· 5 min
Vayu AI Anatomy: a 3D atlas of the human body in your browser
I went looking for a free, interactive 3D atlas of the human body and could not find one. So I built it: 3,753 anatomically named structures across 13 systems, every one of them clickable, fully open-source under CC-BY-SA 4.0.
For the last few weeks I have been looking for a free, interactive 3D atlas of the human body. I could not find one.
What is on the web today is two things. There are YouTube spin-around videos — gorgeous, but you cannot click anything. And there are the commercial atlases like BioDigital, Visible Body, ZygoteBody — beautiful, but locked behind seat licenses and trial paywalls. A medical student or a curious teenager who wants to point at a structure and ask "what is this, and what's it part of?" has nowhere to go.
So I built the version I actually wanted. anatomy.vayuai.ai
3,753 anatomically named structures across 13 systems. Bones, muscles, fascia, vessels, nerves, organs, glands, lymphatics. Click anywhere on the body and a right-hand panel resolves the full anatomical hierarchy — where the structure sits, what it is part of, what it contains. Search the same names in the box at the top and the camera will fly to whatever you typed.
No download. No login. No paywall. Free under CC-BY-SA 4.0, hosted statically, runs in any modern browser.
Why this had to exist
The strange thing is that the open-data world has the assets.
Z-Anatomy ships about 3,600 retopologized meshes under CC-BY-SA, covering skeletal, muscular, cardiovascular, nervous, lymphatic, and respiratory. The SPL Abdominal Atlas out of Brigham and Women's Hospital ships another 94 CT-derived organs under a comparably open license — the liver, kidneys, stomach, pancreas, intestines, adrenals, all segmented from a real patient's CT by physicians at the Surgical Planning Laboratory. BodyParts3D out of the University of Tokyo seeded the whole lineage two decades ago.
These assets are out there. They are just locked inside .blend files
designed for Blender, not for browsers. The missing piece was the
bridge.
So I built it.
The system in a minute
Click a preset and a single system materializes over a ghosted skeleton. Cardiovascular brings up the heart and 680 named arteries and veins. The vessels are anatomically positioned — coronaries on the heart, femorals down the legs, radials and ulnars in the forearms. Click any vessel and you get the full hierarchy.
Pull up Muscular and the body fills with named fascia and bellies. Every superficial sheath, every belly, every named insertion. Hover the neck and the cervical fascia lights up by name.
Toggle Nervous and the body becomes a wireframe of cranial and spinal nerves. 246 of them, traceable from the brainstem down to the left sciatic and every motor branch beyond.
Or strip everything back to a single region. Abdominal Organs brings up the SPL anatomy — CT-derived, physician-segmented, sitting in the cradle of the ribcage. Click the descending colon and the camera flies in.
The third gesture is Layers Depth. Drag the slider and the body peels — skin to muscle to vessels to organs to skeleton. No clipping plane, no toggling fifteen checkboxes. One slider, one fluid reveal.
And the browse tree drills down nine levels deep wherever Z-Anatomy has the data. Nervous system → Brain → Brainstem → Midbrain → Cerebral peduncle → Tegmentum → Grey matter → Oculomotor complex → left accessory nucleus of oculomotor nerve. Every one of those terms is searchable, every one of them clickable, every one of them resolves to a mesh on the body.
How it is built
A short answer.
Stack. Z-Anatomy and SPL data on the asset side. A headless
Blender pipeline that walks the Z-Anatomy .blend archive,
converts the ~950 curve objects (vessels and nerves are stored as
splines, not meshes) into proper triangulated geometry, reconstructs
the TA2 hierarchy from parent chains, and exports one Draco-compressed
glTF per system. A second pipeline handles the SPL atlas — VTK
polydata to triangulated mesh to an affine transform that aligns CT
space with Z-Anatomy world space.
A Node script then dedupes across sources, prefers SPL organs over Z-Anatomy where they overlap, drops misaligned bundles, and emits a manifest the browser reads at boot.
One command rebuilds the whole atlas: npm run pipeline. Re-runnable,
cached, auditable.
Frontend. Vite, React, TypeScript, Three.js, @react-three/fiber,
@react-three/drei, Zustand for state. Total app JavaScript: around
300 KB gzipped. Initial paint with just the skeletal bundle: about 6
MB. Time to first interactive paint: under three seconds on a typical
connection. The full atlas — every system, every bundle —
is 31 MB Draco-compressed.
Click handling is delegated one event listener per ~700-mesh system bundle, not one per mesh. With 3,000+ structures loaded at once, interaction stays smooth.
Data sources. Three of them:
- Z-Anatomy — 3,659 retopologized structures across six systems. CC-BY-SA 4.0.
- SPL Abdominal Atlas (Brigham and Women's Hospital, Surgical Planning Laboratory) — 94 CT-derived organ meshes, physician-segmented. 3D Slicer License Section B.
- BodyParts3D / Anatomography (DBCLS, University of Tokyo) — preserved upstream in the attribution chain. CC-BY-SA 2.1 JP.
All three are CC-BY-SA-compatible. The derivative ships under CC-BY-SA 4.0 with upstream attributions preserved.
Deploy. Static SPA on Vercel. No backend. No login. No analytics in the critical path.
What I left out
No realistic skin texture and no face. Open CC-BY-SA datasets do not include either. I tried a synthetic skin shell derived from the muscular envelope and it looked rough, so it is gone. No clothes either — there is no open clothing asset library.
The reproductive system is sparse (eight entries). That is a known gap in Z-Anatomy's coverage and one of the first things I would like to backfill.
The SPL Abdominal Atlas is one patient's CT. When you toggle Digestive on, the liver and intestines sit in roughly the right place in the Z-Anatomy body — but not pixel-perfectly. They are two different people fitted together with an affine transform. Good enough for teaching, not good enough for surgical planning.
If we add the SPL Brain Atlas next, the brain structure count jumps from 248 to around 600. If we license commercial models, the visual ceiling moves up dramatically. Both are tracked.
What I learned building it
Three things, briefly.
Open anatomical data is far better than it gets credit for. The Z-Anatomy archive is genuinely beautiful work, kept up under CC-BY-SA by a small team. The SPL atlas is real-patient CT segmented by surgeons. The reason most teams pay for commercial atlases is not that the open ones are worse — it is that nobody has bothered to put a browser viewer on top of them.
The hard part is not the modelling, it is the plumbing. Headless Blender, curve-to-mesh conversion, TA2 hierarchy reconstruction, cross-source dedup, affine alignment, manifest emission. The viewer is a couple of weekends. The pipeline is most of the work.
Branding the boring parts. A cream viewport, a warm dark chrome, gold accents, a Fraunces display face, an Inter UI face, the flower-of-life logo in the corner. The atlas does not look like 3D Slicer or Z-Anatomy — it looks like Vayu AI. That choice took as long as some of the engineering.
Why I shipped this
I work at the intersection of mechanical engineering, industrial systems, and applied machine learning. That intersection is where most real-world AI value lives, and most of it is still untapped because the data plumbing is harder than the modelling.
This project is a calling card for the kind of work I do through Vayu AI. Mostly it is a tool I actually wanted to use, that did not exist, that I could build in a weekend.
If it is useful to you, take it. If you find a rough edge, tell me. If you work in medical education, simulation, surgical planning, or patient communication and want a custom version of this for your specific domain, my email is at the bottom of vayuai.ai/contact.
The source is at github.com/RamRam633/Vayu-AI-Anatomy. CC-BY-SA 4.0 — fork it, modify it, even commercialize it. Just preserve attribution and stay open.
Educational use only. Not for clinical or diagnostic purposes.
The future belongs to those who design systems not just with intelligence, but with intention.