Compressed Ultrafast Photography (CUP)

Updated 6 July 2026

CUP is a computational imaging method that captures non-repeatable ultrafast events by encoding a 3D spatiotemporal scene into a single 2D measurement using coded apertures and streak-camera temporal shearing.
It employs a forward model combining spatial encoding and temporal shearing with compressed sensing reconstruction algorithms, such as TV-based iterative solvers, to tackle the inherently ill-conditioned inverse problem.
CUP and its variants have enabled breakthroughs in applications ranging from dense plasma diagnostics to high-speed biological imaging by balancing temporal depth, spatial resolution, and reconstruction fidelity.

Searching arXiv for recent and foundational papers on Compressed Ultrafast Photography and close variants. Search query: "Compressed Ultrafast Photography CUP streak camera compressed sensing arXiv" Compressed Ultrafast Photography (CUP) is a single-shot computational imaging method that captures an ultrafast transient event by compressing a 3D spatiotemporal scene $I(x,y,t)$ into a single 2D measurement using coded aperture and streak imaging, then reconstructing the movie computationally (Guo et al., 2023). Within the broader family of compressed streak imaging (CSI), CUP is the best-known implementation of a framework that combines spatial coding with temporal shearing so that non-repeatable or difficult-to-repeat transients can be recorded in one exposure (Keppler et al., 2024). Its subsequent development has produced hyperspectral extensions, auxiliary-constraint architectures, and alternative optical encodings, while also clarifying the boundary between CUP proper and CUP-like single-shot ultrafast imaging systems that pursue the same objective through different mechanisms (Cheng et al., 27 May 2025).

1. Definition and lineage

CUP belongs to CSI, a class of systems in which a dynamic scene is spatially encoded by a coded aperture and then temporally sheared before being integrated on a 2D detector. In this lineage, CUP is the canonical streak-camera-based implementation. The general CSI formulation described in later work preserves this basic structure while modifying the way temporal and spatial diversity are introduced, or the way compression is distributed across the sensor (Keppler et al., 2024).

The principal motivation for CUP is the acquisition of nonrepeating ultrafast events. In regimes where conventional framing would require either event repetition or a prohibitive trade-off between frame rate and pixel count, CUP instead forms a compressed measurement whose inversion yields a time-resolved movie. This places CUP in the same broad methodological family as other ultrafast computational imaging systems, but its defining characteristic is the coupling of coded spatial modulation with streak-camera temporal shearing rather than ordinary high-speed readout or purely passive optical delay multiplexing (Li et al., 2021).

A persistent conceptual point is that CUP is not merely “very fast video.” Its identity is tied to a specific inverse problem: a 3D scene is not directly sampled frame by frame, but mapped into a compressed 2D observation that must be reconstructed computationally. That distinction becomes important when CUP is compared with related systems such as rolling-shutter compressive video, lensless ultrafast sensing, or passive time-to-space mapping architectures, which may be single-shot and ultrafast yet differ fundamentally in encoding physics and reconstruction philosophy (Antipa et al., 2019).

2. Forward model and inverse reconstruction

A standard CUP forward model is written as

$E(m,n)=TSCI(x,y,t),$

where $C$ is the spatial-encoding operator, $S$ is the temporal-shearing operator, and $T$ is the spatiotemporal-integration operator (Li et al., 2021). The measurement $E(m,n)$ is therefore a 2D detector image containing multiplexed information from the full spatiotemporal scene. The inverse problem is commonly posed in a compressed-sensing form, for example

$\arg \min \left\{ \|E - TSCI\| + \lambda TV(I) \right\},$

with $TV(I)$ imposing a total-variation prior (Li et al., 2021).

The same structure generalizes naturally to higher-dimensional variants. In hyperspectrally compressed ultrafast photography (HCUP), the target becomes a 4D data cube $I(x,y,t,\lambda)$ , and acquisition is expressed as

$E(x',y') = M T S C I(x,y,t,\lambda),$

where $E(m,n)=TSCI(x,y,t),$ 0 is the spatial encoding operator, $E(m,n)=TSCI(x,y,t),$ 1 the spectral shearing operator, $E(m,n)=TSCI(x,y,t),$ 2 the temporal shearing operator, and $E(m,n)=TSCI(x,y,t),$ 3 the spatial-temporal-spectral integration operator (Guo et al., 2023). This formulation makes explicit that CUP-style reconstruction is fundamentally an ill-conditioned inversion problem whose success depends on both measurement design and prior modeling.

Reconstruction algorithms have therefore become a central part of CUP research. HCUP work identifies TwIST and AL as widely used iterative solvers based on TV regularization, while later methods introduce plug-and-play ADMM formulations in which denoisers act as implicit priors (Guo et al., 2023). In the TV-CD framework, the denoising cascade

$E(m,n)=TSCI(x,y,t),$ 4

is embedded into PnP-ADMM so that local smoothness, learned image priors, and temporal neighborhood information are combined within the iterative reconstruction (Guo et al., 2023). Other CUP-adjacent systems retain the compressive inversion but alter the forward operator, as in diffuser-based rolling-shutter video recovery or time-of-flight-based lensless ultrafast sensing (Weinberg et al., 2020).

3. Architectural variants and encoding strategies

Subsequent work has diversified CUP chiefly by changing how time is encoded before reconstruction, how many dimensions are multiplexed, and how much burden is placed on the streak camera. The result is a family of systems that share a compressed single-shot objective but differ in whether the dominant encoding is temporal, spectral, object-motion-based, or entirely passive (Guo et al., 2023).

Variant	Encoding strategy	Distinguishing feature
CUP	Coded aperture + streak-camera temporal shearing	Standard single-shot reconstruction of $E(m,n)=TSCI(x,y,t),$ 5
HCUP	Pseudo-random spatial mask + spectral dispersion + temporal shearing	Single-exposure recovery of $E(m,n)=TSCI(x,y,t),$ 6
TACSI	Standard CSI streak shear + second scanning axis translating the object relative to the coded aperture	Lower streak compression ratio and reduced coded-aperture motion blur
AOD-CUP	FACED-generated discrete illumination + spatial encoding + spectral shearing by diffraction grating	Eliminates the streak camera and enables adjustable frame number and inter-frame interval
Passive time-to-space mapping	Microlens-array image replication + cover-glass path delays + CMOS detection	Computation-free direct ordering of delayed replicas rather than compressed-sensing inversion

HCUP adds an extra spectral dimension to CUP. The scene is first spatially encoded by a pseudo-random binary mask, then different spectral components are dispersed horizontally by a prism or grating, different time frames are sheared vertically by a streak camera or electro-optical deflector, and the resulting mixed information is integrated into a single 2D detector image (Guo et al., 2023). This preserves the CUP logic while increasing dimensionality and compression burden.

Two-axis compressed streak imaging (TACSI) modifies CSI rather than replacing it. A second scanning axis shuttles a conjugate image of the object with respect to the coded aperture during exposure, so that the object moves laterally across the coded mask while the usual temporal shear proceeds vertically. The stated consequence is a lower streak compression ratio and a flash-and-shutter phenomenon that reduces coded-aperture motion blur, especially for stationary or weakly varying fluorescent objects (Keppler et al., 2024).

All-optical discrete illumination compressed ultrafast photography (AOD-CUP) responds directly to two identified CUP bottlenecks: dependence on a streak camera and the image-quality loss caused by excessive compression. It replaces the streak-camera time-shearing stage with a FACED-based discrete illumination scheme that converts a femtosecond pulse into a train of sub-pulses, after which temporal information is mapped into the spectral domain and read out by dispersive gratings. The acquisition chain is written as

$E(m,n)=TSCI(x,y,t),$ 7

with $E(m,n)=TSCI(x,y,t),$ 8 denoting the discrete illumination operator (Cheng et al., 27 May 2025).

A useful boundary case is the fully passive architecture built from a commercial $E(m,n)=TSCI(x,y,t),$ 9 microlens array, a stack of microscope cover glasses, relay optics, and a consumer-grade CMOS sensor. Here each microlens creates a replicated sub-image, selected channels experience different optical path delays, and the time sequence is recovered by channel segmentation and delay ordering rather than by iterative inversion. The paper explicitly contrasts this with CUP: no sparsity prior, no compressed sensing, no streak camera, and no computational reconstruction (Eşlik et al., 30 Apr 2026).

4. Performance envelope, fidelity limits, and mitigation strategies

CUP and its extensions are characterized by extreme temporal compression. HCUP is described as possessing “an incredibly high frame rate of tens of trillions of frames per second and a sequence depth of several hundred,” which is precisely why reconstruction becomes difficult when fine spatial and spectral detail must be recovered from a single 2D observation (Guo et al., 2023). The same tension is visible across the literature: higher sequence depth and stronger multiplexing increase temporal reach but amplify ill-posedness.

Several bottlenecks recur. HCUP work attributes poor reconstruction quality to the ultra-high data compression ratio induced by extremely large sequence depth and to the limited fidelities of traditional reconstruction algorithms; TV-based solvers are said to suffer from staircasing artifacts, texture loss, and poor recovery of weak or low-SNR components under extremely high compression ratio (Guo et al., 2023). AOD-CUP identifies an additional hardware limitation specific to conventional streak-camera CUP: charge-coupling artifacts in the streak tube distort the sheared image and compromise spatial fidelity (Cheng et al., 27 May 2025). In X-ray CUP for laser-fusion diagnostics, spatial resolution becomes especially problematic when coded masks must have relatively large pixels because of fabrication constraints (Li et al., 2021).

Mitigation strategies fall into two broad classes: added measurements and altered encoding. In the hohlraum diagnostic study, supplementing CUP with a simplified space-resolving flux detector (SSRFD) improved the baseline $C$ 0 mask reconstruction from PSNR = 30.3833 dB and SSIM = 0.8797 for CUP alone to PSNR = 36.4277 dB and SSIM = 0.9329 for the proposed method (Li et al., 2021). In dense-gas plasma imaging, CUP reconstruction was enhanced by an auxiliary channel consisting of a spatially integrated and temporally sheared unencoded signal projected onto an unused region of the streak camera, which the authors describe as a new technological improvement (Wang et al., 7 Jul 2025).

Altered encoding can also lower effective compression before inversion. TACSI reports, in simulation, an improvement from PSNR 15.9 dB to 18.3 dB, SSIM 0.524 to 0.740, BUR 3.62% to 15.8%, and CR 40.3 to 6.55 when moving from single-axis CSI to two-axis TACSI (Keppler et al., 2024). AOD-CUP, by shifting to FACED-based discrete illumination and spectral shearing, reports that a USAF 1951 target could resolve element 7-1 while 7-2 was not clearly resolved, with an estimated spatial resolution of approximately 128 lp/mm compared with 45.3 lp/mm for DI-CUP, corresponding to about a 1.82× improvement (Cheng et al., 27 May 2025).

These results suggest that CUP performance is controlled as much by measurement design as by nominal frame rate. A plausible implication is that practical CUP development increasingly hinges on balancing temporal depth, compression ratio, detector artifacts, and prior modeling rather than pushing any single parameter in isolation.

5. Scientific and diagnostic applications

CUP has been used as a direct diagnostic of dense plasma dynamics. In laser breakdown of argon and xenon at pressures up to 40 bar, CUP recorded 2D images of single-event plasma evolution with a spatial resolution of 250 × 100 pixels and an equivalent frame rate of 500 billion frames per second (Wang et al., 7 Jul 2025). The data were collected through red, green, blue, and broad-band filters and used to infer temperature, emissivity, opacity, heat flow, and ionization behavior. The resulting interpretation was that dense plasmas can contract rather than expand, that observed ionization exceeds Saha-equation expectations even with screening corrections, and that during the first few nanoseconds internal plasma processes dominate evolution while heat loss to the surroundings is too weak to explain the observed cooling (Wang et al., 7 Jul 2025).

In inertial-confinement-fusion diagnostics, CUP has been proposed as the high-temporal-resolution backbone of a 2D time- and space-resolved X-ray diagnostic method for hohlraum implosion. The simulated scene comprised 80 frames over 2 ns, corresponding to 25 ps/frame, and the combined CUP + SSRFD architecture was designed to retain the temporal resolution of streak imaging while adding spatial and flux constraints (Li et al., 2021). The significance here is not a new CUP principle but an adaptation to X-ray conditions where mask coarseness and reconstruction ambiguity are especially severe.

CUP-style compressed streak imaging has also been pushed toward biological fluorescence microscopy. TACSI demonstrated microsecond-scale imaging of CHO-K1 cells loaded with FluoVolt voltage-sensitive dye during microsecond pulsed electric field stimulation at about ~800 V/cm, 50 μs pulse width, and 50% duty cycle (Keppler et al., 2024). In the reported comparison, single-axis CSI reconstructions were dominated by artifacts, whereas two-axis TACSI preserved cell boundaries and resolved the expected polarity-dependent fluorescence response across the membrane.

The CUP concept has been extended beyond optical imaging in the usual sense. In the gravity proposal termed T-CUP, CUP serves as the ultrafast event recorder in a passive optical measurement scheme where an orbiting particle periodically blocks laser beams at a detection point, creating a sudden void or drop in the optical signal that is captured by the T-CUP data acquisition unit (Faizal et al., 2020). The proposal claims 10 trillion frames per second and a sensitivity of $C$ 1, using temporal imaging based on the duality between paraxial diffraction and narrow-band dispersion with quadratic phase modulation. This use of CUP turns a short-range gravity problem into a temporal measurement problem rather than a conventional displacement measurement (Faizal et al., 2020).

AOD-CUP preserves CUP’s single-shot ultrafast function while targeting scenes with complex spatial detail. Demonstrations include spatiotemporally chirped light fields, laser-induced stress-wave propagation in LiF crystals, and femtosecond-laser-induced air plasma channel formation, with reported fitted wave velocities of 9.2 km/s along $C$ 2, 8.7 km/s along $C$ 3, 5.1 km/s for the inner stress wave, and plasma expansion velocity of approximately 31.5 km/s (Cheng et al., 27 May 2025).

A common misconception is that any single-shot ultrafast imaging system is a CUP system. The passive architecture assembled from an MLA, cover-glass delay stack, relay lens, and CMOS sensor shows why that is not so. It captures the evolution of a picosecond laser pulse with a temporal sampling interval of 1.46 ps, an effective frame rate of 685 Gfps, and a sequence depth of ten frames, at a total hardware cost below US\$500; however, the time sequence is recovered by channel segmentation and delay ordering, not by solving an inverse problem (Eşlik et al., 30 Apr 2026). The paper explicitly states that the method is much closer to a passive optical delay line array than to CUP.

Rolling-shutter compressive video provides a second boundary case. In one diffuser-based lensless system, sparse recovery reconstructs 140 video frames at over 4,500 frames per second from a single rolling-shutter exposure, with the row-dependent shutter function providing temporal sampling and the diffuser providing spatial multiplexing (Antipa et al., 2019). A related lens-based system with a diffuser in the pupil plane and a conventional rolling-shutter camera reconstructs 54 frames at 104,166 fps from a single frame while maintaining diffraction-limited resolution (Weinberg et al., 2020). These systems are CUP-like in that they compress a spatiotemporal scene into one exposure and recover it computationally, but their coding is primarily spatial and sensor-timing-based rather than streak-camera-based temporal shearing.

Lensless imaging with compressive ultrafast sensing is likewise CUP-adjacent rather than CUP proper. There the measurement is produced by active pulsed structured illumination and time-resolved detection, with each acquisition yielding a time series in which time is a function of the photon’s origin in the scene. The goal is static planar reflectance recovery with fewer illumination patterns than a conventional single-pixel camera, not snapshot recording of a transient movie (Satat et al., 2016). The relationship to CUP is therefore conceptual and methodological—forward modeling plus compressive inversion—rather than architectural identity.

These comparisons clarify what is specific about CUP. CUP is defined not merely by high speed, nor merely by single-shot acquisition, but by the particular conjunction of coded spatial modulation, streak-camera temporal shearing, and compressed reconstruction. Much of the recent literature can be read as either refining this conjunction—through improved priors, auxiliary constraints, or modified shear geometries—or circumventing one of its liabilities, especially streak-camera artifacts and excessive compression, while retaining the broader aim of single-shot ultrafast imaging (Cheng et al., 27 May 2025).