Multi-Frame Diffraction Imaging

• Multi-Frame Diffraction Imaging is a collection of methods using varied diffraction frames (e.g., multi-distance, multi-exposure) to recover phase and object structure.
• It exploits redundancy across changing geometries, illuminations, and exposures to constrain ill-posed inverse problems and improve image reconstruction.
• Applications span astronomy, XFEL, and electron microscopy, with strategies employing coordinated transforms and wavefront modeling to achieve high-resolution results.

Multi-Frame Diffraction Imaging is a class of diffraction-based imaging methods that uses multiple intensity measurements acquired under different conditions—distances, illuminations, angles, exposures, scan positions, or times—to improve phase retrieval, increase information content, or separate multiple unknowns. Across coherent diffractive imaging, Bragg coherent diffraction imaging, 4D-STEM, transmission Kikuchi diffraction, post-adaptive-optics imaging, and ultrafast XFEL experiments, the common principle is the exploitation of redundancy across frames to constrain an otherwise ill-posed inverse problem and recover object structure, wavefront, strain, spectral components, or temporal evolution (You et al., 2024, Yang et al., 2019, O'Leary et al., 2021, Sauppe et al., 2023).

1. Scope and definitional boundaries

The term does not denote a single instrument or a single inverse problem. In the literature it covers multi-distance lensless imaging, where a focus series is recorded along the optical axis; Bragg coherent diffraction imaging, where a 3D Bragg peak is assembled from many 2D detector frames taken as the sample is rocked; scan-series 4D-STEM, where repeated 4D datasets are fused after distortion correction; and post-facto astronomical restoration, where many short-exposure images degraded by different PSFs are jointly inverted to recover a diffraction-limited estimate of the source brightness distribution (You et al., 2024, Yang et al., 2019, O'Leary et al., 2021, Hope et al., 2022).

A frequent misconception is that “multi-frame” is synonymous with time-resolved imaging. The published record is broader. Frames may differ by propagation distance, rocking angle, detector exposure, wavelength mixture, or scan repetition rather than by physical time. In transmission Kikuchi diffraction, for example, multi-exposure fusion is used to extend detector dynamic range and normalize intensity over very wide-angle patterns; in BCDI, the rocking curve itself is already a multi-frame acquisition; in serial CDI, inter-frame similarity serves as a temporal regularizer for dynamic samples (Zhang et al., 2023, Li et al., 2019, Sheng et al., 2024).

2. Acquisition geometries and frame organization

A central divide is between frame diversity generated by changing propagation geometry and frame diversity generated by changing illumination or sample state. In multi-distance lensless imaging, a CMOS detector is translated axially in $1\ \mathrm{mm}$ steps over 51 positions, producing a focus series of diffraction patterns at distances $z_k$; the redundancy is axial rather than lateral (You et al., 2024). In Bragg CDI, the detector records 2D slices through a 3D Bragg peak while the sample is rotated through a rocking curve, so the frame axis is reciprocal-space sampling along the rocking direction (Yang et al., 2019, Li et al., 2019).

Other systems organize frames through repeated scans or exposure bracketing. In 4D-STEM, the raw data are $I_t(x,y,q_x,q_y)$, where $t$ indexes repeated 4D scans of the same field of view; the additional series dimension converts a 4D diffraction measurement into a multi-frame fusion problem (O'Leary et al., 2021). In wide-angle transmission Kikuchi diffraction, the same pattern is recorded at exposure times $0.1$, $0.01$, $0.005$, and $0.001\ \mathrm{s}$, with 50 frames averaged at each exposure before HDR-style fusion (Zhang et al., 2023).

Ultrafast XFEL implementations add a second organizational principle: several temporally distinct diffraction views can be encoded into one detector exposure. “Double diffraction imaging” records two consecutive diffraction patterns of the same free nanoparticle on two opposing detectors using two time-delayed FEL pulses (Sauppe et al., 2023). Two-color XFEL imaging of helium nanodroplets records a single composite pattern from pulses at $\approx 992\ \mathrm{eV}$ and $\approx 1192\ \mathrm{eV}$, delayed by $z_k$0 to $z_k$1, and separates the views using detector energy sensitivity and pattern recognition (Hecht et al., 27 Aug 2025). Dichography goes further by reconstructing two distinct real-space images from one diffraction pattern modeled as an incoherent sum of two scattering signals (Hecht et al., 27 Aug 2025).

3. Forward models, redundancy, and the inverse problem

The mathematical structure is modality-dependent but consistently redundant. In multi-object multi-frame blind deconvolution, the observation model is

$z_k$2

and the MAP objective in Fourier space is

$z_k$3

with $z_k$4 the object, $z_k$5 the PSF or wavefront-derived transfer function, and $z_k$6 a regularizer (Ramos et al., 15 May 2025). In the astronomical post-AO setting, the Kraken MFBD algorithm models each frame in the sequence through estimation of the instantaneous wavefront at the entrance pupil and targets a diffraction-limited estimate of the source brightness distribution (Hope et al., 2022).

In multi-distance lensless imaging, the propagated field for wavelength $z_k$7 and detector plane $z_k$8 is

$z_k$9

and the measured intensity under multi-wavelength illumination is

$I_t(x,y,q_x,q_y)$0

The key constraint is axial redundancy: fields across different detector planes are physically linked by free-space propagation, so phase retrieval and spectral separation are driven by consistency across many $I_t(x,y,q_x,q_y)$1 rather than by lateral overlap (You et al., 2024).

BCDI uses the Fourier modulus of a complex object $I_t(x,y,q_x,q_y)$2, measured on a non-orthogonal reciprocal-space grid tied to detector axes and rocking direction. That non-orthogonality is not a secondary detail: it is the source of shear in both Fourier and conjugate real-space coordinates, and it determines how multi-frame rocking-curve data must be mapped if one wants physically interpretable 3D reconstructions and constraints in an orthogonal sample frame (Yang et al., 2019, Li et al., 2019).

Time-encoded multi-frame CDI introduces an additional layer of ambiguity. In Dichography, the detector measures

$I_t(x,y,q_x,q_y)$3

so the inverse problem is to recover two complex objects from one measured intensity map (Hecht et al., 27 Aug 2025). In serialCDI, the object-plane exit wave is decomposed as

$I_t(x,y,q_x,q_y)$4

where $I_t(x,y,q_x,q_y)$5 is a temporally shared component and $I_t(x,y,q_x,q_y)$6 is frame-specific; the reconstruction uses inter-frame similarity as a temporal constraint rather than treating frames independently (Sheng et al., 2024).

4. Reconstruction strategies and coordinate handling

The principal reconstruction strategies differ by how they exploit frame coupling. MFBD and MOMFBD couple frames through shared objects and frame-dependent PSFs or wavefronts; torchmfbd implements this as a flexible, GPU-accelerated MAP solver with wavefront-based PSF parameterizations, data-driven PSF dictionaries, smoothness and sparsity regularization, patch-based handling of spatially variant PSFs, and analytic object elimination in the classical MOMFBD limit via a Wiener-type update (Ramos et al., 15 May 2025). Kraken adds a Compact MFBD initialization using only a few frames and states that, under suitable physical constraints, numerical convergence is guaranteed (Hope et al., 2022).

In lensless multi-distance imaging, reconstruction proceeds by alternating forward propagation, enforcement of the measured intensity at each plane, and back-propagation. The Spectral Multiplexing Multi-distance Imaging algorithm rescales all wavelength channels at a detector pixel by a common amplitude factor so that the incoherent sum matches the measured intensity. The reported algorithmic innovation is random selection of planes for amplitude substitution rather than fixed sequential ordering, which improves convergence and reconstruction quality (You et al., 2024).

BCDI contributes a distinct class of reconstruction machinery: geometry-aware coordinate transforms. One line of work expresses arbitrary beamline geometry through three rotation matrices and maps between sample space, lab space, detector reciprocal space, and detector-conjugated space; another derives shear-correcting Fourier transforms so that phase retrieval can be carried out entirely in an orthogonal frame, either through a shear-aware 3D DFT for evenly sampled rocking curves or through projection/back-projection formulas for non-evenly sampled data (Yang et al., 2019, Li et al., 2019).

Dynamic and multiplexed CDI has produced two further reconstruction paradigms. SerialCDI computes an inter-frame similarity map from the standard deviation of the current exit-wave estimates, converts it with an adaptive sigmoidal rule, and uses the resulting shared component as a spatio-temporal regularizer (Sheng et al., 2024). Dichography generalizes iterative phase retrieval to the case of two superimposed, non-interfering scattering signals and combines an intensity projector for the two-component amplitude vector with memetic phase retrieval in the Equinox codebase (Hecht et al., 27 Aug 2025).

5. Representative results across domains

Reported applications span astronomy, visible-light CDI, X-ray microscopy, electron microscopy, and materials science. In post-AO astronomy, Kraken reports the high-resolution reconstruction of the spectroscopic binary $I_t(x,y,q_x,q_y)$7 And with $I_t(x,y,q_x,q_y)$8 separation, acquired with the precursor of SHARK-VIS for the Large Binocular Telescope (Hope et al., 2022). In multi-distance lensless imaging, two-color spectral multiplexing with $I_t(x,y,q_x,q_y)$9 and $t$0 was demonstrated, and the experiments reported that when illumination has only two frequency modes, fewer than 10 frames can suffice to reconstruct both exit waves under favorable conditions (You et al., 2024).

In BCDI, coordinate-consistent mapping across reflections yielded shape overlap up to $t$1 once reconstructions were mapped into a common sample frame, illustrating that multi-frame and multi-geometry fusion can be quantitatively consistent when geometry is treated explicitly (Yang et al., 2019). In a multiscale X-ray workflow linking grain mapping to DFXM, motor positions for all grains in an iron polycrystal containing 1100 grains were calculated within seconds, and 3D misorientation fields across grain boundaries were resolved with $t$2 pixel size (Shukla et al., 25 Aug 2025). A non-contact furnace developed for ID03 extends such multi-frame diffraction imaging to high temperature, with stable operation up to $t$3, heating rates exceeding $t$4, and thermal stability better than $t$5 (Lesage et al., 1 Jul 2025).

Electron and X-ray detector-limited modalities show equally strong gains from frame fusion. In 4D-STEM, multi-frame data fusion reduced the ISPCS standard deviation from $t$6 to $t$7, improved ptychographic resolution from $t$8 to $t$9, and reduced the curl of the reconstructed electric field from $0.1$0 to $0.1$1 (O'Leary et al., 2021). In wide-angle TKD with a Timepix3 detector, multi-exposure fusion increased effective dynamic range to $0.1$2 counts and yielded a simulation correlation coefficient of $0.1$3 after flat-fielding (Zhang et al., 2023).

Ultrafast multi-frame CDI provides explicit temporal benchmarks. Two-color X-ray imaging of helium nanodroplets reported recall reaching $0.1$4 with precision $0.1$5–$0.1$6 in sparsely illuminated outer regions, and the sum of separated red and blue patterns recovered $0.1$7 of the total intensity of the experimental outer region (Hecht et al., 27 Aug 2025). Dichography reconstructed two distinct silver nanoparticles from a single diffraction pattern and, in xenon-doped helium nanodroplets, provided evidence of survival of xenon structures up to $0.1$8 after the interaction with the first shot (Hecht et al., 27 Aug 2025). Double diffraction imaging of xenon clusters using two opposing detectors showed that at focus intensities of about $0.1$9 clusters were still largely intact even at delays up to $0.01$0, whereas at five times higher flux the second-frame diffraction showed increased diameters and density fluctuations (Sauppe et al., 2023).

6. Constraints, misconceptions, and directions of development

A second misconception is that adding frames automatically resolves the inverse problem. The literature shows more conditional behavior. In multi-distance spectral multiplexing, incorrect $0.01$1 values degrade propagation accuracy, thick samples violate the single-slice model, and increasing the number of spectral modes demands more frames and higher SNR (You et al., 2024). In two-color XFEL separation, local photon density emerges as the governing control parameter, with recall collapsing sharply when the mean density exceeds $0.01$2 photons per pixel (Hecht et al., 27 Aug 2025). In strong-phase BCDI, split Bragg peaks and dense fringe structure generate many near-degenerate minima; even when reciprocal-space mismatch is low, different initializations can yield different real-space solutions, which motivated an unsupervised Fourier Vision Transformer optimized directly against diffraction-domain losses (Liu et al., 12 Feb 2026).

The published trajectory points toward increasingly explicit use of diversity. SerialCDI shows that inter-frame similarity can serve as a direct reconstruction constraint for dynamic samples, with reconstructions of good quality within a few hundred iterations (Sheng et al., 2024). Two-frame single-shot schemes demonstrate that temporally distinct views can be recovered either from overlapped detector signals or from two detectors, depending on the experimental design (Sauppe et al., 2023, Hecht et al., 27 Aug 2025). Simulations of four-color XFEL generation report time delays of hundreds of femtoseconds and four-color pulses spanning $0.01$3 to $0.01$4 with $0.01$5 intervals, explicitly proposing single-exposure multi-frame diffraction imaging beyond the two-frame case (Liu et al., 18 Sep 2025). This suggests that future multi-frame diffraction imaging will increasingly combine optical diversity, scan diversity, spectral multiplexing, and learned or physics-informed inversion rather than treating them as separate methodological families.