Standard video codecs treat all signals as living on a linear number line.

Phase signals live on a circle, so conventional difference computation, clipping, and RD optimization all produce incorrect results.

We introduced two foundational operations and integrated them into HEVC’s RDO, residual computation, and sample reconstruction processes:

• Shorter Circular Difference: Resolves the distance ambiguity between two phase values by always selecting the shorter arc on the phase circle.
• Circular Clipping: Wraps out-of-range values back onto the valid phase interval instead of hard-clipping to the boundary.

Applied across seven holographic test sequences, the proposed method achieves –63.3% BD-Rate (phase domain) and –65.5% BD-Rate (NR domain) compared to standard HEVC. VVC extensions with SAO and deblocking filters are submitted.

Conventional approaches treat hologram generation and compression as separate pipelines. NHVC is the first unified end-to-end neural holographic video codec, jointly optimizing generation and compression within a single scalable architecture.

Key design choices:

• Deformable convolution for large receptive fields and implicit temporal motion handling.
• Strong band-limiting to suppress high-order diffraction noise.
• Scalable architecture: a single model supports image/video generation and compression, switchable at inference time.

NHVC achieves >33 dB NR-PSNR quality, outperforming cascaded baselines by –75.8% BD-Rate (image) and –75.6% BD-Rate (video).

Phase holograms are shift-invariant: adding a global constant (mod 2π) leaves the 3D reconstruction unchanged. This means two physically identical holograms can show arbitrarily large phase-domain error — making direct phase comparison unreliable.

PDA (Phase Distribution Alignment) maps each hologram to a unique canonical form before computing error, resolving this ambiguity. PDA-based metrics show consistently higher correlation with NR-domain metrics across 13 noise types and 5 strength levels.

PDA has three demonstrated applications:

• Quality evaluation: reliable phase-domain quality assessment without computationally expensive numerical reconstruction.
• Pre-processing for conventional codecs: –14% BD-Rate with HEVC, –25.41% BD-Rate with VVC phase hologram extension.
• Training loss for neural models: +0.37 dB generation quality and –11% BD-Rate vs. NHVC, with faster training convergence.

Existing neural hologram generators are optimized for a fixed propagation distance.

Changing the distance, wavelength, or pixel pitch requires full retraining — impractical for real deployments where hardware specifications vary.

PaD (Propagation as Data) treats the propagation kernel as input data rather than a fixed physical operation baked into the architecture. A dedicated Kernel Encoder ingests the propagation kernel at inference time, enabling a single trained model to generalize to arbitrary propagation conditions without retraining.

Architecture highlights: pre-trained LDM image encoder for stable feature extraction; FFC (Fast Fourier Convolution) for feature–kernel fusion; enlarged receptive field for global scene structure modeling.