An hour into the headset,
your eyes start to ache.
Not from the brightness. From something the screen is quietly asking your eyes to do.
The mismatch is the vergence–accommodation conflict — a known cause of eye-fatigue in near-eye displays. [High confidence]
Every screen you've used
fakes depth the same way.
A phone, a monitor, a headset — each is one flat plane painted to look like a world. Your eyes always knew; at arm's length it never mattered.
Strap one to your face, and it starts to.
A screen your eyes
could relax into.
Lean a little, and see around the side of things. Shift your focus, and the far things soften on their own — the way real light behaves.
What if depth didn't
have to be faked?
A hologram doesn't paint
a picture of light.
It rebuilds the light.
On a chip the size of a fingernail, a few million cells each nudge the timing of a passing wave. Bend enough of them and the light curls, crosses, and comes to a focus in open air — a scene made of nothing else.
So we tell it to match the picture we want.
Aim for the brightness the eye expects, tune the chip until that image appears — and it does. Sharp, complete, convincing.
Get the picture right, and you've got your hologram. Mostly.
But is a flat, right-looking picture
really depth?
Shift your focus, and real light softens the distance.
The picture-matched hologram can't. Its blur comes out harsh and grainy, speckled with noise — because the depth was never really there. A right-looking picture is still just a picture.
Light has a second half.
Not its brightness — its timing: the exact instant each tiny wave peaks. Your eye can't see it, so for years almost no one bothered to get it right. But that hidden timing is where depth hides.
Tune only what you can see, and you throw the depth away.
They started tuning the timing too.
Steer the chip toward a wave whose timing holds together, not just one that looks bright. Describe the scene as a fine mesh, so perspective and overlap come built-in. And learn the quirks of the real lab from a camera, so the maths matches the messy optics.
Mesh rendering + phase-aware optimisation + a camera-learned model of the optics, co-designed. [High confidence]
A sharper picture — and
a real +7 dB from the phase.
Reconstruction quality across methods — higher is better.
And the same quality came 2.39× faster than brute-forcing extra focus planes.
PSNR in dB, axis from 12 · Meng et al. 2026, Table 1 + Fig 3. [High confidence]
Five ways to make a hologram.
Each bar above is a different method. Here's what each actually does — and where it falls short.
Double-phase
Analytic model
Depth-map
Their method, 2D-trained
Their method, the best
the depth comes alive.
The same rabbit, now holding real volume in the air: soft where it should be soft, solid where it should be solid. Light you could lean into.
The three moves, up close.
How they actually rebuilt the light. Swipe through, or keep scrolling — the story holds without it.
POS-WRP
The chip sends light straight out, but a scene is drawn in perspective. POS-WRP "twists" the mesh and its baked textures into the chip's flat space, so perspective, shading and occlusion all survive into the wavefront instead of being flattened away.
PAGD
Older methods match only brightness. Phase-aware gradient descent steers toward a complex target — brightness and timing. The right phase isn't unique (a whole field can shift by 2π), so it centres the phase and optimises toward a consistent one, not a single "true" answer.
Learned optics
Real lenses, lasers and chips never behave like the textbook. A small neural model, trained on what a camera actually sees, learns the lab's quirks and folds them into the maths — closing the gap between the clean simulation and the messy bench.
Flying Planes
To teach that model, they built a dataset: 1,000 random 3D scenes, textured from DIV2K, each paired with its true complex wavefront. Training on real 3D fields — not flat ones — is what lets it handle tilted, off-axis light.
View-dependent light
A blazed grating bends light toward a chosen direction, writing the viewing angle into the phase. An iris at the Fourier plane acts like a pupil, picking out one viewpoint at a time — which is how the image can shift as you lean.
Lean, and look around it.
Scroll, and the near rabbit slides across the cup behind it — a small shift, but a real one. The objects sit at honest distances from each other, the way things do in a window.
Your eyes pick the depth.
The near rabbit is sharp; the far cup softens on its own — the smooth, natural blur that only real depth makes.
The continuous defocus the phase-aware method recovers, shown on a two-object scene (cf. the basketball + multi-depth captures, Fig 6–7). [High confidence]
The same eyes — finally easing.
Where they aim and where they focus land on the same place at last. No fight, no ache. Just a thing in the air, and a pair of eyes resting on it.
It still lives on a table of lasers.
One frame takes about 55 seconds to compute. The look-around is barely a degree and a half wide. And the whole thing sits on a bench you couldn't strap to your face.
54.96 s per frame · ≤1.5° view-dependent tilt · benchtop prototype. Meng et al. 2026. [High confidence]
What it's actually made of.
The prototype that captured every real image in this story.
Phase-only SLM
RGB laser
Eight focus planes
Camera + iris
One RTX 4090
55 s a frame
The hologram stopped
being the hard part.
What's left is time and silicon — making it fast, and making it small. The kind of problem the next ten years are made of.
How long until depth
stops being faked?
Mesh-represented and learning-empowered hologram synthesis for full 3D holographic displays
Xiangyu Meng · Wenbin Zhou · Yifan Peng — The University of Hong Kong
Nature Communications (2026) · open access. The eye-strain problem: Hoffman et al. 2008.