A holographic display has an impossible job. It must reconstruct an entire light field from a finite number of spatial light modulator (SLM) pixels, and it must do it using scalar-wave diffraction from coherent illumination alone.
The consequence is speckle: the grainy, shimmering interference noise that is the signature artefact of computer-generated holography (CGH). Two decades of work have attacked it by making the reconstruction more accurate — better phase retrieval, better optical models, better hardware.
HoOPOE — Holography-Optimized Perceptual ObjectivE function — takes a different line, and it is the kind of move worth paying attention to. The paper, by Ethan Chen, Xi Peng, Alexandre Chapiro, Dario Seyb, Michael Grundmann, Carl Marshall, Sushant Kondguli, Praneeth Chakravarthula and Yuhao Zhu, is to appear at SIGGRAPH Asia 2026.
Optimising for the observer, not the image
Most CGH pipelines optimise a loss between the simulated reconstruction and a target image. Mean squared error, or a perceptual metric borrowed from image compression or super-resolution research.
The borrowing is the problem. Those metrics were tuned on the artefacts that their domains produce — blocking, ringing, blur — and speckle is not one of them. A loss that does not model how speckle is perceived has no particular reason to suppress it where it will be noticed.
HoOPOE is instead CGH-specific and perceptually aware. It accounts for the human visual system’s sensitivity to spatio-temporal frequencies and colour, and — the part that makes it an engineering contribution rather than a modelling one — its parameters are tuned to maximise its performance as a loss function for CGH, not to predict perceptual quality in the abstract.
That distinction matters. A metric that correlates well with human judgement is not automatically a good optimisation target; it may be flat, or jagged, or expensive exactly where the optimiser needs it to be informative. Tuning the parameters for the job of being a loss is a different objective from tuning them for accuracy as a predictor.
The practical effect: the objective hides speckle from human vision rather than attempting to eliminate it from the reconstruction. Noise that falls where the eye is insensitive is noise you do not have to spend SLM capacity suppressing.
The result
The authors used HoOPOE to generate holograms for mesh-derived focal stacks, then ran a user study benchmarking those holograms against existing loss functions.
HoOPOE produced the most visually pleasing results, with users preferring it 68% of the time on average over the next best objective function.
A preference study is the right instrument here, and it is also the honest limit of the claim. There is no fidelity number being beaten, because fidelity is not what is being optimised. The paper is arguing that for a display whose only consumer is a human eye, the eye’s verdict is the measurement — and against the previous best, that verdict lands 68–32.
Why this is interesting beyond holography
Holographic and light-field displays have been a decade away for about three decades, and the reason is rarely the one people expect. The maths of diffraction is settled. The bottleneck is that a finite SLM cannot reproduce a continuous light field, so every practical system is making a lossy choice about which errors to accept.
Framing that choice as a perceptual optimisation problem, with a loss function tuned for the specific artefact the medium produces, is a more tractable question than making the error go away. It is also the same logic that made perceptual audio coding work: not a better reproduction of the waveform, but a careful decision about what the listener will never hear.