Direct-3D Inversion: Methods & Applications

• Direct-3D Inversion is a set of reconstruction strategies that solve full 3D inverse problems using analytical, variational, or learned approaches without 2D decomposition.
• These methods span areas like coherent diffractive imaging, seismic inversion, and tomography, often achieving high precision and computational efficiency.
• Direct-3D Inversion leverages explicit inversion formulas, amortized training, and full-physics coupling to robustly reconstruct complex 3D models from measured data.

Direct-3D Inversion denotes a family of reconstruction strategies in which the target of inference is intrinsically three-dimensional and the inversion is carried out without decomposing the problem into independent 2D reconstructions or a multistage multiview pipeline. Across the literature, the phrase is used in several technically distinct senses: a non-iterative 3D coherent diffractive imaging inversion in the Fraunhofer regime, a direct, non-iterative inversion algorithm for the 3D pseudo-polar Fourier transform, a full-grid direct-3D inversion method in surface-wave tomography, a direct 3D seismic inversion that maps pre-stack seismic data to a 3D subsurface velocity model, and fast CGO-based or D-bar reconstructions for 3D EIT (Podorov et al., 2015, Averbuch et al., 2015, Zhao et al., 21 Jul 2025, Gelboim et al., 2023, Hamilton et al., 2020). This heterogeneity suggests that “direct” is not a single algorithmic property, but a domain-specific claim about how much of the inverse problem is solved analytically, amortized in training, or posed on the full 3D unknown rather than on lower-dimensional intermediates.

1. Semantic range and defining characteristics

In transform-based imaging, “direct” often means that reconstruction follows from an explicit inversion formula rather than from iterative phase retrieval or nonlinear optimization. In "Direct Inversion of Digital 3D Fraunhofer Holography Maps" (Podorov et al., 2015), the reconstruction is obtained by forming the inverse Fourier transform of the measured intensity volume and applying a suitable linear differential operator associated with the reference geometry. In "Direct Inversion of the 3D Pseudo-polar Fourier Transform" (Averbuch et al., 2015), inversion is reduced to structured one-dimensional resampling problems and Toeplitz solves, with total complexity $O(n^3\log n)$. In "Practical Global Backprojection-Convolution in Transmission Cone-beam Computed Tomography" (Grewar et al., 2024), directness refers to a weighted global backprojection followed by a global 3D deconvolution or Fourier filtration.

In PDE-constrained inverse problems, “direct-3D” can mean that the inversion is posed on the full 3D model rather than through a sequential lower-dimensional workflow. "Bayesian Surface Wave Inversion for 3D Shear Wave Velocity Structure Beneath the British Isles: comparing Direct-3D Variational Inversion to Two-step (2D+1D) Inversion Methods" (Zhao et al., 21 Jul 2025) defines the contrast explicitly: the direct-3D method infers a $28 \times 29 \times 20$ shear-wave velocity model from observed inter-receiver Love-wave travel times, whereas the two-step methods first solve period-dependent 2D tomography and then many independent 1D inversions.

In learned inverse problems, “direct” usually denotes a data-to-model regression or estimator evaluation at deployment. "Deep Compressed Learning for 3D Seismic Inversion" (Gelboim et al., 2023) uses a learned operator $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$ to map pre-stack seismic data directly to a $192\times 64\times 496$ velocity volume, with iterative reconstruction at test time: no. "Deep Bayesian Inversion" (Adler et al., 2018) uses deterministic networks to approximate the posterior mean and pointwise posterior variance in ultra low dose 3D helical CT, again in one forward pass per estimator.

A plausible implication is that Direct-3D Inversion is best treated as a family resemblance term. The common denominator is not a single solver class, but the refusal to collapse a 3D inverse problem into a chain of 2D surrogates or to depend on expensive per-instance iterative reconstruction at deployment.

2. Analytic and transform-based direct inversion

A canonical analytic example is 3D Differential Fourier Holography. Under the Fraunhofer approximation, the measured data satisfy

$$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$

and the inverse Fourier transform of intensity yields the autocorrelation of the total object (Podorov et al., 2015). If the object is measured together with a suitably designed reference object, and if an operator $A$ is chosen so that

$$A\chi_{\mathrm{ref}}(\mathbf r)=\sum_i c_i\delta(\mathbf r-\mathbf a_i),$$

then differentiating the autocorrelation collapses the reference to delta functions and produces shifted copies of the unknown object. For a line reference along $\mathbf p$,

$$A=(\mathbf p\cdot\nabla), \qquad \widehat{A}(\mathbf s)= i(\mathbf p\cdot\mathbf s),$$

so the differentiation is implemented by multiplication in reciprocal space rather than by finite-difference differentiation in real space. The paper summarizes the practical reconstruction as

$$\text{object} + \text{c.c. mirrored object} + \text{noise} \propto 2\iiint (\mathbf p,\mathbf s)\,I(\mathbf s)e^{i\mathbf s\cdot \mathbf r}\,d\mathbf s.$$

The method is mathematically exact under its model assumptions, but the reconstruction contains shifted replicas, mirrored/twin image terms, and autocorrelation noise (Podorov et al., 2015).

The 3D PPFT inversion literature gives a different form of exact directness. "Direct Inversion of the 3D Pseudo-polar Fourier Transform" (Averbuch et al., 2015) first resamples the PPFT data from the 3D pseudo-polar grid to a decimated Cartesian frequency grid $28 \times 29 \times 20$0 by an onion-peeling procedure, and then inverts a separable decimated Fourier operator along each axis. The technical core is the reduction of 3D inversion to repeated one-dimensional trigonometric-polynomial resampling and Toeplitz-structured linear algebra, accelerated by the Gohberg–Semencul formula and FFT-based circulant embedding. The total inversion cost is

$28 \times 29 \times 20$1

and the reported reconstructions for $28 \times 29 \times 20$2, $28 \times 29 \times 20$3, and $28 \times 29 \times 20$4 volumes have RMSE near machine precision (Averbuch et al., 2015).

In cone-beam CT with multidimensional source loci, the same theme appears as global backprojection-convolution. "Practical Global Backprojection-Convolution in Transmission Cone-beam Computed Tomography" (Grewar et al., 2024) shows that, after a suitable volumetric weighting, weighted cone-beam backprojection over a cylindrical source locus becomes equivalent to a shift-invariant parallel-beam backprojection on a common viewing sphere support. Exact inversion then reduces to multiplying each Fourier component by

$28 \times 29 \times 20$5

where $28 \times 29 \times 20$6 is the integral of the solid-angular density of source points along the great circle perpendicular to $28 \times 29 \times 20$7. For the cylindrical specialization, the paper derives both an explicit volumetric weighting and a closed-form 3D Fourier multiplier (Grewar et al., 2024).

Direct inversion in 3D EIT uses tailor-made nonlinear Fourier transforms rather than standard FFT geometry. "3D EIT Reconstructions from Electrode Data using Direct Inversion D-bar and Calderón Methods" (Hamilton et al., 2020) transforms

$28 \times 29 \times 20$8

into a Schrödinger equation

$28 \times 29 \times 20$9

constructs approximate scattering data $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$0, recovers $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$1 by inverse Fourier transform, and then reconstructs $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$2 by solving a single elliptic boundary value problem. "Fast 3D Partial Boundary Data EIT Reconstructions using Direct Inversion CGO-based Methods" (Hamilton et al., 2024) extends this logic to partial boundary data by deriving ND-map formulas such as

$$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$3

and by using the corresponding inverse Fourier and Schrödinger back-end. In these works, low-pass filtering in the nonlinear Fourier domain is the stabilizing mechanism, not iterative Tikhonov or TV optimization (Hamilton et al., 2020, Hamilton et al., 2024).

3. Full-physics 3D inversion with coupled models and variational formulations

Some Direct-3D Inversion papers remain iterative, but are “direct” because they solve the full 3D inverse problem in a single coupled parameterization instead of via sequential reductions. "Joint Hydrogeophysical Inversion: State Estimation for Seawater Intrusion Models in 3D" (Steklova et al., 2016) couples a 3D variable-density groundwater model to a 3D direct current resistivity imaging model through Archie's law,

$$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$4

and solves the constrained joint inverse problem with ADMM. The unknown hydrologic state is the initial solute mass fraction field $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$5, while the geophysical state is the bulk conductivity $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$6. Analytical sensitivities are derived for both blocks, and the method was reported to improve recovered salinity relative to separate or simple coupled inversions in both homogeneous and heterogeneous synthetic cases (Steklova et al., 2016).

"A numerical method for efficient 3D inversions using Richards equation" (Cockett et al., 2017) gives another deterministic PDE-constrained example. The forward model is the mixed-form Richards equation with van Genuchten-Mualem constitutive laws, discretized by a staggered finite-volume method and fully implicit backward Euler time stepping. The inverse problem is

$$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$7

and the core numerical contribution is a matrix-free exact discrete sensitivity method. In the 3D example, the mesh has $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$8 cells, i.e. $$\hat{\mathbf{m}}=\mathcal{R}(\Phi \odot \mathbf{d};W)$$9 unknowns for distributed $192\times 64\times 496$0; a full inversion with matrix-free sensitivity required 34.5 hours, whereas finite-difference sensitivity would have required ~8.5 years (Cockett et al., 2017).

In surface-wave tomography, direct-3D variational inversion is positioned against the conventional $192\times 64\times 496$1 decomposition. "Bayesian Surface Wave Inversion for 3D Shear Wave Velocity Structure Beneath the British Isles" (Zhao et al., 21 Jul 2025) formulates the posterior over a fixed-grid $192\times 64\times 496$2 model with 16,240 parameters, defines a Gaussian travel-time likelihood, and uses physically structured variational inference (PSVI) because direct-3D Monte Carlo inversion was judged too high-dimensional. The forward map is still factored computationally through local 1D dispersion and 2D Eikonal travel-time calculation, but the inference is posed on the full 3D model. The paper reports that direct-3D preserved better lateral continuity and produced synthetic data simulations that align more closely with observed data than the two-step inversions (Zhao et al., 21 Jul 2025).

CSSI adds a different form of full-physics coupling. "Nonuniform Iterative Phasing Framework and Sampling Requirements for 3D Dynamical Inversion from Coherent Surface Scattering Imaging" (Donatelli et al., 20 Apr 2026) reconstructs a single 3D density from all grazing-incidence rotation-series measurements at once, with the distorted-wave Born approximation (DWBA) built directly into the measurement projection. Each intensity datum is

$192\times 64\times 496$3

so the inversion must account simultaneously for nonuniform reciprocal-space sampling, phase loss, and dynamical mixing of four Fourier components. The paper combines iterative-projection phasing with a direct inverse NDFT and shows accurate 3D reconstructions from one moderate incident angle or from two low incident angles (Donatelli et al., 20 Apr 2026).

4. Learned and amortized direct inversion

In learned inversion, the defining pattern is an amortized map from measurement space to a 3D model. "Deep Compressed Learning for 3D Seismic Inversion" (Gelboim et al., 2023) takes pre-stack shot-gather data as input and predicts a full 3D velocity model as output. The inversion operator is

$192\times 64\times 496$4

where $192\times 64\times 496$5 is a learned binary source-selection mask and $192\times 64\times 496$6 is a 3D convolutional encoder–decoder. The input tensor is $192\times 64\times 496$7, the output velocity model is $192\times 64\times 496$8, and the held-out test-set 3D SSIM values are 0.9265 without compressed sensing, 0.9262 at 30% CS rate, and 0.9225 at 10% CS rate (Gelboim et al., 2023). The method is direct because there is no explicit wave-equation solve at inference time and no intermediate reconstruction of the discarded shot gathers.

"Deep Bayesian Inversion" (Adler et al., 2018) defines directness differently. Rather than learning posterior samples, it trains one network to estimate $192\times 64\times 496$9 and a second to estimate the pointwise posterior variance

$$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$0

The application is ultra low dose 3D helical CT, but for computational reasons the method is applied slice-wise. The reported runtime is about 80 milliseconds per slice for direct estimation, versus about 40 seconds to estimate posterior moments from 1000 samples with the sampling-based method (Adler et al., 2018). This is direct Bayesian inversion for a 3D imaging task, but not a full 3D CNN inversion backbone.

"Self-Supervised Knowledge-Driven Deep Learning for 3D Magnetic Inversion" (Li et al., 2023) occupies an intermediate position. At inference time, the model is a one-pass 2D-to-3D neural inverter: $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$1 mapping 2D surface magnetic anomaly $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$2 and a 1D depth guideline $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$3 to a 3D susceptibility volume $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$4. But $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$5 is optimized on the target data itself through a closed loop,

$$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$6

with loss

$$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$7

The field-data anomaly fit was reported as MAE 0.0035, SNR 29.3839, and XCOR 0.9993, compared with 0.0711, 8.1859, and 0.9282 for the supervised Jia baseline (Li et al., 2023). This suggests a form of direct inversion after target-specific self-supervised optimization, rather than a universal pretrained direct map.

"Neural Wavelet-domain Diffusion for 3D Shape Generation, Inversion, and Manipulation" (Hu et al., 2023) adds a direct inversion pathway to a 3D generative model. Shapes are converted to a wavelet-domain TSDF representation, and an encoder predicts a latent code

$$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$8

from the coarse coefficient volume $$I(\mathbf s) \sim \left| \mathcal{F}\{\chi(\mathbf r)\} \right|^2,$$9. A conditional diffusion model reconstructs the coarse volume, a detail predictor estimates $A$0, and the final TSDF and mesh are recovered by inverse wavelet transform and marching cubes. The system is not purely feed-forward in its best-performing form, because it uses 400 iterations of latent refinement, but it is also not an optimization-only latent fitting scheme (Hu et al., 2023).

5. Sampling, priors, and identifiability

Across these methods, directness does not remove sampling requirements. In 3D DFH, a single far-field detector image gives only a 2D slice or projection in reciprocal space, so the 3D method assumes a sampled 3D intensity map assembled from many rotated detector images and interpolated onto a 3D Cartesian grid (Podorov et al., 2015). The authors explicitly stress that the method assumes full enough 3D sampling of the diffraction volume. In CSSI, direct-3D inversion depends on incident-angle regime, exit-angle coverage, rotation increment

$A$1

and the avoidance of destructive troughs in the interference factor

$A$2

through the object support (Donatelli et al., 20 Apr 2026).

Direct methods also remain vulnerable to numerical precision and discretization error. The DFH paper warns that all calculations must use double precision, and that GIF and TIFF are inadequate for storing diffraction data because the dynamic range is too small (Podorov et al., 2015). The GBC cone-beam CT paper identifies four specific error sources: accurate discretisation of a multidimensional locus, infinite range of the convolution kernel, absence of a closed-form discrete kernel, and aliasing artefacts in the backprojection that are enormously magnified by the convolution step (Grewar et al., 2024). Its practical remedies include accumulated-weight normalization, softening of backprojections, modified discrete filtering, and low-pad correction (Grewar et al., 2024).

The literature also repeatedly shows that “direct” does not imply unconstrained. The magnetic inversion method needs a user-provided 1D guideline $A$3 to mitigate the volume effect and skin effect inherent in recovering a 3D susceptibility volume from 2D magnetic anomalies (Li et al., 2023). Partial-boundary EIT direct inversion works by truncating full-boundary formulas to $A$4, and the paper states plainly that the resulting images are best interpreted as fast localization and contrast maps; conductivity contrast is underestimated and reconstructions degrade near inaccessible regions (Hamilton et al., 2024). Hydrogeophysical ADMM inversion assumes an Archie-type coupling relation and is most appropriate when the constraint uncertainty is low (Steklova et al., 2016).

A plausible implication is that direct-3D inversion methods trade iterative parameter search for stronger forward-model commitments, stronger data-geometry requirements, or stronger priors. The ill-posedness is displaced, not eliminated.

6. Adjacent usages and common misconceptions

A recurring misconception is that any model named “Direct3D” or “DIRECT-3D” is an inversion method in the strict sense. "Direct3D: Scalable Image-to-3D Generation via 3D Latent Diffusion Transformer" (Wu et al., 2024) is a native 3D generative model whose “direct” claim means that it generates 3D latents from image conditions without requiring a multiview diffusion model or SDS optimization. The inference path still starts from Gaussian noise and performs iterative diffusion denoising conditioned on DINO-v2 and CLIP features; the paper explicitly states that it is not a deterministic one-shot inverse encoder from image to 3D latent (Wu et al., 2024). "DIRECT-3D: Learning Direct Text-to-3D Generation on Massive Noisy 3D Data" (Liu et al., 2024) is even further from inversion: it is a text-conditioned tri-plane diffusion model trained on noisy and unaligned 3D assets, and its only explicit latent-variable inference concerns training-time pose estimation, not test-time recovery of a 3D asset from observations (Liu et al., 2024).

A second misconception is that any mapping from lower to higher dimension with an auxiliary variable is a true inverse. "A Very Simple Approach for 3-D to 2-D Mapping" (Dey et al., 2010) makes the opposite point. The forward map

$A$5

cannot be uniquely inverted from $A$6 alone because there are three unknowns and only two equations. The paper’s inverse is exact only on the augmented space

$A$7

with

$A$8

This is a constrained inverse, not a general inverse of projection (Dey et al., 2010).

Taken together, these adjacent usages sharpen the technical meaning of Direct-3D Inversion. In its strict form, the phrase denotes methods that reconstruct a 3D unknown from measurements or conditions by a direct map, a closed-form transform identity, or a full 3D variational formulation, while avoiding lower-dimensional decomposition as the primary inversion mechanism. In weaker or neighboring usages, “direct” may instead refer to the absence of multiview intermediates, the presence of a feed-forward initialization, or an augmented representation that retains the missing degrees of freedom.