Diffractive Snapshot Spectral Imaging

• Diffractive Snapshot Spectral Imaging (DSSI) is a paradigm that encodes wavelength information onto sensor pixels using engineered diffractive elements.
• It employs varied designs—such as micro-structured filters, pseudo-random diffusers, and diffractive lenses—to map spectral variations into spatial patterns.
• DSSI enables efficient multispectral imaging with trade-offs in spectral resolution, sensitivity, and required computational reconstruction.

Searching arXiv for recent and foundational DSSI papers to ground the article. Diffractive Snapshot Spectral Imaging (DSSI) denotes a family of snapshot spectral imaging architectures in which a diffractive optical element encodes wavelength-dependent information into a single sensor measurement, so that a three-dimensional spatio-spectral scene can be recovered computationally or, in some designs, routed directly to designated sensor pixels without digital reconstruction. Across the literature, DSSI includes transmissive diffractive filters mounted in close proximity to monochrome sensors, diffuser-based lensless systems with on-sensor spectral filter arrays, diffractive-lens and photon-sieve systems that encode wavelength into controlled blur, angular-spectral lensless imagers, and multi-layer diffractive optical networks that create a virtual spectral filter array at the output plane (Wang et al., 2017, Monakhova et al., 2020, Wang et al., 2017, Mengu et al., 2022, Majumder et al., 2024).

1. Conceptual scope and system classes

DSSI is not a single hardware template but an optical-computational paradigm. In one widely cited form, a micro-structured diffractive filter replaces the usual Bayer filter and converts spectral information to a spatial code on the sensor pixels, after which calibration and regularization-based linear algebra are used to compute the multi-spectral image (Wang et al., 2017). In another form, a thin pseudo-random phase diffuser is placed close to a sensor overlaid by a tiled spectral filter array, and the hyperspectral volume is recovered by solving a sparsity-constrained inverse problem with sub-super-pixel resolution (Monakhova et al., 2020). Diffractive-lens implementations instead rely on a wavelength-dependent focal length so that each spectral component is encoded as a distinct blur or focus state, with reconstruction formulated as sparse recovery or learned inversion (Kar et al., 2019, Oktem et al., 2020).

A distinct branch of the literature removes numerical inversion almost entirely. Diffractive optical network-based multispectral imagers are trained to route pre-determined spectral channels onto spatially repeating pixel subsets, thereby creating a virtual spectral filter array at the output image field-of-view and converting a monochrome focal plane array into a multispectral imaging device without spectral filters or image recovery algorithms (Mengu et al., 2022). Related diffractive processors have also been trained to perform multispectral quantitative phase imaging in a snapshot, again with passive dielectric layers and a monochrome detector (Shen et al., 2023).

This diversity makes one terminological point important: “snapshot” does not imply a uniform recovery strategy. Some DSSI systems use optimization-based matrix inversion or deep unfolding; others perform feed-forward spectral routing; still others, such as diffractive-lens systems, may require a small stack of measurements unless implemented with a multi-lens array or a fast programmable diffractive element that superimposes multiple PSF states in a single exposure (Kar et al., 2019). A plausible implication is that DSSI is best understood as a wavelength-to-space encoding framework rather than a single camera topology.

2. Optical encoding mechanisms

The central optical principle in DSSI is the deliberate engineering of wavelength-dependent point-spread functions or spectral routing patterns. In the compact single-shot multi-spectral video-imaging camera, a transmissive diffractive filter mounted at $d\approx0.5\,\rm mm$ from a monochrome CMOS sensor is implemented as a 2-D “super-lattice” of period $P=18\,\mu\rm m$, with square pixels of side $\Delta=3\,\mu\rm m$ and relief up to $h_{\max}=1.2\,\mu\rm m$. Its pseudo-random height profile is chosen to maximize spectral–spatial decorrelation of the transmitted wavefront, and light from each object point and wavelength produces a wavelength-dependent, spatially shifted PSF whose centroid remains at the geometrical image location, so conventional blur is replaced by coded spectral blur (Wang et al., 2017).

Lensless diffuser-based DSSI uses a different encoding mechanism. In Spectral DiffuserCam, each point in the world maps through a thin pseudo-random phase diffuser to a wavelength-invariant caustic pattern, while an overlaid tiled spectral filter array applies a narrowband mask $m(\mathbf p,\lambda)$ at each pixel. The measurement therefore combines spatial multiplexing by the diffuser PSF $h(\mathbf p-\mathbf r)$ with spectral masking by $m(\mathbf p,\lambda)$, yielding a flexible design that can use contiguous or non-contiguous spectral filters (Monakhova et al., 2020).

Diffractive-lens systems encode wavelength into focus diversity. For a photon sieve of outer diameter $D$ and smallest hole diameter $\Delta$, the first-order focal length is

$f(\lambda)=\frac{D\,\Delta}{\lambda},$

so each wavelength is sharply imaged at a different plane while other bands appear defocused. In this class of systems, the chromatic focal shift is the encoding mechanism that converts spectral content into a superposition of wavelength-dependent blurs (Oktem et al., 2020). The earlier compressive spectral imaging architecture using diffractive lenses combines such dispersion with a coded aperture and a monochrome detector, and models the diffractive lens by a wavelength-dependent PSF derived from the photon-sieve pupil function (Kar et al., 2019).

All-optical diffractive networks extend the same idea into multi-layer passive processing. In the visible-band diffractive optical network, $P=18\,\mu\rm m$0 phase-only diffractive layers spanning $P=18\,\mu\rm m$1 are optimized so that at each target wavelength the interference pattern concentrates on a designated sub-array of output pixels while minimizing leakage to other pixels (Mengu et al., 2022). In the HD snapshot diffractive spectral imaging system, a periodic diffractive filter array fabricated by nanoimprint lithography is placed $P=18\,\mu\rm m$2 above a cover-glass-removed sensor; each wavelength then produces a unique local spatial shift of order a few pixels, which encodes spectral information into a single grayscale diffractogram (Majumder et al., 2024).

3. Forward models and calibration strategies

Despite substantial hardware variation, DSSI is usually expressed as a linear forward operator mapping a spectral cube to a 2D sensor image. In vectorized notation, one common form is

$P=18\,\mu\rm m$3

where $P=18\,\mu\rm m$4 is the discretized multispectral datacube, $P=18\,\mu\rm m$5 is the raw sensor image, and $P=18\,\mu\rm m$6 is the system matrix whose columns are calibrated PSFs for specific spatial and spectral coordinates (Wang et al., 2017). In diffuser-based notation, the same principle appears as

$P=18\,\mu\rm m$7

and in discrete form as $P=18\,\mu\rm m$8 (Monakhova et al., 2020). HD DSSI similarly uses $P=18\,\mu\rm m$9, with $\Delta=3\,\mu\rm m$0 described as highly sparse and block-Toeplitz with circulant blocks due to local space-invariance (Majumder et al., 2024).

Calibration is correspondingly central. The 2017 sensor-adjacent diffractive-filter camera constructs $\Delta=3\,\mu\rm m$1 by scanning a $\Delta=3\,\mu\rm m$2 pinhole over a $\Delta=3\,\mu\rm m$3 object-space grid with $\Delta=3\,\mu\rm m$4 step size, while a supercontinuum laser and tunable bandpass filter select 25 bands from $\Delta=3\,\mu\rm m$5 to $\Delta=3\,\mu\rm m$6; the resulting $\Delta=3\,\mu\rm m$7 measurements populate the columns of the system matrix, which is then fixed for subsequent imaging (Wang et al., 2017). The HD DFA system instead calibrates the PSF within one $\Delta=3\,\mu\rm m$8 unit cell and at 10 off-axis sensor locations, then interpolates neighboring PSFs to form spatially varying kernels across the full frame (Majumder et al., 2024).

Other variants use more specialized calibration. Wang and Menon’s angular-spectral lensless imager builds a matrix of “angular-spectral point-spread-functions” by scanning both incidence angle and wavelength; the system discretizes $\Delta=3\,\mu\rm m$9 over 31 angles from $h_{\max}=1.2\,\mu\rm m$0 to $h_{\max}=1.2\,\mu\rm m$1 and 61 wavelengths from $h_{\max}=1.2\,\mu\rm m$2 to $h_{\max}=1.2\,\mu\rm m$3 (Wang et al., 2017). The Aperture Diffraction Imaging Spectrometer (ADIS) emphasizes a weakly calibration-dependent data form: calibration requires monochromatic PSF measurement at each band and the spectral response of each mosaic pixel, while lateral mask shifts preserve PSF amplitudes because the induced phase factor changes only the field phase, not its magnitude (Lv et al., 2023).

4. Inverse reconstruction and computational inference

For systems that do not optically demultiplex the spectral channels, reconstruction is an ill-posed inverse problem. Early DSSI work used Tikhonov regularization. In the compact multispectral video-imaging camera, the estimate is obtained by

$h_{\max}=1.2\,\mu\rm m$4

implemented by SVD and a filtered pseudoinverse, with an empirically chosen $h_{\max}=1.2\,\mu\rm m$5 and a direct MATLAB solver taking $h_{\max}=1.2\,\mu\rm m$6 ms/frame (Wang et al., 2017).

Subsequent systems adopted stronger priors. Spectral DiffuserCam solves

$h_{\max}=1.2\,\mu\rm m$7

where the regularizers are anisotropic 3D TV and the nuclear norm; optimization is performed with FISTA in 500–1000 iterations, requiring 12–24 min on an RTX 2080Ti (Monakhova et al., 2020). Diffractive-lens systems use sparse-recovery formulations with transform-domain $h_{\max}=1.2\,\mu\rm m$8 priors and ADMM or alternating minimization, exploiting FFT-based convolution to avoid explicit matrix construction (Kar et al., 2019, Oktem et al., 2020).

Deep reconstruction has become a distinct research direction. ADIS reconstructs $h_{\max}=1.2\,\mu\rm m$9 from $m(\mathbf p,\lambda)$0 through a deep-unfolding architecture called the Cascade Shift-Shuffle Spectral Transformer (CSST), which alternates a gradient step with a learned denoiser guided by a lightweight quantitative parameter estimation network (Lv et al., 2023). The 2025 Diffractive Deep Unfolding (DDU) framework argues that many unfolding methods used on related tasks are not fully compatible with DSSI because of its distinct optical encoding mechanism; it derives an analytical solution for the data-fidelity term in the frequency domain and combines it with a network-based initialization strategy and a learnable prior module, reporting improved performance with comparable parameter counts and computational complexity (Zhuge et al., 7 Jul 2025). In parallel, the integrated near-infrared framework combines a designed DOE with NIRSA-Net, a U-shaped Transformer that reconstructs 31 spectral bands from a single encoded image (Ma et al., 20 Aug 2025).

An important misconception is therefore that DSSI always requires heavy post-processing. That is true for many compressive systems, but diffractive optical network-based imagers explicitly aim for computation-free, power-efficient and polarization-insensitive forward operation, with only standard demosaicing at readout (Mengu et al., 2022). Conversely, the computational burden remains a defining limitation for high-dimensional inverse DSSI, particularly for HD reconstructions and low-light scenes (Majumder et al., 2024).

5. Reported performance and trade-offs

Reported operating points span visible, visible–infrared, near-infrared, extreme ultraviolet, and terahertz regimes. The table below summarizes representative systems and quantitative claims directly reported in the literature.

System	Optical front end	Reported operating point
Compact multispectral video imager (Wang et al., 2017)	Micro-structured diffractive filter near monochrome CMOS	25 bands, $m(\mathbf p,\lambda)$1, $m(\mathbf p,\lambda)$2 spectral resolution; over $m(\mathbf p,\lambda)$3 spatial-resolution enhancement
Spectral DiffuserCam (Monakhova et al., 2020)	Diffuser + tiled spectral filter array	64 channels, $m(\mathbf p,\lambda)$4; recovered peaks within $m(\mathbf p,\lambda)$5; two-point spectral resolution $m(\mathbf p,\lambda)$6
HD DFA system (Majumder et al., 2024)	Nanoimprinted DFA $m(\mathbf p,\lambda)$7 above sensor	25 bands, $m(\mathbf p,\lambda)$8, $m(\mathbf p,\lambda)$9 pixels; $h(\mathbf p-\mathbf r)$0 FWHM
NIR DOE + NIRSA-Net (Ma et al., 20 Aug 2025)	Designed DOE + learned reconstruction	31 bands, $h(\mathbf p-\mathbf r)$1, $h(\mathbf p-\mathbf r)$2 resolution; $h(\mathbf p-\mathbf r)$3 PSNR, $h(\mathbf p-\mathbf r)$4 SSIM

The 2017 compact camera reported spectral resolution of $h(\mathbf p-\mathbf r)$5 within the visible band $h(\mathbf p-\mathbf r)$6, object-space spatial resolution $h(\mathbf p-\mathbf r)$7, an MTF cutoff of $h(\mathbf p-\mathbf r)$8 cycles/mm, and a $h(\mathbf p-\mathbf r)$9 improvement over a conventional Bayer-filtered camera under identical optics; it also showed that a $m(\mathbf p,\lambda)$0 raw frame could be traded between $m(\mathbf p,\lambda)$1 pixels $m(\mathbf p,\lambda)$2 bands and $m(\mathbf p,\lambda)$3 pixels $m(\mathbf p,\lambda)$4 bands by recalibrating on coarser grids without changing hardware (Wang et al., 2017). The same system was extended to visible–infrared imaging by adding one $m(\mathbf p,\lambda)$5 calibration band, with crosstalk below $m(\mathbf p,\lambda)$6 for non-adjacent bands and about $m(\mathbf p,\lambda)$7 for nearest neighbors (Wang et al., 2017).

Spectral DiffuserCam emphasized the decoupling of spatial and spectral sampling: because the spectral filters are overlaid on a diffuser-based lensless encoder rather than acting as direct sampling pixels, the system reported two-point spatial resolution of $m(\mathbf p,\lambda)$8 super-pixels, multi-point resolution of $m(\mathbf p,\lambda)$9 super-pixels, and reconstruction SNR above $D$0 in bright scenes, with exposures of $D$1 under laboratory illumination (Monakhova et al., 2020). The HD DFA system prioritized field size and throughput, reporting $D$2 spatial pixels over 25 channels, average transmission exceeding $D$3, $D$4 more sensitivity compared to filter-based multispectral cameras, and average reconstruction SNR exceeding $D$5 in bright regions (Majumder et al., 2024).

Performance trade-offs recur throughout the field. In diffractive-filter cameras, larger sensor support can be exchanged against either more spectral bands or wider field of view (Wang et al., 2017). In HD DFA systems, pushing below $D$6 spectral resolution would demand larger DFA–sensor gaps or taller micro-pixels, increasing spatial blur (Majumder et al., 2024). In diffractive-lens systems, the number of distinct PSF states $D$7 trades off SNR and reconstruction error against acquisition complexity (Kar et al., 2019). In diffractive optical networks, increasing the number of spectral bands is accompanied by a trade-off with output spatial resolution unless the number of diffractive degrees of freedom is increased (Mengu et al., 2022).

6. Applications, misconceptions, and current directions

DSSI has been demonstrated across a wide application range. The HD DFA platform reported proof-of-concept experiments in biological tissue classification, food quality assessment, and simulated stellar photometry; ex vivo chicken lung and trachea were separated by an LDA classifier with classification accuracy $D$8 on held-out pixels, a strawberry-aging regressor achieved $D$9 with mean absolute error $\Delta$0 days, and a logistic classifier distinguishing fruit younger than 5 days from older fruit achieved AUC $\Delta$1 (Majumder et al., 2024). Wang and Menon proposed compact angular-spectral imaging for fast phenomena, remote sensing, medical diagnostics, and machine vision, while ADIS demonstrated single-frame hyperspectral video of a combustion flame at 35 fps (Wang et al., 2017, Lv et al., 2023). Diffractive optical networks were positioned for biomedical microscopy, remote sensing, environmental monitoring, terahertz imaging, security screening, and material characterization (Mengu et al., 2022).

Several misconceptions can be addressed directly from the literature. DSSI is not necessarily lensless: some systems use a standard photographic lens or an imaging lens plus ultra-thin aperture mask (Wang et al., 2017, Lv et al., 2023). It is not necessarily tied to absorptive filters: the compact diffractive-filter camera, the HD DFA system, and diffractive optical network-based imagers all emphasize the absence of lossy absorptive filters, with corresponding gains in sensitivity or photon throughput (Wang et al., 2017, Majumder et al., 2024, Mengu et al., 2022). Nor is DSSI restricted to contiguous visible-band spectroscopy: Spectral DiffuserCam explicitly supports non-contiguous spectral filter choices, visible–IR operation has been shown with an added $\Delta$2 band, and near-infrared operation from $\Delta$3 to $\Delta$4 has been demonstrated at $\Delta$5 resolution (Monakhova et al., 2020, Wang et al., 2017, Ma et al., 20 Aug 2025).

Current research directions are split between optical design and reconstruction. On the optics side, the literature points to multi-height designs, non-periodic patterns, adaptive or scene-specific diffractive filter arrays, extension to ultraviolet or near-infrared materials, and joint end-to-end optimization of diffractive elements with reconstruction networks (Majumder et al., 2024, Ma et al., 20 Aug 2025). On the algorithmic side, the 2025 DDU framework identifies reconstruction as comparatively underdeveloped relative to optical front-end design and proposes DSSI-specific deep unfolding with an analytical fidelity solver and network-based initialization (Zhuge et al., 7 Jul 2025). This suggests that the field is moving from proof-of-concept inverse solvers toward reconstruction methods that are explicitly matched to diffractive encoding physics, rather than borrowed from generic compressive imaging.