Optimizing Unsigned Distance Fields at the Surface Interface

17 concepts15 papers1 proposal

When representing 3D geometry implicitly, researchers face a fundamental choice between Signed Distance Fields (SDFs) and Unsigned Distance Fields (UDFs). SDFs are mathematically elegant; they define a surface as the zero-level set of a function that transitions smoothly from negative values (inside the object) to positive values (outside). This smooth transition means the function is differentiable at the surface, allowing neural networks to easily locate the boundary using gradient-based optimization. However, SDFs require a strict definition of "inside" and "outside," making them incapable of representing open surfaces, thin structures, or complex internal geometries like clothing or car interiors.

To overcome this limitation, Chibane et al. introduced Neural Unsigned Distance Fields (NDF), which discard the sign and map every spatial coordinate to its absolute, non-negative distance to the closest surface. While UDFs successfully represent arbitrary, non-watertight topologies, they introduce a severe mathematical penalty: a non-differentiable "cusp" or "kink" exactly at the surface interface where the distance reaches zero.

A growing body of research focuses on resolving this zero-level set non-differentiability to stabilize training, eliminate geometric artifacts, and enable high-fidelity surface reconstruction.


The Geometry of the Cusp

To understand why standard UDFs fail during gradient-based optimization, consider a one-dimensional example. A signed distance function behaves like the linear function $f(x) = x$. Its derivative is a constant $1$ everywhere, providing a clear, continuous signal that guides an optimizer directly to the zero-crossing at $x = 0$.

An unsigned distance function behaves like the absolute value function $f(x) = |x|$. As you approach the surface from either side, the distance decreases linearly. But exactly at the surface ($x = 0$), the function forms a sharp $V$-shaped cusp.

Signed Distance: f(x) = x           Unsigned Distance: f(x) = |x|
           ^                                   ^
           |   /                               | \   /
           |  /                                |  \ /
----------0---------->                     ----0---------->
        /  |                                   |
       /   |                                   |

This geometric kink introduces three severe challenges for neural networks:

  1. Non-Differentiable Singularity: As noted by Long et al. and Hu et al., the derivative is undefined exactly where we care most—at the surface boundary. Neural networks struggle to fit these sharp mathematical discontinuities, often resulting in noisy, "shattered," or fragmented surface extractions, as documented by Ma et al. and Zhang et al..
  2. Vanishing and Ambiguous Gradients: Because the function is strictly non-negative, the gradient magnitude drops or oscillates unpredictably near zero. As Fainstein et al. and Kong et al. show, this vanishing gradient pathway prevents the network from accurately localizing the surface and makes it incredibly difficult to calculate reliable surface normals for shading or rendering.
  3. Eikonal Constraint Incompatibility: True distance fields must satisfy the Eikonal equation, meaning the gradient magnitude must equal one ($| \nabla f | = 1$) almost everywhere. However, Walder et al. and Xu et al. demonstrate that enforcing a unit gradient globally is fundamentally incompatible with the instantaneous direction flip required at a UDF's zero-distance interface, leading to severe training instabilities.

To solve these issues, recent papers follow three main strategies: shifting from scalar distances to vector fields, leveraging off-surface geometric properties, or mathematically reformulating the distance metric itself.


Shifting from Distance to Direction: Vector Fields

One of the most intuitive ways to eliminate the zero-level set cusp is to change what the network predicts. Instead of forcing the network to learn a scalar distance that bottoms out at zero, we can train it to predict a continuous vector field that points directly to the nearest surface point.

This is the core intuition behind the Closest Surface-Point (CSP) representation introduced by Venkatesh et al. and the Gradient Distance Function (GDF) proposed by Stella et al..

Returning to our 1D example, instead of predicting $f(x) = |x|$, the network predicts the displacement vector $g(x) = -x$. At any point $x$, adding $g(x)$ to your position lands you exactly on the surface ($x + (-x) = 0$).

Scalar UDF Cusp:                        Vector Field Smoothness:
           ^                                       ^
          / \                                     /
         /   \                                   /
--------0--------->                     --------0--------->
                                               /
                                              /

Because $g(x) = -x$ is a perfectly smooth, linear function, its derivative is continuous everywhere—including at zero. By predicting vectors rather than scalar magnitudes, these frameworks maintain stable, non-vanishing gradients directly at the surface interface, enabling faster convergence and preventing the formation of holes or fragmented meshes.


Borrowing Strength from the Neighborhood

If the mathematical instability of a UDF is strictly confined to the zero-level set, then the field just a fraction of a millimeter away from the surface remains perfectly differentiable and well-behaved. A second class of methods exploits this property by using "off-surface" information to guide the boundary's optimization.

For instance, Chibane et al. bypassed the boundary discontinuity by sampling gradients at a slight distance from the surface during training. Taking this concept further, Ma et al. introduced Level Set Projection.

Instead of forcing the network to directly resolve the non-differentiable kink at $f(x)=0$, Level Set Projection samples points in the surrounding neighborhood where the gradients are clean and stable. It uses these reliable off-surface gradients to mathematically "pull" and project the zero-level set into a continuous, high-fidelity shape.

Similarly, Li et al. developed Implicit Filtering (I-Filtering), which applies smoothing operations across the implicit field to preserve sharp geometric features while mitigating the unoriented gradient noise that typically plagues raw UDF boundaries.


Reformulating the Distance Metric

A third approach alters the underlying mathematical formulation of the distance function to smooth out the cusp entirely, ensuring the representation is inherently differentiable.

Hyperbolic Scaling

In DUDF, Fainstein et al. apply a hyperbolic scaling transformation to the unsigned distance field. By transforming the distance metric through a non-linear lens, they soften the sharp mathematical cusp into a smooth, differentiable curve at the zero interface, restoring the reliable gradients necessary for computing accurate surface normals and shading.

Monge-Ampère Regularization

Rather than using raw distances, Yang et al. propose the Scaled-Squared Distance Function ($S^2DF$). Squaring the distance function ($f(x) = x^2$) naturally eliminates the sharp corner at zero, turning the $V$-shape into a smooth parabola. However, squaring the distance breaks the Eikonal constraint ($|\nabla f| = 1$). To counteract this, the authors introduce a regularizer based on a Monge-Ampère-type partial differential equation, balancing boundary smoothness with global geometric accuracy.

Raw UDF Cusp (|x|):                     Squared Distance Parabola (x²):
           ^                                       ^
          / \                                     / \
         /   \                                   /   \
--------0--------->                     --------(---)----->
                                                 \_/

Relaxing Non-Negativity

Xu et al. target the root cause of the gradient oscillation: the strict non-negativity constraint ($f(x) \ge 0$). In their framework, DEUDF, they relax this constraint near the surface interface and pair it with local normal alignment. Allowing the function to smoothly transition without a hard floor at zero prevents gradient vanishing and preserves fine, high-frequency geometric details.


Summary

The evolution of neural implicit surfaces highlights a clear trajectory. While Signed Distance Fields provided the initial foundation for implicit modeling, their topological rigidity prompted the shift toward Unsigned Distance Fields. Today, the field is moving past the raw, non-differentiable formulations of early UDFs. Whether through vector field reparameterization, off-surface gradient projection, or smooth metric reformulations, modern architectures ensure that the surface interface remains differentiable—combining the topological freedom of unsigned fields with the optimization stability of signed ones.

Key Foundational Work

Neural Unsigned Distance Fields for Implicit Function Learning (2020) — contributed 1 concept elements to this direction.
Implicit surface modelling as an eigenvalue problem (2005) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 14 papers
  2. Version 1 · 2026-10-04 · 27 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_unsigned_distance_field_optimization_54f57e95,
  title = {unsigned distance field optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsigned-distance-field-optimization/rdr_54f57e959ff24d85b64bfe96c28aad8b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_unsigned_distance_field_optimization_54f57e95_v2,
  title = {unsigned distance field optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unsigned-distance-field-optimization/rdr_54f57e959ff24d85b64bfe96c28aad8b/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake