Conditional diffusion model for high quality unseen hologram synthesis

Unseen hologram synthesis enables the generation of phase-only holograms (POHs) directly from user-specified instructions, eliminating reliance on amplitude inputs and enabling unprecedented creative flexibility for immersive mixed reality applications. However, existing non-end-to-end methods rely on amplitude-domain generative models trained on massive but physically inconsistent datasets, introducing phase cumulative errors that degrade hologram quality with noise and artifacts. Although end-to-end methods based on autoencoders avoid phase error accumulation, their entangled latent representations hinder the manipulation of holographic attributes. This paper proposes holo-LDM, an end-to-end conditional diffusion model for unseen hologram synthesis directly from user provided semantic and physical constraints with only 2 KB parameters. Physical model-driven complex-valued UNets are employed to generate speckle-suppressed holographic datasets. For precise holographic attribute control, cross-attention transformers dynamically align labels with holographic features during the denoising process, which enables the synthesis of novel holograms with user-specified attributes during inference. Simulation and optical experiments confirm that holo-LDM is capable of synthesizing high-quality holograms across diverse semantic categories with accurate control of diffraction depth. Compared to existing diffusion-based non-end-to-end baselines, holo-LDM achieves over 30% improvement in terms of Fréchet inception distance (FID), while reducing training costs by an order of magnitude.