Contents

I was scrolling X when a post stopped my hand mid-swipe.

A 3D model of a human body, right there in the browser. Toggle the skeleton, the muscles, the organs, look at it from any angle. Spin it around with a finger.

At a glance it looks like a demo video for some expensive anatomy software. Except this one was built using GPT-6 Astra.

Voxel art of Hanageruge pulling a part out of a 3D anatomy viewer in a browser and swapping the label to Japanese

"I want to poke at this in Japanese"

Seeing the original post, my first reaction was simply that I wanted to try it.

The second was that if you could check the part names in Japanese, it would be a lot more useful for clinicians and students here.

Going through the repository, the app code is MIT licensed, and the anatomy data it uses is BodyParts3D 4.0 under a licence that requires attribution. So as long as you keep to the conditions, you can take it, change it and publish it.

So I did the Japanese version.

→ human-atlas.pages.dev

To be clear about what I did: I did not build a 3D anatomy viewer from scratch. The idea and the implementation were already public, so I got to try the next step. I'm genuinely grateful to the original poster for showing this kind of use and, more importantly, for publishing it in a form anyone can pick up.

What exactly got translated

Not just the part names. The whole interface.

The systems panel in English on the left and in Japanese on the right, with the same item names and counts

In the systems panel on the left, Systems becomes 系統, Skeleton becomes 骨格, Muscles becomes 筋, Sensory organs becomes 感覚器官. The control panel at the bottom too: Explode anatomy becomes 解剖モデルを分解, Assembled becomes 組み立て. Even the hint text in the corner — drag to orbit, pinch to zoom, tap to inspect — is in Japanese.

For anyone who wants the original wording, the EN button in the top right switches back.

The Japanese version of Human Atlas: the systems panel on the left with skeleton 296, muscles 402, arteries 639, the full-body 3D model in the centre, and the explode slider at the bottom

By the numbers: 2,234 modelled pieces across 15 systems. Skeleton 296, muscles 402, arteries 639, veins 404, and so on. All of that runs in a browser.

I put it on Cloudflare, not Vercel

The repository README says to import the project into Vercel, and it ships a Vercel config file. The original demo runs on Vercel too.

But this app builds down to a folder of static files (dist), so any static host will serve it. The README says as much: it can be served by another static host.

So I put it on Cloudflare Pages. The reason is simple: this blog runs on Cloudflare too. Keeping things in one place makes both the admin and the billing easier to follow.

None of which is a knock on Vercel. The point is that being able to choose where it lives comes from the build output being static files. That was a good call in the original design.

Astra did not "draw" the human body

Looking at this viewer, it's tempting to think AI can now generate accurate anatomical illustrations.

The actual mechanism is a bit different.

Astra did not imagine a human body and draw it. It read existing 3D data of the human body and assembled it into something you can operate in a browser. What the AI is doing isn't the drawing. It's the conversion and the moving parts around it.

That distinction matters to me. If an AI had imagined the anatomy, then however plausible it looks, you couldn't trust what's inside. If the source data comes from research institutes and is based on a real body, the story changes completely.

So what is the source data?

An anatomy database called BodyParts3D.

Then I found out the model itself came from Japan

Reading up on BodyParts3D, I got surprised a second time.

The team that developed and published this data was centred on Japanese research institutes.

A demo of the most talked-about AI in the world. Follow its foundation down, and you land on research that has been building up in Japan for years. "People here made something like this?" — the feeling was different from being impressed by Astra's performance.

And the deeper I dug, the clearer it became that BodyParts3D wasn't born as an anatomy teaching resource at all.

At the start of it were research into how radio waves from mobile phones affect the human body, and whole-body MRI scans of an actual Japanese man.

  • Why did a body model built for radio-wave research turn into an anatomy database with over 2,000 parts?
  • Who made it, using which sources, and how?
  • And how far can you trust it?

Digging into that comfortably outgrew one post, so it became Part 2.

→ Who actually made this 3D human body? Following the trail led to Japanese radio-wave research

Old research data, moving again because of AI

The most interesting thing here isn't that an AI generated medical knowledge from nothing.

It's that data specialists spent years building and publishing was read by an AI and turned into a shape that's usable today.

Until now, even when specialist 3D data was public, the number of people who could actually handle it was small. You had to understand the data format and be able to write rendering code before you were even at the starting line. AI is lowering that wall between the data and the people who want it, and fast.

A researcher leaves data behind. An educator turns it into teaching material. A clinician uses it to explain something to a patient. A student spins it around and learns. This viewer shows that possibility unusually clearly.

I don't think AI erases what specialists have accumulated. It takes knowledge and data that only a few people could handle, and turns it into something more people can use. That direction probably matters more, in the end.

Worth knowing before you poke at it

It's fun, so please do go and touch it. But it helps to know what kind of thing it is.

  • It's a viewer for education and study, not a tool for diagnosis or surgery (the original repository says so explicitly)
  • Inside is a standard anatomy based on one adult male body. It does not match everyone's body
  • The shapes are simplified so they run smoothly in a browser
  • Showing the 3D model means loading a fair amount of data, so give your connection a little room

How far you can trust the data is something Part 2 goes into properly.

Part 2 has the rest

This isn't only a story about how impressive a new AI is. It's about knowledge that has been building for years meeting new technology and starting to move again.

Part 2 covers how that foundation was made: cutting whole-body MRI into 2 mm cubes, the medical staff who spent more than three years classifying tissue by hand, and the question of how correct this data actually is.

→ Part 2: Who actually made this 3D human body? Following the trail led to Japanese radio-wave research

Part counts, system counts and licence terms were checked against the official repository and the BodyParts3D documentation on 6 September 2026.