Significance Holographic 3D displays reconstruct the complete wavefront information of 3D objects and are therefore regarded as a key direction in the development of next-generation naked-eye 3D displays. However, traditional hologram generation methods struggle to achieve a balance between holographic reconstruction quality and computational speed. Among these challenges, the suppression of speckle noise remains a central issue in holographic 3D displays. Arising from coherent illumination and random phase distributions, speckle noise represents a primary factor that affects the quality of holographic reconstructions. In recent years, deep learning has provided a new computational paradigm for generating low-noise holograms. Therefore, this study systematically reviews deep learning-based methods for low-noise hologram generation and provides a detailed discussion of the fundamental principles, technical advancements, and future trends of data-, model-, and hybrid-supervised methods, with the aim of providing a reference for the further optimization of low-noise hologram generation methods. Progress In recent years, low-noise hologram generation methods based on deep learning have experienced rapid development, and innovations in data-, model-, and hybrid-supervised methods have driven advances in holographic 3D displays. Data-supervised methods depended on paired datasets of target images and holograms, and low-noise holograms were generated by learning the mapping relationships between them. Early data-supervised studies focused on applying deep neural networks to hologram generation, in which tensor holography enabled the real-time synthesis of low-noise holograms, whereas diffraction-engineered networks further enhanced the visual realism of holographic 3D displays. One noteworthy strategy directly generated low-noise holograms from 2D target images, which simplified the construction of training datasets. However, data-supervised methods remained highly dependent on large-scale, high-quality datasets, and generalization ability and model performance were closely associated with dataset quality. Model-supervised methods embedded differentiable physical models into the training process, which avoided reliance on paired datasets and enabled unsupervised or self-supervised learning of low-noise holograms by minimizing the intensity difference between reconstructed and target images. Hardware-based model-supervised methods calibrated optical aberrations through system feedback and achieved high-quality color holograms. Software-based model-supervised methods introduced complex-valued neural networks or lightweight architectures and generated holograms with high accuracy and high frame rates. In addition, model-supervised methods were applied to address occlusion problems in complex 3D scenes. However, model-supervised methods required repeated diffraction propagation calculations, which resulted in high training costs. Hybrid-supervised methods integrated the advantages of data and model supervision and generated high-fidelity, low-noise holograms by combining data and physical losses. For example, the two-stage hybrid training method first pretrained networks on high-quality datasets and performed unsupervised fine-tuning using physical models, surpassing the performance ceiling of data-supervised methods. Under data-scarce conditions, hybrid-supervised methods based on nonpaired learning achieved object-to-hologram mapping through cycle-consistency structures. Hybrid-supervised methods were also applied to optical system optimization, propagation error compensation, and hologram generation under low spatial coherence. However, hybrid-supervised methods still encountered challenges associated with complex training processes and limited stability. Conclusions and Prospects Although deep learning has driven the rapid development of holographic 3D displays, several challenges remain: 1) in data-supervised methods, the training efficiency of deep neural networks remains low, with a strong dependence on training datasets; 2) in model-supervised methods, the speed of hologram generation does not yet satisfy the requirements for real-time interaction; 3) in hybrid-supervised methods, the performance gains achieved by simply increasing the depth or width of neural networks are limited. Therefore, future research is expected to proceed in the following directions: 1) transfer learning methods can be applied to reduce the training time required for data-supervised methods and alleviate dependence on datasets; 2) lightweight neural networks can be designed to improve the generation speed of model-supervised methods; 3) novel architectures, such as Transformer models, self-attention mechanisms, and diffusion models, can be explored to extend the performance boundaries of neural networks. With continued improvements in training efficiency, inference speed, and reconstruction quality, holographic 3D displays are expected to facilitate transformative applications in medical imaging, industrial inspection, education, and entertainment.
RivensonY, WuYichen, OzcanA.Deep learning in holography and coherent imaging[J].Light(Science & Applications),2019,8:85. doi:10.1038/s41377-019-0196-0
[2]
LiuKexuan, WuJiachen, HeZehao,et al.Progress of learning-based computer-generated holography[J].Chinese Journal of Liquid Crystals and Displays,2023,38(6):819-828. doi:10.37188/CJLCD.2023-0081
CaoLiangcai, HeZehao, LiuKexuan,et al.Progress and challenges in dynamic holographic 3D display for the metaverse(Invited)[J].Infrared and Laser Engineering,2022,51(1):20210935.
GerchbergR W, SaxtonW O.A practical algorithm for the determination of phase from image and diffraction plane pictures[J].Optik,1972,35:237-246. doi:doi:http://dx.doi.org/
[7]
ChakravarthulaP, PengYifan, KollinJ,et al.Wirtinger holography for near-eye displays[J].ACM Transactions on Graphics,2019,38(6):1-13. doi:10.1145/3355089.3356539
[8]
PengYifan, ChoiS, PadmanabanN,et al.Neural holography with camera-in-the-loop training[J].ACM Transactions on Graphics,2020,39(6):1-14. doi:10.1145/3414685.3417802
ZhangYixin, ZhangMingkun, LiuKexuan,et al.Progress of the computer-generated holography based on deep learning[J].Applied Sciences,2022,12(17):8568. doi:10.3390/app12178568
[13]
YuXuan, ZhangHaomiao, ZhaoZhe,et al.On the use of deep learning for computer-generated holography[J].iScience,2025,28(5):112507. doi:10.1016/j.isci.2025.112507
[14]
ShimobabaT, BlinderD, BirnbaumT,et al.Deep-learning computational holography:A review[J].Frontiers in Photonics,2022,3:854391. doi:10.3389/fphot.2022.854391
LeeJ, JeongJ, ChoJ,et al.Deep neural network for multi-depth hologram generation and its training strategy[J].Optics Express,2020,28(18):27137. doi:10.1364/oe.402317
[17]
ZhengHuadong, HuJianbin, ZhouChaojun,et al.Computing 3D phase-type holograms based on deep learning method[J].Photonics,2021,8(7):280. doi:10.3390/photonics8070280
[18]
KhanA, ZhangZhijiang, YuYingjie,et al.GAN-holo:Generative adversarial networks-based generated holography using deep learning[J].Complexity,2021,2021:6662161. doi:10.1155/2021/6662161
[19]
ShiLiang, LiBeichen, KimC,et al.Towards real-time photorealistic 3D holography with deep neural networks[J].Nature,2021,591(7849):234-239. doi:10.1038/s41586-020-03152-0
[20]
YangD, SeoW, YuH,et al.Diffraction-engineered holography:Beyond the depth representation limit of holographic displays[J].Nature Communications,2022,13:6012. doi:10.1038/s41467-022-33728-5
[21]
ChangChenliang, WangDi, ZhuDongchen,et al.Deep-learning-based computer-generated hologram from a stereo image pair[J].Optics Letters,2022,47(6):1482-1485. doi:10.1364/ol.453580
[22]
LiuShengchi, ChuDaping.Deep learning for hologram generation[J].Optics Express,2021,29(17):27373. doi:10.1364/oe.418803
[23]
YolalmazA, YüceE.Comprehensive deep learning model for 3D color holography[J].Scientific Reports,2022,12:2487. doi:10.1038/s41598-022-06190-y
[24]
ChangChenliang, DaiBo, ZhuDongchen,et al.From picture to 3D hologram:End-to-end learning of real-time 3D photorealistic hologram generation from 2D image input[J].Optics Letters,2023,48(4):851-854. doi:10.1364/ol.478976
[25]
QinHaifeng, HanChao, ShiXuan,et al.Complex-valued generative adversarial network for real-time and high-quality computer-generated holography[J].Optics Express,2024,32(25):44437. doi:10.1364/oe.543792
[26]
FangQiwei, ZhengHuadong,et al.Diffraction model-driven neural network with semi-supervised training strategy for real-world 3D holographic photography[J].Optics Express,2024,32(26):45406. doi:10.1364/oe.538649
[27]
YuXuan, TengZhilin, FanXuhao,et al.IncepHoloRGB:Multi-wavelength network model for full-color 3D computer-generated holography[J].Opto-Electronic Advances,2025,8(10):250130. doi:10.29026/oea.2025.250130
[28]
FuJiahui, WanWenqiang, WangYunrui,et al.High-quality hologram generation based on a complex-valued hierarchical multi-fusion neural network[J].Chinese Optics Letters,2025,23(9):090501. doi:10.3788/col202523.090501
[29]
LiuQingwei, ChenJing, YaoYongwei,et al.Frequency aware high-quality computer-generated holography via multilevel wavelet learning and channel attention[J].Optics Letters,2024,49(19):5559-5562. doi:10.1364/ol.532049
[30]
YuGuangwei, WangJun, YangHuan,et al.Asymmetrical neural network for real-time and high-quality computer-generated holography[J].Optics Letters,2023,48(20):5351. doi:10.1364/ol.497518
[31]
ZhongChongli, SangXinzhu, YanBinbin,et al.Real-time 4K computer-generated hologram based on encoding conventional neural network with learned layered phase[J].Scientific Reports,2023,13:19372. doi:10.1038/s41598-023-46575-1
[32]
YanXingpeng, LiuXinlei, LiJiaqi,et al.Generating multi-depth 3D holograms using a fully convolutional neural network[J].Advanced Science,2024,11(28):2308886. doi:10.1002/advs.202308886
[33]
ZhangYuanzhe, ChengDewen, BeijingN N L,et al.Real-time multi-depth holographic display using complex-valued neural network[J].Optics Express,2025,33(4):7380. doi:10.1364/oe.551943
[34]
WangDi, LiZhaosong, ZhengYi,et al.Liquid lens based holographic camera for real 3D scene hologram acquisition using end-to-end physical model-driven network[J].Light:Science & Applications,2024,13:62. doi:10.1038/s41377-024-01410-8
[35]
LiZhaosong, LiuChao, LiXiaowei,et al.Real-time holographic camera for obtaining real 3D scene hologram[J].Light:Science & Applications,2025,14:74. doi:10.1038/s41377-024-01730-9
[36]
ShiLiang, LiBeichen, MatusikW.End-to-end learning of 3D phase-only holograms for holographic display[J].Light:Science & Applications,2022,11:247. doi:10.1038/s41377-022-00894-6
[37]
GopakumarM, LeeG Y, ChoiS, et al. Full-colour 3D holographic augmented-reality displays with metasurface waveguides[J].Nature,2024,629(7911):791-797. doi:10.1038/s41586-024-07386-0
[38]
SunXiaotian, XiongChencen, RenZhenbo,et al.Unpaired learning for digital holographic reconstruction and generation[J].Optics Express,2025,33(2):3414. doi:10.1364/oe.551211
[39]
ZhangWanbin, FengYijian, RenYanchen,et al.Cycle-consistent physical constraint deep learning for incoherent self-interference digital holographic reconstruction[J].Optics and Lasers in Engineering,2026,196:109385. doi:10.1016/j.optlaseng.2025.109385
[40]
YuanGanzhangqin, ZhouMi, PengYifan,et al.Error-compensation network for ringing artifact reduction in holographic displays[J].Optics Letters,2024,49(11):3210-3213. doi:10.1364/ol.519519
[41]
TongXin, XuRenjun, LiuKe,et al.A deep-learning approach for low-spatial-coherence imaging in computer-generated holography[J].Advanced Photonics Research,2023,4:2200264. doi:10.1002/adpr.202200264