Conditional Geometric Reconstruction

• Conditional geometric reconstruction is a method that uses auxiliary conditions, such as sparse views, learned priors, and temporal context, to mitigate ambiguity in 3D inverse problems.
• It combines classical geometric formulations with probabilistic and generative approaches to recover accurate shapes from limited or uncertain data.
• Recent advances integrate explicit geometric carriers and diffusion-based models to enforce regularization, ensuring improved fidelity and structural consistency.

Searching arXiv for papers relevant to conditional geometric reconstruction. Tool call: arxiv_search with query "conditional geometric reconstruction 3D reconstruction diffusion sparse views canonical anchors geometry" Conditional geometric reconstruction can be understood as a family of reconstruction problems in which the target geometry is inferred under explicit conditioning by auxiliary observations, latent variables, acquisition geometry, temporal context, or learned priors. In the literature, this conditioning ranges from classical formulations with uncertain projection geometry, sparse cross-sections, and point–normal measurements to contemporary feed-forward and diffusion-based systems that condition on sparse views, canonical anchors, time, age, ray bundles, or generative intermediates. Across these settings, the central technical objective is similar: to reduce ambiguity in ill-posed inverse problems by coupling geometric inference to additional structure rather than reconstructing shape from raw observations alone (Pedersen et al., 2022, Zhang et al., 2024, Chen et al., 28 Mar 2025).

1. Conceptual scope and historical lineages

Early work framed reconstruction conditionally in explicitly geometric terms. In “Geometric reconstruction from point-normal data” (Rieffel et al., 2010), the input is a sparse set of point–normal samples

$$\{ (p_i, n_i) \in \mathbb{R}^3 \times S^2 : i=1 \ldots N \},$$

and the goal is to recover planar faces and combinatorial adjacencies consistent with those samples. For a convex polyhedron with one marker per face, the reconstruction is given by the intersection of half-spaces

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$

which yields exactly the polyhedron (Rieffel et al., 2010). In “Geometric Tomography With Topological Guarantees” (Amini et al., 2010), reconstruction is conditioned on a family of cutting planes and their cross-sections, and the reconstructed object is defined by a nearest-point rule relative to the arrangement cells of those planes. Under the Density Condition

$h_C < \mathrm{reach}_C(O)$

and the Transversality Condition

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$

the output preserves the homotopy type, and is further shown to be homeomorphic and isotopic to the original object (Amini et al., 2010).

A second lineage arises in inverse problems with uncertain acquisition geometry. In “A Bayesian Approach to CT Reconstruction with Uncertain Geometry” (Pedersen et al., 2022), the unknown image $x$ and the unknown geometry parameters $\theta$ are reconstructed jointly under the forward model

$$y = A(\theta)x + \epsilon, \qquad \epsilon \sim N(0,\lambda^{-1}I).$$

Here the condition is not merely the measured sinogram $y$, but the joint probabilistic structure tying image, projection geometry, and hyperparameters into one posterior (Pedersen et al., 2022).

Recent deep-learning work generalizes this idea from explicit geometry to learned conditioning. “Conditional Single-view Shape Generation for Multi-view Stereo Reconstruction” (Wei et al., 2019) models $p(S\mid I)$ with a latent variable $r\sim \mathcal N(0,I_d)$, while “Pragmatist” (Zhang et al., 2024) reformulates reconstruction from sparse, unposed observations as conditional novel view synthesis, using generated complete observations to facilitate subsequent reconstruction. This suggests that contemporary conditional reconstruction inherits two older themes: uncertainty modeling and geometric regularization, but relocates them into learned representations and generative priors.

2. Mathematical forms of conditioning

The literature instantiates conditioning in several mathematically distinct ways. A probabilistic formulation appears in Bayesian CT, where the posterior

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$0

couples reconstruction and geometry estimation directly (Pedersen et al., 2022). The key conditional density,

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$1

makes the reconstruction explicitly dependent on uncertain geometry (Pedersen et al., 2022).

A generative conditional formulation appears in single-view shape generation. There the target is the conditional density

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$2

with a latent-conditioned generator

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$3

so that multi-view reconstruction is posed as the intersection of the per-view shape manifolds

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$4

(Wei et al., 2019). In this setting, conditioning is used to represent uncertainty in unseen geometry rather than to collapse it into a single deterministic predictor.

Conditioning can also be temporal or demographic. In “Conditional Temporal Attention Network (CoTAN)” (Ma et al., 2023), the time-varying velocity field is conditioned jointly on integration time $$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$5 and post-menstrual age $$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$6. The attention weights

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$7

define a conditional time-varying velocity field

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$8

which is then integrated to deform a template mesh diffeomorphically (Ma et al., 2023).

In dynamic monocular 4D reconstruction, conditioning appears as arbitrary spatio-temporal querying. “4RC” (Luo et al., 10 Feb 2026) encodes the entire video once,

$$H_i = \{ x \in \mathbb{R}^3 : (x - p_i)\cdot n_i \le 0 \},$$9

and answers geometry–motion queries with

$h_C < \mathrm{reach}_C(O)$0

using the factorization

$h_C < \mathrm{reach}_C(O)$1

to separate base geometry from time-dependent relative motion (Luo et al., 10 Feb 2026). A plausible implication is that conditional geometric reconstruction is not tied to a single output geometry; it can also denote conditional access to geometry across time, view, or deformation state.

3. Representations and conditional carriers

Conditioning is realized through the representation chosen for geometry. Several recent systems use explicit geometric carriers whose parameters are then modulated by learned context. In “360-GeoGS” (Yao et al., 5 Jan 2026), a single 360° RGB panorama is processed by a SphereCNN backbone to produce a spherical cost volume, while high-level image features are aggregated into a global conditioning code $h_C < \mathrm{reach}_C(O)$2. Feature-wise Linear Modulation applies

$h_C < \mathrm{reach}_C(O)$3

before a U-Net decoder regresses per-pixel 3D Gaussian parameters at panorama resolution $h_C < \mathrm{reach}_C(O)$4 (Yao et al., 5 Jan 2026). The Gaussian covariance is parameterized as

$h_C < \mathrm{reach}_C(O)$5

which makes position, scale, and rotation available for direct geometric regularization (Yao et al., 5 Jan 2026).

A different explicit carrier is the triplane signed distance field. In “GCRayDiffusion” (Chen et al., 28 Mar 2025), a transformer-based image encoder produces three feature planes $h_C < \mathrm{reach}_C(O)$6, and the SDF is decoded as

$h_C < \mathrm{reach}_C(O)$7

Camera poses are over-parameterized as neural bundle rays

$h_C < \mathrm{reach}_C(O)$8

and the denoiser is conditioned on local image features and the current SDF values at ray-hit points,

$h_C < \mathrm{reach}_C(O)$9

(Chen et al., 28 Mar 2025). Here the representation itself carries the condition: the evolving SDF regularizes pose diffusion, and the denoised rays provide explicit on-surface samples back to the SDF.

Conditional reconstruction can also target topological rather than purely metric structure. “ComplexGen” (Guo et al., 2022) treats a CAD boundary representation as a chain complex

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$0

with a probabilistic model

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$1

for vertices, edges, faces, incidences, and geometric parameters conditioned on an input point cloud $$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$2. A tri-path transformer decoder predicts existence and incidence probabilities, and a global optimization enforces structural validness constraints such as $$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$3 (Guo et al., 2022). This broadens the notion of geometric reconstruction: the reconstructed quantity may be a valid B-Rep complex rather than a point cloud, mesh, or radiance field.

4. Coupling generative priors to explicit geometry

A major contemporary direction couples conditional generation to explicit geometric reconstruction. “Pragmatist” (Zhang et al., 2024) states the problem as follows: sparse, unposed observations are first completed by a multiview conditional diffusion model, then a feed-forward large reconstruction model predicts a mesh, and the poses of the input views are subsequently recovered by inverting the obtained 3D representations and using them to optimize the reconstructed object. The paper’s stated motivation is that direct prediction from unposed sparse views does not utilize geometric priors and cannot hallucinate the appearance of unseen regions, making fine geometric and textural details difficult to reconstruct (Zhang et al., 2024).

“GeoRect4D” (Wu et al., 22 Apr 2026) addresses dynamic sparse-view 3D reconstruction by coupling an explicit dynamic 3DGS substrate with a single-step diffusion rectifier in a closed-loop optimization. The rendered frame

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$4

is rectified by

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$5

and the substrate is optimized with a hybrid objective

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$6

Structural Locking injects encoder features into decoder UpBlocks,

$$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$7

while Spatio-Temporal Coordinated Attention refines latent states jointly across spatial views and short temporal windows (Wu et al., 22 Apr 2026).

“UniRecGen” (Huang et al., 1 Apr 2026) resolves the tension between deterministic reconstruction and stochastic generation by forcing both modules into a shared canonical space. A VGGT-based reconstruction mapping outputs multi-view point maps $$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$8 in an object-centric frame, and a diffusion generator is conditioned on the same canonical anchor $$h_C < \tfrac12(1-\sin \alpha_C)\,\mathrm{reach}_C(O),$$9. During inference, the denoiser applies

$x$0

while the conditioning tokens are augmented as

$x$1

(Huang et al., 1 Apr 2026). This suggests that conditional reconstruction increasingly depends on explicit coordinate alignment between geometric and generative modules, not merely on feature fusion.

A more extreme case is “Mind the Gap: Geometrically Accurate Generative Reconstruction from Disjoint Views” (Wilczynski et al., 8 May 2026), which defines reconstruction from disjoint inputs $x$2 with $x$3. GLADOS bridges the inputs with a generated intermediate view, globally aligns dense point maps into a coarse scaffold, and then iterates rendering, inpainting, depth estimation, and consistency optimization. Its multiview consistency term is

$x$4

(Wilczynski et al., 8 May 2026). The broader implication is that generative conditioning is being used not only to fill unseen regions but also to establish overlap where none exists.

5. Geometry-aware regularization and optimization

Conditional reconstruction does not reduce to adding a conditioning token; in much of the literature, the conditioning signal is made operational through explicit geometric regularization. “360-GeoGS” (Yao et al., 5 Jan 2026) introduces Depth-Normal geometric regularization to couple Gaussian position, scale, and orientation to rendered depth and normals. The scale flattening loss

$x$5

encourages ellipsoids to flatten onto surfaces; rendered depth is composed by

$x$6

and the rendered normal

$x$7

is constrained against a proxy normal $x$8 through

$x$9

(Yao et al., 5 Jan 2026). The full loss

$\theta$0

shows how conditioning can be embedded into the training objective rather than only the network input.

In pose-free reconstruction, “GCRayDiffusion” (Chen et al., 28 Mar 2025) imposes on-surface regularization from denoised ray-hit points: $\theta$1 supplemented by the eikonal term

$\theta$2

The total objective,

$\theta$3

makes the reconstructed surface and the conditionally estimated camera rays mutually constraining (Chen et al., 28 Mar 2025).

In microscopy, “Three-Step Conditional Diffusion 3D Reconstruction for Light-Field Microscopy” (Zhao et al., 24 May 2026) redesigns diffusion itself into a deterministic three-step reconstruction process. The forward process is

$\theta$4

the network predicts $\theta$5, and the denoised estimate

$\theta$6

supports a DDIM-style deterministic reverse update over three fixed timesteps (Zhao et al., 24 May 2026). A condition encoder injects multi-scale features from the light-field measurement into each U-Net block, and an Inter-Class Detection module uses a Mahalanobis score over feature statistics to identify out-of-distribution inputs (Zhao et al., 24 May 2026).

A recurring misconception is that conditional methods are primarily appearance-driven. The cited regularizers indicate otherwise: conditioning is often made geometrically binding through SDF constraints, normal agreement, structural locking, manifold-validity constraints, or topological sampling conditions.

6. Domains, evaluation regimes, and persistent limitations

Conditional geometric reconstruction now spans substantially different data modalities. In cortical surface reconstruction, CoTAN predicts multi-resolution stationary velocity fields from neonatal MRI and reports “0.12mm geometric error and 0.07% self-intersecting faces,” with inference taking “0.21 seconds to deform an initial template mesh” per hemisphere on the dHCP dataset (Ma et al., 2023). In 360° scene reconstruction, “360-GeoGS” reports improved depth metrics, point-cloud accuracy, and Chamfer distance under Depth-Normal regularization, while noting evaluation “primarily on indoor scenes with known camera poses” and dependence on “an initial depth estimate and reliable normal proxy” (Yao et al., 5 Jan 2026). In sparse unposed object reconstruction, “UniRecGen” evaluates on Toys4K and GSO under “4-view sparse unposed inputs,” reporting improvements in Chamfer-L$\theta$7, Precision, Recall, F-Score, Normal Consistency, and IoU relative to ReconViaGen (Huang et al., 1 Apr 2026). In zero-overlap scene reconstruction, GLADOS introduces “Generative Reconstruction from Disjoint Views” and specialized metrics including CLIP Score, FID, Photometric error, GeCo, MEt3R, and Reconstruction Failure Rate (Wilczynski et al., 8 May 2026).

Evaluation protocols therefore remain domain-specific. Surface tasks use Chamfer Distance, Hausdorff Dist, Normal Consistency, F-score, ASSD, HD$\theta$8, or self-intersection rates; view-synthesis-integrated methods report PSNR, SSIM, LPIPS, or photometric error; zero-overlap methods introduce GeCo and MEt3R; topology-oriented work emphasizes homotopy type, homeomorphism, isotopy, or manifold-validity constraints (Chen et al., 28 Mar 2025, Ma et al., 2023, Wilczynski et al., 8 May 2026, Amini et al., 2010).

The limitations are equally varied but thematically consistent. “Pragmatist” identifies the difficulty of reconstructing fine geometric and textural details from sparse unposed views without geometric priors and unseen-region hallucination (Zhang et al., 2024). “360-GeoGS” notes limitations for “fully unposed or single-image use” (Yao et al., 5 Jan 2026). GLADOS states that in under-constrained regions, generative priors can hallucinate “plausible” but incorrect geometry, and that current metrics “still rely on proxies” (Wilczynski et al., 8 May 2026). “GeoRect4D” is motivated by the observation that naive integration of generative priors can cause structural drift and temporal inconsistency (Wu et al., 22 Apr 2026). These results suggest that the central unresolved issue is not whether conditioning helps, but how to make conditioned predictions geometrically faithful when evidence is sparse, uncertain, or non-overlapping.

Taken together, the literature defines conditional geometric reconstruction not as a single algorithmic family but as a reconstruction principle: geometry is estimated while being constrained, queried, regularized, or completed by auxiliary conditions. Those conditions may be probabilistic, geometric, temporal, demographic, topological, canonical, or generative. The field’s most technically mature directions combine explicit geometric substrates with conditional generative models, and the most rigorous classical results show that when the conditioning assumptions are precise enough, one can obtain not only better reconstructions but also formal guarantees on topology and uncertainty (Pedersen et al., 2022, Amini et al., 2010, Huang et al., 1 Apr 2026).