engine, VRM)	Main program (Windows only)

UDP is a simple way of sending that throws data at the other end one piece after another. A port number is a number that tells which app on a PC incoming data is for. 38013 written in hexadecimal is 947d, the same sequence as 947D-Tech, the name the developer uses for its domain and GitHub account.

On the reason for the split, the official announcement says it is to divide the processing load. You can leave the reading to a phone or another PC and do only the display on a Windows PC. If you run both on the same PC, the manual says it works with the destination left at 127.0.0.1 (the address that points to itself).

Is the camera video sent to the internet?

Since this is software that points a camera at you, you wonder whether the video is sent somewhere. I read the public source of the Windows tracking app (mediapipe_transceiver_app) and checked three things.

• The AI model (holistic_landmarker.task) and the components that run it are loaded from files placed inside the app. The lines that load them from an online distributor were disabled as comments (src/capture.js)
• package.json describes a step that downloads the model file while the app is being built and puts it into the app
• The part that handles communication (src-tauri/src/lib.rs) uses only UDP sending, and I saw no code that connects to other servers. What it sends is numbers such as the positions of the read points

As far as the source goes, the AI estimation is completed inside the PC and the results reach only the destination you specify. Its structure differs from something like ChatGPT, where video goes to an AI in the cloud, and it looks like an example of AI running locally. But what I read was the public source. I did not inspect the communication of the distributed executables on this blog.

The same source also contained a setting to estimate on the GPU (delegate: “GPU"). Next to it is a note that estimating on the GPU causes an error when outputting facial expression values (blendshapes). That agrees with the manual’s statement that expressions are calculated by another method.

Which OS does it run on? Windows, Android, Mac, Linux

The manual’s operating environment is written separately for the main program and the tracking software. The main program needs a Windows PC, and the tracking software runs on any of a Windows PC, Android, or a PC that runs Python.

Role	Systems it is said to run on	How to get it
Main program (displays the avatar)	Windows PC only. A graphics card is recommended	Just unzip the distribution. No installation needed
Tracking (Windows version)	Windows 11. Other OSes may work, as stated	Bundled in the same zip as the main program
Tracking (Android version)	Android phones. Snapdragon 8 series or equivalent processing power recommended	Get the source and build it yourself in Android Studio
Tracking (Python version)	Checked on Ubuntu 24.04. With Python and uv, other OSes may work, as stated	Get the source. For developers, and said to be fairly difficult

The official test environment was a desktop PC with Windows 11, Ryzen 7 5700X, 32GB of memory and a GeForce RTX 4060 Ti (16GB). It does not say how far it runs on a lighter PC than that.

What is in the distribution?

The official release is a single file, HIROMEIRO-v20260917.zip, about 412MB. Checking the list of files inside the zip (177 items), the main executables are these three (the zip also contains an Unreal Engine prerequisite installer and the shipping executable inside the main program).

File	Role
Windows/HIROMEIRO.exe	The main program (the Unreal Engine app under Binaries/Win64)
mediapipe_transceiver_app.exe	The Windows tracking app
mediapipe_receiver_utility.exe	The app for recording and playing back motion

All are for Windows; there were no executables for Mac or Linux. The Android app (apk) is not bundled either, so on a phone you build it yourself.

If you only have a Mac or Linux?

A Mac or Linux PC might be able to take the role of reading motion. Linux is officially confirmed with the Python version, and Mac falls under “other OSes may work". Either way you need a separate Windows PC for the main program that displays the avatar.

There is no official way to run the main program on anything but Windows. According to the FAQ, the main program uses Unreal Engine, so the source cannot be published in full because of licensing. Building your own Mac or Linux version also looks out of reach for now.

What camera and how much space do you need?

Item	Official description
Camera on a PC	An ordinary webcam. Tested with a Logitech (Logicool in Japan) C920n. The Python version’s description says a UVC-type camera
Camera on a phone	The phone’s own camera (front and back can be switched). Checked on Xperia 1 IV and Galaxy S26
Capture size	Fixed at 1280×720 in the Windows version’s source
Values you enter	The camera’s diagonal field of view (catalog value) and the camera’s orientation
Room size	The whole body must be in frame. Roughly 2m x 2m

UVC (USB Video Class) is a common rule for USB cameras that lets a PC treat them as cameras without installing a separate driver. In the Windows tracking app you pick, from a list, a camera the OS recognizes.

Why enter field of view and orientation

Video from one camera is flat, so depth cannot be known as it is. By entering how widely the lens sees (field of view) and which way the camera faces, it works out the body’s position in three dimensions. If the camera sits level, it is said to work with the orientation left at its initial value.

The orientation is entered as three numbers that express which way gravity points (a gravity vector). When the bottom of the image is the floor it is (0, 9.8, 0); when the camera is turned on its side and the left of the image is the floor, it is (9.8, 0, 0). The Android version uses the phone’s sensor to decide portrait or landscape.

Phone processing power

The Android README recommends a Snapdragon 8 series or better phone. With the 8 series even somewhat older devices run comfortably, while with the 7 series there were devices, even new ones, that ran heavily. The Torque G07 was named as an example it does not recommend.

Also, the Android version processes only while its screen is in front and stops when it goes to the background. While in use you need to leave the app open.

How does it compare with other camera-only full-body software?

There is other software that moves an avatar from camera video. Similar ones are TDPT (ThreeDPoseTracker), XR Animator and Webcam Motion Capture. As a reference I also lined up Sony’s mocopi, which uses body-worn sensors. All values are from each official page, and I have not run them on this blog.

Software / device	Price	What you need	Points read	Leg movement	OS
HIROMEIRO	Free	One camera	Body 33, face 478, both hands 42 (output of the MediaPipe Holistic Landmarker it uses)	Read	Main program Windows only. Reading on Windows / Android / Python
TDPT (Windows version)	Free	One camera	Body 24	Read	Windows (iOS version also exists)
XR Animator	Free (CC BY-NC-SA 4.0)	One webcam	Not stated (uses MediaPipe)	Read	Web browser; Windows / macOS / Linux apps
Webcam Motion Capture	Paid plan exists (the reading features can be tried free)	One webcam	Not stated	Feature list goes up to upper body and hands	Windows 10 or later / macOS Catalina or later
mocopi (reference, sensor type)	$449 at its US launch (Sony Electronics)	6 body-worn sensors + a phone	6 sensor points	Read	Phone app

Of the four camera-based tools, the three that cost nothing and read as far as the legs are HIROMEIRO, TDPT and XR Animator. Comparing the number of body points, leaving out face and hands, the 33 in the model HIROMEIRO uses is 9 more than TDPT’s 24. More points does not always mean finer reading, and accuracy cannot be compared by the number of points alone.

The big difference is the OS on which the avatar can be displayed. HIROMEIRO’s main program is Windows-only. XR Animator also runs in a web browser, and Webcam Motion Capture supports Mac. The sensor-based mocopi does not use a camera, in exchange for the cost of the hardware ($449 at US launch).

What PC specs do the vendors give?

Software	Specs the vendor gives
HIROMEIRO	Test environment = Ryzen 7 5700X, 32GB memory, GeForce RTX 4060 Ti (16GB). A graphics card is recommended
TDPT (Windows version)	Reference values = High: Core i9, 16GB, RTX 4090 / Medium: Core i7, 16GB, RTX 2070 Super / Low: Core i5, 4GB, no GPU (only the light reading setting works)

HIROMEIRO does not say how it runs on a lighter PC. For a PC without a graphics card, TDPT was the only one among those compared that gives an official guide.

What to watch before using HIROMEIRO

Bugs the vendor lists

The manual lists two bugs confirmed as of 14 September 2026.

• The Windows tracking app cannot output facial expression values because of MediaPipe’s behavior. The avatar’s expressions themselves are said to move because they are calculated another way
• The Android version stops tracking when the face is hidden. If you want to turn your back to the camera, you are directed to use other tracking software

The Python version is written as running all features stably. But it is meant to be operated by commands, and you must do the camera calibration yourself.

Will you show up on the stream?

The Windows tracking app has a “privacy mode" that keeps the camera video off the screen, and it is on from the start. If you turn it off to check the video, the manual calls for taking great care that you don’t appear on the stream. There is no point in an avatar stream if your real face gets out.

A note on gamepad input

Using a gamepad is recommended for operating the main program. Buttons move the viewpoint, reset the avatar’s position, and open and close the settings menu.

The thing to watch is that it accepts gamepad input even when the main program’s window is not selected. If you use the gamepad to play a game, as in a gameplay stream, you are directed to disable the gamepad in the settings menu.

Price, and what you may do

According to the FAQ, HIROMEIRO is free. Earning income from streaming with it is also allowed. What is prohibited is redistributing the main program and using it to slander others.

The main program is not open source, but the tracking software and the part that turns points into joint angles are published under the Apache-2.0 license. Bug reports and requests are accepted as GitHub issues, with a note that they cannot always be acted on.

How to decide whether to try HIROMEIRO

From what I looked into so far, I sorted out who it suits and who needs care. All are guides based on the official descriptions and the source, not results confirmed by running it on this blog.

If you are someone who	Guide
Has a Windows PC and a webcam and wants to move a VRM avatar with the whole body	A candidate. The main program and the tracking app are in the same zip, and one PC is enough
Doesn’t want camera video sent to a server on the internet	A candidate. In the public source, estimation runs on the PC and results are sent by UDP to the destination you specify
Wants to stream seated, or can’t find 2m x 2m of space	An upper-body-only mode is available
Has only a Mac or Linux PC	The main program is Windows-only, so you can’t use it as it is. You need a Windows PC for display
Wants to use a phone as the camera	Android only. No apk is distributed, so you need to build it in Android Studio. Snapdragon 8 series recommended
Wants to try on a PC without a graphics card	HIROMEIRO recommends a graphics card. The one with an official no-GPU guide is TDPT (light settings only)
Wants to display the avatar on a Mac alone	HIROMEIRO can’t be used. Among similar tools, XR Animator and Webcam Motion Capture support Mac
Often turns their back to the camera	The vendor says the Android version stops when the face is hidden. Choose the Windows or Python version

As a first step, connecting a webcam to a Windows PC and running the two bundled apps on the same PC takes the fewest steps. You can leave the destination at 127.0.0.1. A sound order seems to be: first check in upper-body-only mode whether the avatar follows your movement, and if you have the space, move on to the whole body.

Besides reading the whole body from a camera alone, there are methods using body-worn trackers. I looked into the characteristics of each and lined them up in a separate article.

Full-body and face tracking compared (in Japanese)

What I Haven’t Confirmed in This Article

• I did not run HIROMEIRO (the main program is Windows-only). Everything here is what the official documents, the public source and the distribution file list say
• I did not inspect the network communication of the distributed executables
• I did not verify the claims of the vendors in the comparison table (TDPT, XR Animator, Webcam Motion Capture, mocopi). They are each official page’s description

Summary: how HIROMEIRO works and where it runs

HIROMEIRO is free software that moves a VRM avatar with the whole body from a single webcam or phone camera. MediaPipe’s Holistic Landmarker reads body movement, and as far as the public source shows, that estimation happens inside the PC or phone. The main program that displays the avatar is Windows-only, and the app that reads motion comes in three kinds: Windows, Android and Python. Against similar camera-based software, its body point count is 33 against TDPT’s 24, and reading as far as the legs is shared with XR Animator and TDPT.

Before using it, check the space where the whole body fits (roughly 2m x 2m), entering the camera’s field of view and orientation, and the building and bugs of the Android version. Earning income from streaming is allowed, and redistributing the main program is prohibited.

AI that reads body movement from camera video is now running on PCs and phones too. Seen together with the text-writing AI this blog usually covers, the uses of AI that runs locally may be wider than you would think.

What runs by memory size (VRAM)

Sites I consulted

• HIROMEIRO official site and manual (Kushinada Kiko): quickstart, tracking software, recording software, FAQ
• HIROMEIRO official release (GitHub, v20260917)
• GitHub sources: mediapipe_transceiver_app, HolisticTransceiver, mediapipe_python_transceiver, mediapipe_receiver_cpp
• PR TIMES: announcement of the official HIROMEIRO release; MoguLive: article on the official release
• Holistic Landmarker (Google MediaPipe, number of points)
• ThreeDPoseTracker (Digital Standard official); XR Animator (GitHub); Webcam Motion Capture (official); mocopi QM-SS1 (Sony Store)