---
title: 'Signed Distance Field (SDF): An Overview'
url: https://www.emergentmind.com/topics/signed-distance-field-sdf
type: topic
---

# Signed Distance Field (SDF): An Overview

A **signed distance field (SDF)** is a scalar function that assigns every point in space its signed Euclidean distance to the nearest point on a surface. Its magnitude gives distance, its sign identifies the side of the surface, and its zero level set represents the surface itself. Under the common inside-negative convention, for a solid $\mathcal S$,

$$
d(\mathbf x,\mathcal S)=
\begin{cases}
-\min_{\mathbf y\in\mathcal S}\|\mathbf x-\mathbf y\|, & \mathbf x\text{ inside }\mathcal S,\\
0, & \mathbf x\in\mathcal S,\\
+\min_{\mathbf y\in\mathcal S}\|\mathbf x-\mathbf y\|, & \mathbf x\text{ outside }\mathcal S.
\end{cases}
$$

The represented surface is the zero level set,

$$
\mathcal S=\{\mathbf x\in\mathbb R^3\mid f(\mathbf x)=0\}.
$$

SDFs provide continuous geometry, metric proximity, differentiable normals, and a representation suitable for reconstruction, rendering, collision detection, motion planning, shape generation, compression, medical imaging, and geometric reasoning. Neural, voxel, polynomial, radial-basis, and hybrid SDFs differ primarily in how the field is parameterized, supervised, updated, and converted into an explicit surface.

## 1. Mathematical structure and geometric properties

An SDF differs from an occupancy or indicator field because it provides a metric signal throughout space rather than only a binary inside/outside value. The zero level set identifies the surface, while nonzero values encode distance to that surface. This allows arbitrary-resolution queries and supports surface extraction through Marching Cubes, Dual Contouring, or other isosurfacing methods.

At differentiable points, the spatial gradient points toward increasing signed distance. On a regular surface, it corresponds to the surface normal, and an exact Euclidean SDF satisfies the Eikonal equation almost everywhere away from singularities:

$$
\|\nabla_{\mathbf x}f(\mathbf x)\|=1.
$$

The SDF is not differentiable at medial-axis or singular points where multiple surface points are equally close. These singularities are significant in learning and optimization because enforcing unit-gradient behavior indiscriminately can oversmooth corners and other geometrically valid nonsmooth regions. The SDF gradient nevertheless provides a normal field and a direction of increasing clearance, making it useful for collision costs and trajectory optimization [2204.02296].

For a robot configuration $\mathbf q$ and a body point $\mathbf x(\mathbf q)$, a differentiable collision quantity can be differentiated by the chain rule:

$$
\frac{\partial f(\mathbf x(\mathbf q))}{\partial \mathbf q}
=
\nabla_{\mathbf x}f(\mathbf x(\mathbf q))
\frac{\partial\mathbf x(\mathbf q)}{\partial\mathbf q}.
$$

This relationship allows the same representation to encode geometry and the derivatives required by optimization. In articulated robots, link-level SDFs can be transformed through forward kinematics and combined with a minimum operation to obtain a configuration-dependent robot field [2307.00533].

An SDF sample also has a nonlocal geometric interpretation. A sample $(\mathbf p_i,s_i)$ defines a sphere centered at $\mathbf p_i$ with radius $|s_i|$. The true surface must lie on the side prescribed by the sign, must not penetrate the corresponding open ball, and must touch the sphere at a closest point. This tangent-sphere interpretation explains why distance samples away from the zero level set contain information about sharp features and global surface placement [2308.09813]. Recent work further treats global nonintersection, tangency, sidedness, and uncovered-point conditions as hard constraints for constructing mutually realizable discrete SDFs [2605.01919].

## 2. Parameterizations and representations

SDFs can be represented explicitly in a grid, implicitly with a neural network, through analytic basis functions, or with hybrids that combine persistent discrete information and continuous learned fields.

### Neural implicit fields

The standard neural representation uses an MLP,

$$
f_\theta:\mathbb R^3\times\mathbb R^m\rightarrow\mathbb R,
$$

where $\mathbf x$ is a spatial query and $\mathbf z$ is a shape latent code. The reconstructed surface is extracted from $f_\theta(\mathbf x,\mathbf z)=0$. DeepSDF-style models use one global decoder and one shape code, providing compact shape representations but placing the entire geometric burden on one function. Reducing decoder capacity can cause over-smoothed or incomplete geometry, particularly for high-frequency or localized structures [2108.08593].

Several systems augment or replace the basic MLP:

- **Local latent decomposition**: Local Geometry Code Learning distributes a global code over a mesh graph using a Global-to-Local graph network. A shared small decoder evaluates local SDFs, increasing effective capacity without proportionally enlarging the decoder [2108.08593].
- **SIREN fields**: iSDF uses a four-hidden-layer MLP with Softplus activations and off-axis positional encoding for online reconstruction from posed depth images [2204.02296]. HIO-SDF uses a SIREN with sinusoidal activations and combines it with a persistent coarse voxel field [2310.09463].
- **Hash-grid fields**: RISE-SDF and SplatSDF use multiresolution hash encodings followed by compact geometry MLPs. RISE-SDF shares SDF features with radiance and material networks for glossy inverse rendering [2409.20140]. SplatSDF injects Gaussian-splatting features at estimated surface anchors during training while retaining ordinary SDF inference [2411.15468].
- **Geometry-conditioned fields**: gSDF conditions hand and object SDF decoders on articulated kinematic features and pixel-aligned visual features, using joint transformations to express query points in local coordinate systems [2304.11970].
- **Semantic fields**: SurroundSDF jointly predicts a continuous SDF and semantic logits from surround-view images. A query-based architecture uses voxel features internally but produces continuous geometry and semantics at arbitrary coordinates [2403.14366].

### Discrete and voxel fields

Voxel SDFs store distance values in a regular or sparse grid. They are naturally incremental and preserve previously observed geometry, but memory and resolution scale with the mapped volume. Classical TSDF fusion is therefore effective for local mapping but becomes expensive for large environments with fine details. RTSDF uses voxelization and 3D jump flooding to generate a coarse field, then selectively refines near-surface voxels with ray tracing for real-time soft-shadow approximation [2210.06160].

SDF-Pack uses a truncated, discretized SDF as a local geometric state for robotic bin packing. The field assigns positive values to free space, negative values to occupied space, and small positive values to free space near existing geometry. Minimizing the SDF sampled throughout a candidate object encourages compact placement rather than merely maximizing contact area [2307.07356].

### Analytic and scattered-data representations

For articulated robots, Bernstein polynomial bases provide compact link-level SDFs with analytic spatial derivatives. Link fields are transformed by forward kinematics and combined through a minimum operation, producing task-space and joint-space derivatives suitable for collision avoidance and whole-body manipulation [2307.00533].

Radial basis function methods reconstruct a continuous SDF from scattered samples and inferred tangent points. A cubic polyharmonic kernel,

$$
\phi(r)=r^3,
$$

is combined with a linear polynomial and fitted to both original off-surface samples $(\mathbf x_i,d_i)$ and zero-valued tangent-point constraints $(\mathbf y_i,0)$ [2609.06209]. Partition-of-unity decomposition reduces the practical cost of local RBF solves.

## 3. Learning, supervision, and regularization

SDF learning requires more than fitting a zero crossing. A network can place an approximately correct surface while producing incorrect distance magnitudes, gradients, signs, or topology. Consequently, SDF systems combine distance regression with Eikonal, normal, geometric, photometric, semantic, or structural constraints.

### Distance and Eikonal objectives

A common Eikonal regularizer is

$$
\mathcal L_{\mathrm{eik}}
=
\mathbb E_{\mathbf x}
\left[
\left(\|\nabla_{\mathbf x}f(\mathbf x)\|-1\right)^2
\right].
$$

MVSDF combines direct signed-distance supervision from multi-view stereo with feature consistency, rendered-image losses, Eikonal regularization, and a surface-validity indicator. Direct supervision throughout the volume helps initialize topology, whereas feature and rendering losses refine local geometry [2108.09964].

iSDF does not receive ground-truth SDF values. It samples posed depth images along camera rays and constructs distance bounds from nearby observed surface points. Free-space samples are constrained to have positive values up to a bound, near-surface samples are directly supervised, gradients are aligned with approximate geometric directions, and the Eikonal constraint is applied away from surfaces [2204.02296].

HIO-SDF combines global voxel-derived SDF samples with local high-resolution samples from the current point cloud. A persistent coarse voxel map prevents catastrophic forgetting, while local ray-based samples provide detail near currently observed surfaces. The SIREN is optimized with SDF and Eikonal losses [2310.09463].

### Weak and indirect supervision

Dense ground-truth SDFs are often unavailable. SurroundSDF introduces the Sandwich Eikonal formulation: LiDAR points anchor the zero level set, occupied voxels provide negative-side constraints, and free voxels provide positive-side constraints. This supplies dense two-sided supervision without treating occupied voxel centers as exact surface points [2403.14366].

VesselSDF reformulates sparse-slice vascular segmentation as volumetric SDF regression. A binary occupancy U-Net first identifies likely vessels, and a second 3D U-Net refines the detached occupancy prediction into an SDF. Its loss includes SDF regression, occupancy classification, anisotropic Eikonal regularization, an adaptive Gaussian regularizer, and a surface regularizer. The Gaussian term suppresses distant field fluctuations while applying weaker smoothing near the zero level set [2506.16556].

### Geometric and graph regularization

Local SDF decompositions can produce discontinuities when regions receive insufficient samples. LGCL introduces a geometric similarity loss that encourages neighboring graph vertices to have related local codes without forcing all codes to be identical. Excessive similarity can oversmooth local variation, whereas insufficient regularization can produce surface vibrations at region boundaries [2108.08593].

In multi-shape latent representations, a generalization branch captures global shape structure and an overfitting branch uses spatial feature grids to preserve high-frequency detail. Surface-focused balanced sampling reduces cross-shape interference in shared grid features [2511.14539].

Some approaches impose hard geometric consistency rather than soft regularization. Greedy SDF interpolation tests candidate signed radii against global sphere constraints and guarantees mutual validity when each newly added sample preserves realizability [2605.01919].

## 4. Surface reconstruction and extraction

The zero level set must generally be converted into an explicit mesh. Marching Cubes remains common because it is simple and efficient, but it uses local sign changes and interpolation and may ignore distance information carried by samples far from the surface. This can cause oversmoothing, missing thin structures, disconnected components, and loss of sharp features.

Tangency-aware reconstruction addresses this limitation by interpreting each sample as a sphere that the surface must touch. *Reach For the Spheres* begins with an enclosing triangle mesh and minimizes a tangency-violation energy. Closest points on the current mesh are paired with target tangent points on the sample spheres, producing a sparse quadratic problem solved through implicit mesh-flow steps and remeshing [2308.09813].

RBF Your SDF fits the original SDF samples and inferred zero-valued tangent points. Tangent points are obtained either from collapsed exposed regions, which are especially informative near sharp corners, or by iterative optimization on each sample sphere. The resulting continuous field is extracted with Dual Contouring [2609.06209].

Greedy SDF interpolation instead augments coarse samples with values that are globally consistent with all previously assigned samples. It can refine SDF grids, repair pseudo-SDFs produced by CSG operations, and provide globally informed data for subsequent mesh extraction [2605.01919].

Neural systems commonly apply Marching Cubes to dense field evaluations. MVSDF extracts a mesh from a $512^3$ SDF volume and then removes invalid extrapolated surfaces using a learned validity indicator and graph cut. gSDF evaluates separate hand and object fields on a 3D grid and extracts their meshes with Marching Cubes. SplatSDF uses Marching Cubes at resolution $512$ after training-time Gaussian fusion.

Mesh extraction quality depends on field validity. A visually plausible zero level set does not guarantee accurate distance values away from the surface, a unit gradient, correct topology, or safe sphere-tracing behavior. This distinction is central to work on pseudo-SDF repair and globally consistent interpolation.

## 5. Rendering and inverse rendering

SDFs support both explicit surface rendering and SDF-to-density volume rendering.

### Surface rendering

MVSDF uses sphere tracing to locate zero-level-set intersections and implicit differentiation to propagate gradients through the intersection. Appearance is represented by a surface light field conditioned on position, SDF normal, and viewing direction. The method is explicitly surface-based and does not convert SDF values into NeRF-style densities [2108.09964].

RTSDF sphere-traces a ray through a voxel SDF. The field value determines a safe step size, while the minimum distance encountered along a light ray provides a heuristic for soft-shadow attenuation. Its approximate penumbra is smoother than the evaluated shadow-mapping configuration but its umbra is less physically accurate than distributed ray tracing [2210.06160].

### SDF-to-density rendering

NeuS-style systems convert SDF samples into opacity concentrated around the zero crossing. RISE-SDF uses NeuS-style normalized S-density for glossy-object rendering, deriving normals directly from the SDF gradient. Its neural SDF supplies surface location, rendering weights, normals, reflected directions, and geometry features to radiance, BRDF, environment-light, and indirect-illumination networks [2409.20140].

SplatSDF uses a logistic SDF-to-density mapping and volumetric rendering. Gaussian splatting provides a training-time surface anchor and fused embedding, but Gaussian information is not required during inference [2411.15468].

### Differentiable visibility

Visibility derivatives are difficult because silhouettes and occlusion boundaries are discontinuous. “A Simple Approach to Differentiable Rendering of SDFs” retains an explicit SDF surface for forward rendering and approximates the lower-dimensional visibility boundary with a thin band of near-silhouette samples. Relaxed silhouette points satisfy a near-zero SDF condition and a ray-directional local-minimum condition. The method accepts controlled bias in exchange for low variance and simple implementation [2405.08733].

### Probabilistic surface rendering

Stochastic Signed Distance Processes model the SDF along a ray as a stochastic process. The first surface intersection becomes a random first-passage time, and pixel colors are modeled as a mixture over possible intersection intervals. An Ornstein–Uhlenbeck residual process produces correlated SDF behavior along the ray, while Bayesian filtering estimates survival-conditioned crossing probabilities. NeuS is recovered as a perfectly correlated limiting case with a shared logistic residual [2606.20856].

This formulation provides an explicit distribution over intersection depth and supports uncertainty metrics such as CRPS, MAE, ECE, and sharpness. Its practical parallel approximation assumes that once the process becomes negative it remains negative, which can fail for rays crossing multiple surfaces or objects.

## 6. Applications

### Shape reconstruction, completion, and compression

Neural SDFs represent objects continuously without storing dense voxel grids. DeepSDF and related models support shape reconstruction, generation, and compression. LGCL improves compact-decoder reconstruction by distributing latent capacity over local graph-associated codes [2108.08593]. A common limitation is the tradeoff between decoder capacity, latent storage, and high-frequency detail.

SplatSDF improves neural SDF reconstruction using training-time 3D Gaussian guidance. It reports a mean DTU Chamfer distance of $0.58$ mm and a mean NeRF Synthetic Chamfer distance of $0.86$ mm in the supplied evaluations [2411.15468]. Compact latent-space models combine a shared generalization branch with a surface-focused overfitting branch, reporting high-fidelity reconstructions across Stanford, ShapeNet, and D-FAUST data [2511.14539].

### Multi-view and autonomous-scene reconstruction

MVSDF uses stereo-derived signed distances, multi-view feature consistency, and differentiable rendering to reconstruct complex scene topology without object masks [2108.09964]. SurroundSDF predicts continuous geometry and semantics from surround-view images, using LiDAR, occupancy labels, and the Sandwich Eikonal formulation. Its reported nuScenes reconstruction results include an absolute-relative depth error of $0.174$ [2403.14366].

HIO-SDF targets large, previously unknown robot workspaces. A coarse Voxfield map provides persistent memory and a SIREN provides continuous queries and gradients. It reports a mean global SDF error of $5.57$ cm, compared with $10.27$ cm for iSDF and $7.98$ cm for a $10$ cm Voxfield representation [2310.09463].

### Robotics and planning

iSDF supports online collision queries, collision costs, surface normals, and spatial gradients from a stream of posed depth images. It reports final SDF error below $6$ cm on each evaluated sequence and generally lower collision-cost error than voxel baselines [2204.02296].

Robot Distance Fields use Bernstein-polynomial link SDFs and forward kinematics to produce differentiable geometry in task and joint spaces. The representation has been applied to collision avoidance, contact planning, and whole-body manipulation with 7-axis Franka Emika robots [2307.00533].

SDF-Pack uses a truncated SDF to rank feasible object placements in robotic bin packing. Its SDF-minimization heuristic produced higher packed volume and compactness than deepest-bottom-left, maximum-touching-area, and heightmap heuristics in the reported 1,000-case evaluation [2307.07356].

### Rendering and graphics

RTSDF generates approximate dynamic SDFs through voxelization, jump flooding, selective ray tracing, and temporal accumulation. In the reported configuration it achieved approximately $97$ frames per second, while producing smoother penumbras than the evaluated shadow-mapping method [2210.06160].

SDFs also support differentiable inverse rendering. RISE-SDF reconstructs glossy-object geometry, albedo, metallicity, roughness, environment illumination, and indirect reflections using a shared SDF backbone [2409.20140]. The thin-band differentiable renderer supports geometry, material, lighting, and camera optimization without requiring an explicit silhouette data structure [2405.08733].

### Medical imaging and geometric reasoning

VesselSDF uses a continuous SDF to reconstruct thin, branching vascular networks from sparse CT slices. On the Hepatic Vessels dataset it reports Dice $0.72$, Chamfer distance $0.68$, and Hausdorff distance $4.1$, outperforming the compared voxel-classification baselines in the supplied table [2506.16556].

GeoSDF represents plane-geometry diagrams as differentiable fields whose elements and relationships are optimized numerically. Its symbolic language expresses incidence, equality, angles, parallelism, perpendicularity, area, order, containment, and convexity. The optimized field can be rendered and directly self-verified, with the reported GeoQA solving accuracies of $94.5\%$ in completion and $95.9\%$ in multiple-choice settings [2506.13492].

## 7. Limitations, trade-offs, and research directions

SDFs do not eliminate geometric ambiguity. A field can have a plausible zero level set while having incorrect distances, gradients, topology, or extrapolation. Neural fields may oversmooth fine structures, forget previously observed regions, or confidently hallucinate unobserved geometry. Voxel fields preserve observations incrementally but require memory proportional to spatial extent and resolution. Hybrid systems such as HIO-SDF explicitly combine these complementary properties [2310.09463].

Regularization introduces competing objectives. Eikonal losses encourage distance-like behavior but can oversmooth singularities. Graph similarity losses reduce local fluctuations but can collapse legitimate variation. Gaussian smoothness suppresses floating structures but can erase thin geometry if over-weighted. Surface-focused sampling improves detail but may leave far-field regions dependent on a learned prior.

Topology is another central limitation. Implicit fields can represent disconnected and concave surfaces, but optimization and extraction do not guarantee the correct topology. MVSDF depends on initial MVS topology, SDF reconstruction through tangency flow preserves the topology of its enclosing initialization, and neural systems can produce disconnected components or spurious zero crossings.

Computational cost varies substantially. A voxel lookup is inexpensive but discrete; neural evaluation provides continuous queries and automatic differentiation but is more expensive. Global RBF systems are dense, although partition-of-unity methods reduce practical scaling. Differentiable rendering adds sphere tracing, derivative evaluation, and visibility handling. Probabilistic SDF rendering adds sequential or approximate Bayesian filtering.

Several research directions follow from these trade-offs:

- **Globally valid interpolation**: enforce sphere nonpenetration, tangency, sidedness, and uncovered-point conditions rather than only local smoothness or Eikonal behavior [2605.01919].
- **Sharp-feature reconstruction**: retain off-surface distance information and avoid single-normal assumptions near corners and creases [2609.06209].
- **Uncertainty-aware fields**: model uncertainty in geometry and first-intersection depth rather than treating neural extrapolation as certified distance [2606.20856].
- **Incremental large-scale mapping**: combine persistent coarse memory, local detail, continuous queries, and bounded history [2310.09463].
- **Geometry–appearance factorization**: use SDF geometry jointly with radiance, BRDF, environment illumination, semantics, and indirect-light models [2409.20140].
- **Compact multi-shape latent spaces**: preserve high-frequency detail while reducing per-shape storage and cross-shape interference [2511.14539].
- **Task-specific geometric priors**: incorporate articulated kinematics, vascular continuity, packing constraints, or symbolic geometric relations without abandoning continuous implicit geometry.

The defining strength of an SDF is the conjunction of surface representation, metric distance, sign, and differentiability in one field. Its practical effectiveness depends on whether those properties are preserved by the chosen parameterization, learning objective, discretization, interpolation, rendering model, and extraction procedure.

Source: https://www.emergentmind.com/topics/signed-distance-field-sdf