HoloMamba: Efficient CGH for Full-Color Video

• HoloMamba is a lightweight neural architecture that uses an asymmetric Mamba-UNet with multiscale spatio-temporal modeling to enable high-speed, full-color holographic video generation.
• It integrates Spectrum-Guided Depth Division Multiplexing (SGDDM) to broaden the phase spectrum and mitigate color crosstalk in depth-division multiplexing setups.
• HoloMamba achieves state-of-the-art performance with 35.44 dB PSNR, 0.95 SSIM, and 267 FPS while drastically reducing network parameters and memory usage.

Searching arXiv for the HoloMamba paper and a few related CGH baselines to ground the article in current papers. HoloMamba is a lightweight asymmetric Mamba-UNet for video computer-generated holography (CGH), introduced as the computational core of a high-speed, full-color video CGH pipeline in “High-Speed FHD Full-Color Video Computer-Generated Holography” (Zhang et al., 27 Aug 2025). It is paired with Spectrum-Guided Depth Division Multiplexing (SGDDM) to address two limitations in practical holographic video: color fidelity degradation in depth-division multiplexing (DDM) caused by over-smoothed phase holograms with narrow angular spectra, and the computational inefficiency of frame-by-frame optimization methods that ignore spatial-temporal correlations across video sequences. Within that system, HoloMamba is designed to exploit temporal redundancy explicitly while maintaining low parameter count, low memory usage, and FHD full-color holographic video generation at over 260 FPS (Zhang et al., 27 Aug 2025).

1. Position within full-color video CGH

HoloMamba is not presented as an isolated neural architecture, but as one component of a system-level solution for high-speed full-color holographic video (Zhang et al., 27 Aug 2025). The underlying problem formulation begins from two constraints. First, learning-based CGH models often produce over-smoothed phases with narrow angular spectra. In Fourier or angular spectrum method terms, the phase’s spectral energy concentrates near low frequencies, so little energy remains in the high-frequency region. Since the diffraction angle satisfies

$\sin\theta=\lambda\sqrt{f_x^2+f_y^2},$

low spatial frequencies diffract at small angles, leading to a large depth of field (DOF). In a full-color DDM setup, where RGB channels are assigned to different focal depths, a large DOF causes the color images to bleed into one another and produce depth replicas or color crosstalk (Zhang et al., 27 Aug 2025).

Second, existing frame-by-frame optimization methods typically optimize frames independently. The paper identifies this as a failure to exploit the strong spatial-temporal correlations between consecutive frames, producing computationally inefficient solutions and weaker temporal coherence in video holography (Zhang et al., 27 Aug 2025). HoloMamba addresses the computational and temporal modeling side of this problem, while SGDDM addresses the optical consequence of over-smoothed phase prediction.

A common misconception is that minimizing reconstruction loss alone is sufficient for high-quality full-color display. The paper argues the opposite: methods that optimize only image fidelity can worsen full-color display quality because they do not explicitly enforce a sufficiently broad angular spectrum or adequate depth separation between color channels (Zhang et al., 27 Aug 2025). This distinction is central to understanding why HoloMamba is coupled to SGDDM rather than proposed as a standalone replacement for prior CGH networks.

2. Optical motivation and the role of SGDDM

The optical model introduced alongside HoloMamba is Spectrum-Guided Depth Division Multiplexing. Its purpose is to modulate the hologram in the frequency domain so that the phase distribution is encouraged to carry more spectral diversity and therefore a wider angular spectrum (Zhang et al., 27 Aug 2025). The target intensity is written as

$I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$

Here $H(f_x,f_y)$ is the angular spectrum propagation transfer function, and $\mathcal{M}_C$ is a channel-specific circular binary mask centered at $(c_x,c_y)$ with radius $r$.

Physically, shifting the mask off-axis is equivalent to adding a linear phase ramp in the spatial domain:

$\phi_{\text{eq}(x,y)=\phi_{\text{ori}(x,y)+2\pi(c_x x+c_y y),$

which boosts the phase gradient,

$\nabla \phi_{\text{eq}(x,y)=\nabla \phi_{\text{ori}(x,y)+2\pi(c_x,c_y),$

and therefore widens the angular bandwidth (Zhang et al., 27 Aug 2025). A wider bandwidth implies a larger numerical aperture and a smaller DOF. The paper states the approximate relation

$$\mathrm{DOF}_i \approx \frac{\lambda_i}{\mathrm{NA}_i^2},$$

and uses the depth-separation condition

$$|z_i-z_j|>\frac{1}{2}\left(\frac{\lambda_i}{\mathrm{NA}_i^2}+\frac{\lambda_j}{\mathrm{NA}_j^2}\right),\quad i\neq j,$$

to guide the learning of the RGB masks so that adjacent color planes remain sufficiently separated (Zhang et al., 27 Aug 2025).

Because the mask must be binary in the forward pass while remaining trainable, the paper uses the differentiable surrogate

$I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$0

with $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$1 initialized at $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$2, doubled after each epoch, and capped at $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$3. This progressively sharpens the soft mask into a physical binary aperture during training (Zhang et al., 27 Aug 2025).

The significance of SGDDM for HoloMamba is direct. HoloMamba improves computational efficiency and temporal modeling, but SGDDM compensates for the tendency of learning-based CGH to produce smooth phases whose spectra are too concentrated at low frequency for robust DDM. This suggests that the paper views optical spectrum control and sequence modeling as complementary requirements rather than interchangeable ones.

3. Architecture of HoloMamba

HoloMamba takes a video clip as a spatio-temporal tensor

$I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$4

with RGB amplitudes as input and zero phase initialization, and predicts the phase hologram sequence $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$5 (Zhang et al., 27 Aug 2025). Its backbone is a three-level asymmetric U-Net. The encoder uses cascaded 3D convolutions for patch embedding and downsampling, whereas the decoder uses residual depthwise convolution, upsampling, and skip connections.

The core building block is the Multi-Receptive Field Interaction (MRFI) module. MRFI splits channels into two branches: a local CNN branch and a global Mamba branch (Zhang et al., 27 Aug 2025). The local branch captures wavefront detail through multi-scale 3D convolutions with kernel sizes $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$6 and $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$7. The global branch reshapes the feature tensor into a 1D sequence over $I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$8 and applies a Bidirectional Spatial-Temporal Mamba (BSTMamba) block for long-range dependency modeling with linear complexity.

The layer structure is summarized as

$I_{\textrm{target}(x,y)=\left|\mathcal{F}^{-1}\left\{H(f_x,f_y)\,\mathcal{M}_C(f_x,f_y)\,\mathcal{F}\{e^{i\phi(x,y)}\}\right\}\right|^2, \quad C\in\{R,G,B\}.$9

followed by an FFN. Inside BSTMamba, the model employs separate forward and backward selective state-space scans:

$H(f_x,f_y)$0

and fuses them with a gating signal through element-wise modulation and linear projection (Zhang et al., 27 Aug 2025).

The scan strategy traverses height, width, and time in sequence, and then scans the reversed sequence as well. According to the paper, this bidirectional design allows the model to access both past and future temporal context while preserving intra-frame spatial continuity (Zhang et al., 27 Aug 2025). A plausible implication is that HoloMamba is tailored not merely to generic video prediction, but specifically to phase hologram synthesis, where coherent propagation structure extends across both spatial and temporal dimensions.

4. Local-global modeling and training objective

The paper attributes HoloMamba’s performance to the balance between local and global modeling (Zhang et al., 27 Aug 2025). The local CNN branch is responsible for recovering fine spatial detail in the wavefront, while the Mamba branch provides efficient long-sequence modeling that avoids the quadratic cost associated with Transformers and the limited receptive field of plain 3D CNNs. The hybrid design is explicitly described as better balanced than using only CNNs or only global sequence models.

Training uses the hybrid loss

$H(f_x,f_y)$1

The MSE term measures reconstruction fidelity in the propagated intensity domain, while the focal frequency loss (FFL) emphasizes high-frequency discrepancies:

$H(f_x,f_y)$2

with frequency-aware weight

$H(f_x,f_y)$3

The paper characterizes this choice as sensible for CGH because high-frequency content is precisely what tends to disappear in over-smoothed phase predictions (Zhang et al., 27 Aug 2025).

Implementation details reported in the paper are limited but specific. The model is trained in PyTorch 2.0 with Adam at learning rate $H(f_x,f_y)$4 for 50 epochs on an NVIDIA RTX 8000, using DIV2K for full-color images and DAVIS2017 for video (Zhang et al., 27 Aug 2025). These details locate HoloMamba within a supervised learning workflow rather than an iterative per-frame optimization regime.

5. Reported performance and empirical behavior

On FHD benchmarks, HoloMamba achieves 35.44 dB PSNR and 0.95 SSIM, outperforming HoloNet (29.69/0.90), CCNN-CGH (32.01/0.92), Divide-Conquer-and-Merge (32.83/0.93), and CVMNet (30.28/0.90) (Zhang et al., 27 Aug 2025). The paper also reports 44.7K parameters and 708 MB memory, compared with HoloNet’s 2868.7K parameters and 7984 MB memory.

For video CGH, the central throughput result is 267 FPS for FHD output, reported as more than 2.6× faster than the prior state-of-the-art, Divide-Conquer-and-Merge (Zhang et al., 27 Aug 2025). Temporal consistency is also quantified: the warping error is 0.022, compared with 0.048 for DCM and 0.031 for CVMNet. The paper states that flow visualizations show smoother inter-frame trajectories and reduced flicker.

The following table organizes the principal benchmark figures reported for HoloMamba and the comparison methods named in the paper.

Method	Reported quality	Reported efficiency
HoloMamba	35.44 dB PSNR, 0.95 SSIM	44.7K parameters, 708 MB memory, 267 FPS
HoloNet	29.69 dB PSNR, 0.90 SSIM	2868.7K parameters, 7984 MB memory
CCNN-CGH	32.01 dB PSNR, 0.92 SSIM	—
Divide-Conquer-and-Merge	32.83 dB PSNR, 0.93 SSIM	slower than HoloMamba by more than 2.6×
CVMNet	30.28 dB PSNR, 0.90 SSIM	warping error 0.031

The paper’s qualitative interpretation is that HoloMamba improves both reconstruction fidelity and temporal coherence while remaining computationally lightweight (Zhang et al., 27 Aug 2025). A plausible implication is that these results are particularly relevant for deployment scenarios in which FHD holographic video must be generated continuously rather than as isolated still frames.

6. Ablation evidence and design interpretation

The ablation studies are used to justify both the choice of backbone and the local-global decomposition in MRFI (Zhang et al., 27 Aug 2025). Replacing HoloMamba with a 3D CNN yields 13,976 MB memory, 17 FPS, and 24.78 dB PSNR, which the paper characterizes as a poor efficiency-quality balance. A ViT variant is reported as even less practical, with 202,348 MB memory, 0.20 FPS, and 29.01 dB PSNR.

The MRFI ratio study further indicates that the balance between local and global branches matters materially. Pure local modeling (0:1) yields only 19.44 dB, while pure global modeling (1:0) improves quality to 29.03 dB but increases memory to 2964 MB. The chosen 0.8:0.2 split is reported as the best trade-off, giving 28.74 dB with 2734 MB memory (Zhang et al., 27 Aug 2025).

These ablations support two claims made in the paper. First, neither a purely convolutional model nor a purely global sequence model is optimal for this task. Second, HoloMamba’s contribution is not simply the inclusion of Mamba blocks, but their integration into an asymmetric U-Net with explicit local-global partitioning. This suggests that the architecture is tuned to a CGH-specific operating point in which phase prediction requires both local texture recovery and long-range spatio-temporal dependency modeling.

The SGDDM ablation is equally important. Without spectral guidance, the output phase remains too smooth, the spectrum stays concentrated at low frequency, and severe color crosstalk appears under DDM. With SGDDM, the spectrum broadens and full-color reconstructions become clean and faithful (Zhang et al., 27 Aug 2025). This evidences that HoloMamba alone does not resolve the optical failure mode associated with narrow-band phase predictions.

7. Significance, scope, and limitations of interpretation

Within the paper’s framing, HoloMamba contributes to a broader redefinition of video CGH as a jointly optical and sequence-modeling problem rather than a per-frame image-to-phase regression problem (Zhang et al., 27 Aug 2025). Its significance lies in showing that explicit spatio-temporal modeling with a lightweight asymmetric Mamba-UNet can improve reconstruction quality, reduce memory consumption, and increase throughput for FHD full-color holographic video.

At the same time, the paper’s claims should be interpreted in the context of the full pipeline. The reported high-fidelity full-color display without compromise in frame rate depends on the combination of HoloMamba with SGDDM, not on HoloMamba in isolation (Zhang et al., 27 Aug 2025). A related misconception would be to treat HoloMamba as solving color crosstalk directly. The evidence presented instead indicates that HoloMamba addresses efficiency and temporal blindness, whereas SGDDM addresses the spectral and DOF conditions required for DDM.

The paper’s overall conclusion is therefore system-level: SGDDM fixes the optical issue created by learning-based over-smoothing, and HoloMamba fixes the computational inefficiency and temporal blindness of prior video CGH methods. Together they enable high-speed FHD full-color holographic video with strong fidelity, low memory, and real-time-or-better throughput (Zhang et al., 27 Aug 2025).