Towards Edge Holography via
Implicit Neural Representation & Compression

Hyunmin Ban, Wenbin Zhou, Yifan Peng
The University of Hong Kong (HKU)
IEEE Transactions on Visualization and Computer Graphics (IEEE VR), 2026
🎉 Best Paper Honorable Mention
Paper Supplementary Video (TBD)
Abstract overview figure

Abstract

Holographic displays offer realistic 3D visualization, but practical deployment remains challenging due to the high computational cost of computer-generated holography and the difficulty of efficiently compressing hologram data. To address this, we present a lightweight, display-aware framework that combines implicit neural representations (INRs) with camera-calibrated wave propagation for phase-only hologram generation and compression.

Our method represents holograms as continuous neural functions, enabling compact modeling of high-frequency holographic structure, and further improves deployment efficiency through quantization-aware training and entropy coding. Experiments on an unfiltered holographic display prototype show that the proposed INR-CGH achieves competitive 2D and 3D reconstruction quality, up to 11× compression with minimal degradation, and decoding speeds of at least 250 fps.

Captured 3D Planes

Zoomed-in captured videos illustrating depth-dependent reconstruction changes across focal planes.

System Pipeline

Our INR-CGH framework represents a phase-only hologram as a compact implicit neural function. On the server side, coordinates are input to the INR model and optimized through a camera-calibrated propagation model with quantization-aware training. The resulting parameters are then entropy coded for efficient transmission.

On the edge side, the compressed parameters are decoded and directly inferred into a phase-only hologram, enabling fast holographic display with high-fidelity reconstruction under unfiltered optical conditions.

System pipeline figure

Experimental Results

Hardware Prototype

We built a benchtop unfiltered holographic display prototype to validate the proposed method in real optical experiments. The setup uses a phase-only SLM, RGB laser illumination, relay optics, and a focus-tunable capture process for evaluating reconstructions across multiple depths.

Hardware prototype figure

Captured Results

We compare SGD-based and INR-based hologram generation in both 2D and 3D settings, for both hologram representation and compression. Each comparison block can be browsed interactively to inspect the differences between the two approaches.

2D Hologram Representation

2D Hologram Compression

3D Hologram Representation

3D Hologram Compression

BibTeX

@article{ban2026edgeholography,
  title   = {Towards Edge Holography via Implicit Neural Representation and Compression},
  author  = {Ban, Hyunmin and Zhou, Wenbin and Peng, Yifan},
  journal = {IEEE Transactions on Visualization and Computer Graphics},
  year    = {2026},
  note    = {IEEE VR 2026}
}