Neural Holography measured its fixes with a camera, not simulation

What you see

The Stanford Computational Imaging Lab's own project page shows the work captured through a benchtop near-eye display prototype, not a finished headset: a comparison figure on the page contrasts a classic algorithm's speckled, artifact-heavy reconstruction against the paper's own camera-in-the-loop result, imaged through the same optical rig. A viewer of that prototype would look through the display's eyepiece to see the holographic image; the page states the method reaches full-color, 1080p-resolution images at real-time rates, a claim about computed image quality captured by camera, not a report from a human wearing a finished product.

How it works

The paper, credited on the project page to Yifan Peng, Suyeon Choi, Nitish Padmanaban, and Gordon Wetzstein of Stanford University, and titled “Neural Holography with Camera-in-the-loop Training,” is computer-generated holography: a phase pattern is computed and displayed on a spatial light modulator, then illuminated coherently. Its distinguishing mechanism, per the project page's abstract, is “camera-in-the-loop training,” in which a physical camera photographs the display's actual output and that measurement is fed back to either optimize a hologram directly or train a neural network, called HoloNet in the paper, to model the real optical system's quirks rather than an idealized simulation of it.

Viewing conditions and limits

The paper's own comparison figure, in the PDF retrieved from the lab's site, reports specific image-quality numbers for its prototype: a classical Gerchberg-Saxton reconstruction at roughly 16 decibels PSNR and 0.52 SSIM, against about 19 decibels and 0.68 SSIM for the paper's camera-in-the-loop optimization, and about 17.6 decibels and 0.64 SSIM for the faster neural-network variant, HoloNet, all measured on the authors' own prototype display. Those figures describe one lab rig under the authors' own test conditions; the source does not state how the numbers would transfer to a different optical design or a mass-produced headset.

What it is not

This is not a claim of a shipped consumer headset: the project page lists the SIGGRAPH Asia 2020 technical paper and a separate, shorter SIGGRAPH 2020 Emerging Technologies demonstration, both research outputs with downloadable PDFs and source code, not a product announcement. It is also not a fix for the display's fundamental optics; the method corrects for a mismatch between a simulated and an actual optical system through measurement and learning, it does not change the near-eye display's field of view or form factor.

• Are the reported PSNR and SSIM figures from the authors' own prototype, and under what illumination or content were they measured?
• Does “real-time” refer to the full camera-in-the-loop optimization or to the trained neural-network inference alone?
• Is the demonstration described a SIGGRAPH technical paper, an emerging-technologies exhibit, or a product roadmap?

Taken on its own terms, the paper's contribution is a measured, camera-verified correction to computer-generated holography's image quality on one prototype, a narrower and more checkable claim than “neural holography solves holographic displays.”