TensoIS: Feed-Forward Tensorial Inverse Scattering

• TensoIS is a feed-forward method that estimates heterogeneous subsurface scattering parameters from six images using a low-rank tensor representation.
• It employs a synthetic HeteroSynth dataset and Fractal Perlin noise to model complex volumetric variations in extinction and albedo.
• The approach achieves rapid, accurate recovery of scattering fields, outperforming iterative methods with significantly lower computation times.

Searching arXiv for the primary paper and closely related tensorial inverse scattering / tensor-factorization work. Tensorial Inverse Scattering (TensoIS) is a learning-based feed-forward framework for estimating subsurface scattering parameters of heterogeneous media from sparse multi-view image observations, introduced in the paper "TensoIS: A Step Towards Feed-Forward Tensorial Inverse Subsurface Scattering for Perlin Distributed Heterogeneous Media" (Tiwari et al., 4 Sep 2025). In that formulation, the target is the recovery of volumetric extinction coefficient $\sigma_t(\mathbf{x})$ and volumetric albedo $\alpha(\mathbf{x})$ inside a bounded volume, where the internal heterogeneity is modeled using Fractal Perlin noise rather than a homogeneous prior. The method couples a synthetic dataset, HeteroSynth, with a low-rank tensor parameterization that reconstructs dense $3$D scattering fields from six input images in a single forward pass, thereby contrasting with analysis-by-synthesis and differentiable volume rendering pipelines that solve the inverse problem iteratively (Tiwari et al., 4 Sep 2025).

1. Problem setting and physical formulation

TensoIS addresses the inverse problem of recovering heterogeneous subsurface scattering parameters from images of an object whose appearance is governed by multiple scattering in a volumetric medium. The forward model assumes a per-voxel extinction coefficient $\sigma_t$, volumetric albedo $\alpha$, and phase function $f_p$, with

$\sigma_t = \sigma_s + \sigma_a$

and

$\alpha = \frac{\sigma_s}{\sigma_t}.$

The paper uses the Henyey-Greenstein phase function with $g=0$ for isotropic scattering, and images are rendered in Mitsuba 3 by simulating radiative transfer through a heterogeneous volume bounded by a shape (Tiwari et al., 4 Sep 2025).

The inverse task is defined as follows: given a set of six images $\mathcal{I}=\{I_k\}_{k=1}^6$ of a heterogeneous object $\alpha(\mathbf{x})$0, estimate the per-voxel fields $\alpha(\mathbf{x})$1 and $\alpha(\mathbf{x})$2 inside the object's bounding volume $\alpha(\mathbf{x})$3. The paper characterizes this mapping as mathematically ill-posed because distinct parameter configurations can yield similar image observations, particularly under heterogeneity (Tiwari et al., 4 Sep 2025).

This formulation places TensoIS within inverse scattering rather than conventional surface inverse rendering. Its target variables are volumetric transport parameters, not merely surface BRDFs or homogeneous bulk coefficients. A central premise of the work is that most existing approaches either approximate complex path integrals through analysis-by-synthesis or use differentiable volume rendering techniques to account for heterogeneity, whereas prior learning-based estimation methods largely assume homogeneous media (Tiwari et al., 4 Sep 2025).

2. Perlin-distributed heterogeneity and the HeteroSynth dataset

A defining element of TensoIS is its use of Fractal Perlin noise as a procedural model for heterogeneous scattering parameters. The paper states that no specific distribution is known to the authors that can explicitly model heterogeneous scattering parameters in the real world, and proposes Perlin and Fractal Perlin noise as effective models for intricate heterogeneities of natural, organic, and inorganic surfaces (Tiwari et al., 4 Sep 2025). The stated motivation is not that Perlin noise is a measured physical law, but that it is a usable procedural prior in the absence of a well-defined empirical distribution. This suggests a pragmatic prior rather than a closed account of real-world heterogeneity.

To operationalize that prior, the authors construct HeteroSynth, a synthetic dataset of photorealistic images paired with ground-truth volumetric scattering parameters. HeteroSynth contains 103 varied 3D meshes from the VOLMAP dataset, with 90 used for training and 13 for testing. Parameter volumes are generated on an $\alpha(\mathbf{x})$4 grid using Fractal Perlin noise with five octaves; a modulus operation on $\alpha(\mathbf{x})$5 introduces sharp, high-frequency variations, and $\alpha(\mathbf{x})$6 is varied in the interval $\alpha(\mathbf{x})$7 to model different scattering and absorption regimes (Tiwari et al., 4 Sep 2025).

Rendering is performed in Mitsuba 3 under both point and environment lighting, producing approximately 1.1 million images in total. Each configuration includes six view angles for full appearance sampling. For every image, the dataset provides exact $\alpha(\mathbf{x})$8D volumes of $\alpha(\mathbf{x})$9 and $3$0, along with foreground masks and, where available, meshes or signed distance functions for shaping the object in the grid (Tiwari et al., 4 Sep 2025).

HeteroSynth is therefore not merely an image corpus; it is a supervised inverse-scattering benchmark with paired volumetric ground truth. In the logic of the paper, the dataset is a necessary complement to the feed-forward formulation, because it supplies the structured supervision required to regress heterogeneous scattering volumes directly from image observations (Tiwari et al., 4 Sep 2025).

3. Low-rank tensor representation and network design

Instead of directly predicting dense $3$1 parameter volumes, TensoIS represents each volume as a sum of learnable low-rank tensor components. The reconstruction is written as

$3$2

where $3$3 is the $3$4D parameter volume, $3$5 are axis-aligned vectors, $3$6 are plane-aligned matrices, and $3$7 is the decomposition rank, set to $3$8 in the reported experiments (Tiwari et al., 4 Sep 2025).

The image encoder processes each of the six observations with a dedicated $3$9D convolutional encoder, producing per-view features $\sigma_t$0 that are concatenated into a latent code: $\sigma_t$1 This latent representation is then passed to decoder branches that predict tensor factors for each physical parameter: $\sigma_t$2 The outer-product composition of these predicted vectors and matrices yields the final $\sigma_t$3D grids for $\sigma_t$4 and $\sigma_t$5 (Tiwari et al., 4 Sep 2025).

The paper also notes that the network estimates environment lighting through spherical harmonic coefficients when necessary. Architecturally, the significance of the tensor parameterization is computational rather than merely notational: it avoids the memory and compute burden of direct dense volume regression while preserving a structured volumetric representation (Tiwari et al., 4 Sep 2025).

This design is conceptually adjacent to tensor-factorized scene representations in inverse rendering. "TensoIR: Tensorial Inverse Rendering" uses a tensor factorization-based neural scene representation to estimate scene geometry, surface reflectance, and environment illumination from multi-view images under unknown lighting conditions (Jin et al., 2023). The commonality is the use of low-rank tensor structure as an efficient and regularized representation; the distinction is that TensoIS targets heterogeneous subsurface scattering parameter volumes rather than radiance-field geometry and surface appearance.

4. Optimization, supervision, and feed-forward inference

Training is supervised at the level of volumetric parameter fields. The principal loss reported in the paper is a masked $\sigma_t$6 objective over object voxels: $\sigma_t$7 where $\sigma_t$8 denotes the object mask and $\sigma_t$9 is the network prediction. The paper further reports auxiliary lighting and feature consistency regularization terms in training (Tiwari et al., 4 Sep 2025).

The inference mode is strictly feed-forward. Once trained, the network predicts dense heterogeneous scattering fields for a new six-view image set with a single forward pass and requires no iterative optimization. This is one of the paper's central contrasts with optimization-based inverse scattering pipelines, which are described as slow, ambiguous, and susceptible to local minima (Tiwari et al., 4 Sep 2025).

The role of multi-view and multi-light supervision is explicitly tied to ambiguity reduction. According to the reported ablations and discussion, multi-view and multi-light training, together with feature regularization loss, help disambiguate parameter configurations that produce visually similar images (Tiwari et al., 4 Sep 2025). In that sense, TensoIS does not eliminate ill-posedness in a mathematical sense; rather, it constrains the inverse map through architectural bias, procedural priors, and supervised data coverage.

A useful comparison emerges with recent electromagnetic inverse-scattering systems that also separate an intermediate representation from the final material field. "Generalizable Neural Electromagnetic Inverse Scattering" uses induced current as a physical bridge between scattered fields and relative permittivity, enabling generalizable feed-forward prediction on unseen data (Cheng et al., 26 Jun 2025). "Physics-Informed Deep Contrast Source Inversion" similarly models current distributions with a residual MLP while treating medium parameters as learnable tensors in a differentiable framework (Sun et al., 14 Aug 2025). These works address different physics and measurement modalities, but they illustrate a broader pattern in contemporary inverse-scattering research: direct end-to-end reconstruction becomes more tractable when the unknown field is structured through an intermediate or low-rank representation.

5. Evaluation protocol, empirical behavior, and reported limitations

The evaluation reported for TensoIS covers unseen heterogeneous variations over shapes from the HeteroSynth test set, smoke and cloud geometries obtained from open-source realistic volumetric simulations, and some real-world samples. The metrics include Mean Absolute Error (MAE) and MSE between predicted and ground-truth volumes, as well as rendered-image similarity through MSE and $\alpha$0MS-SSIM computed from images synthesized with predicted parameters (Tiwari et al., 4 Sep 2025).

The paper describes ablations over direct volume regression versus tensor decomposition, the number of tensor components, separate versus shared encoders and decoders, and a volume-optimization baseline implemented with Pytorch3D. It reports that TensoIS achieves low average errors for both parameter volumes and produces highly realistic rendered images from predicted parameters, including on unseen Perlin-distributed heterogeneities (Tiwari et al., 4 Sep 2025).

Efficiency is a prominent empirical result. The reported runtime comparison is that TensoIS is orders of magnitude faster than optimization-based methods, described as a few milliseconds versus approximately 30 minutes per scene (Tiwari et al., 4 Sep 2025). The same section states that traditional iterative optimization can reproduce images but may yield physically inaccurate or artifact-ridden parameter fields, which is presented as evidence that photometric reproduction alone is not a sufficient criterion for inverse-scattering quality (Tiwari et al., 4 Sep 2025).

For real-world data, the paper reports plausible heterogeneous scattering volumes, but it also identifies limitations. The main stated issues are geometry estimation and the lack of surface reflectance modeling (Tiwari et al., 4 Sep 2025). These limitations are significant because they delimit the scope of the reported real-world applicability: the framework is evaluated on real samples, but its training prior and rendering assumptions remain centered on synthetic Perlin-distributed volumetric heterogeneity.

6. Position within inverse rendering and broader tensorial inverse scattering research

Within computer graphics, TensoIS is positioned against two families of prior approaches: analysis-by-synthesis methods that approximate complex path integrals and differentiable volume rendering methods for heterogeneous media, as well as learning-based estimators that assume homogeneous scattering parameters (Tiwari et al., 4 Sep 2025). The paper's stated contribution is a feed-forward architecture for procedurally heterogeneous inverse scattering, combined with a low-rank tensor representation and a synthetic supervision pipeline.

The broader phrase "tensorial inverse scattering" has a different and older history in applied mathematics, electromagnetics, and computational imaging. Inverse medium scattering with heterogeneous scattering coefficients has been formalized in terms of scattering coefficients $\alpha$1 with symmetry and tensorial properties, and the exponential decay of these coefficients has been linked directly to the exponentially ill-posed character of fixed-frequency inverse medium scattering (Ammari et al., 2013). For Maxwell systems, inverse scattering has been reduced to a Fredholm second-kind integral equation with a scalar weakly singular kernel, enabling reconstruction of complex permittivity in a bounded region from scattering amplitude data under a Born-type approximation (Ramm, 2012). In polarization-sensitive optical coherence tomography, the recovery of an orthotropic susceptibility tensor has been formulated as a three-dimensional inverse scattering problem for Maxwell's equations, with reconstruction based on the second-order Born approximation (Elbau et al., 2017).

These usages do not define TensoIS as a single standardized framework across fields. Rather, they show that the adjective tensorial may refer to tensor-valued unknowns, multi-indexed scattering coefficients, tensor-space liftings, or low-rank tensor parameterizations. For example, "Non-convex regularization of bilinear and quadratic inverse problems by tensorial lifting" introduces dilinear mappings and diconvex regularization by lifting nonlinear operators to linear representatives on tensor spaces (Beinert et al., 2018). By contrast, the graphics method TensoIS uses tensoriality primarily as a low-rank volumetric representation for inverse subsurface scattering (Tiwari et al., 4 Sep 2025).

A common misconception is therefore to treat all "tensorial inverse scattering" papers as instances of the same problem class. The available literature supports a narrower conclusion: the 2025 TensoIS paper defines a specific feed-forward inverse subsurface scattering framework for Perlin-distributed heterogeneous media (Tiwari et al., 4 Sep 2025), while related work in inverse rendering, electromagnetic imaging, and mathematical inverse problems uses tensorial structure in distinct senses and under different forward operators [(Jin et al., 2023); (Cheng et al., 26 Jun 2025); (Sun et al., 14 Aug 2025); (Ammari et al., 2013)]. Within that narrower definition, TensoIS is best understood as an attempt to make heterogeneous inverse scattering tractable by combining a procedural prior, synthetic paired supervision, and low-rank tensor decoding in a single-pass neural pipeline (Tiwari et al., 4 Sep 2025).