Implicit geometric representations have revolutionized 3D shape modeling. For years, Signed Distance Fields (SDFs) were the gold standard. An SDF defines a shape by mapping every point in space to its closest distance to the surface, using negative signs for the inside and positive signs for the outside. To extract a polygonal mesh from an SDF, standard algorithms like Marching Cubes look for "sign flips" along grid edges. If one grid corner is negative and its neighbor is positive, the surface must cross the edge between them. By linearly interpolating these values, we can precisely pinpoint the surface boundary.
However, SDFs possess a fundamental limitation: they require a well-defined interior and exterior. This restricts them to "watertight" manifolds. If you want to represent an open surface—such as a piece of clothing, a single leaf, an internal car panel, or complex nested geometry—SDFs break down.
Unsigned Distance Fields (UDFs) solve this by discarding the sign entirely. A UDF outputs only the absolute distance to the nearest surface ($f(x) \ge 0$). While this allows UDFs to represent arbitrary topologies and open boundaries, it introduces a severe computational challenge: without sign transitions, traditional iso-surface extraction algorithms cannot locate the surface. The surface is no longer a zero-crossing; it is a local minimum valley where the field hits exactly zero and bounces back up.
SDF (Sign Flip): Positive (+) ---> [ Surface (0) ] ---> Negative (-)
UDF (No Flip): Positive (+) ---> [ Surface (0) ] ---> Positive (+)
Converting these non-negative fields into clean, single-layered polygonal meshes without introducing topological errors or massive computational overhead is an active area of research.
The Instability of the Zero Level-Set
Why can't we simply extract the voxels where $f(x) = 0$? In practice, distance fields are represented via continuous neural networks or discrete voxel grids. Near the target surface, the mathematical properties of a UDF degrade. As the distance approaches zero, the field's gradients become highly unstable, noisy, or vanish entirely.
As noted by Junsheng Zhou in UDiFF: Generating Conditional Unsigned Distance Fields with Optimal Wavelet Diffusion, this extreme sensitivity to noise means that even minor field imperfections cause standard extraction techniques to leave holes or fragmented geometry. Furthermore, because a UDF lacks orientation data, naive extraction methods often suffer from "double-layer" artifacts—mistakenly generating two distinct mesh surfaces pressed against each other instead of a single, infinitely thin open sheet.
To overcome these challenges, modern research splits into three main strategies: adapting Marching Cubes using local gradients or pseudo-signs, leveraging global geometric optimization like double covering, and formulating hybrid implicit fields.
Strategy 1: Adapting Marching Cubes via Gradients and Pseudo-Signs
Instead of abandoning Marching Cubes, one approach is to modify how the algorithm evaluates grid cells. While UDF values do not change sign across a surface, their gradients do. As you cross an open surface, the vector pointing to the closest surface point suddenly flips 180 degrees.
Benoît Guillard et al. leveraged this insight in MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks by extending Marching Cubes to reason about gradient direction flips rather than scalar sign changes. This allows for fast, differentiable mesh extraction directly from raw UDFs.
Gradient Vector Behavior Crossing a UDF Surface:
Point A (Right) ---> [ Surface ] <--- Point B (Left)
Another way to preserve the efficiency of Marching Cubes is to dynamically estimate local signs. Federico Stella et al. introduced this concept in Neural Surface Detection for Unsigned Distance Fields. Instead of relying on hand-crafted heuristics, they trained a neural network to predict valid local sign configurations for individual grid cells. This effectively converts the UDF into a local "pseudo-SDF," allowing standard Marching Cubes to run safely without sign conflicts.
To combat the noise inherent to high-resolution grids, Federico Stella et al. later developed High Resolution UDF Meshing via Iterative Networks. This method uses an iterative network that allows neighboring voxels to communicate over multiple passes, resolving local ambiguities and filling in missing gaps where noise would otherwise cause holes.
Strategy 2: Double Covering and Spatial Partitioning
A second school of thought avoids local voxel-level decisions entirely, opting instead for global geometric optimization to guarantee topological consistency.
In Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering, Fei Hou et al. proposed a two-step "double covering" technique. Instead of trying to find the exact zero-surface immediately, the algorithm extracts an iso-surface at a small positive distance $\epsilon$. This creates a thin, watertight volumetric shell (a double cover) that safely encloses the true open surface. The algorithm then collapses this shell back down onto the zero level-set using optimization and graph cuts, eliminating non-manifold artifacts.
The authors of DCUDF2: Improving Efficiency and Accuracy in Extracting Zero Level Sets from Unsigned Distance Fields (Xuhui Chen et al.) advanced this paradigm by introducing self-adaptive weights and dynamic topology correction, significantly accelerating the projection phase while preserving sharp geometric features.
Double Covering Pipeline:
[Raw UDF]
│
▼
[Extract ε-Shell] ───> (Encloses open surface as a watertight volume)
│
▼
[Graph-Cut Collapse] ───> (Projects shell inward to form a single thin mesh)
However, even double-covering methods can struggle with complex non-manifold structures, such as T-junctions or intersecting planes. To address this, Xuhui Chen et al. introduced MIND in MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction. MIND reframes surface extraction as a spatial partitioning problem. Instead of tracking the surface itself, it treats the 3D space as a collection of distinct "material" regions. The target surface is then extracted naturally as the interface where these different material zones meet, completely bypassing the risk of double-layer errors.
Strategy 3: Hybrid and Alternative Field Formulations
A third approach modifies the underlying representation itself, embedding orientation or structural cues directly into the implicit field during training so that extraction is inherently cleaner.
- Hybrid Fields: Li Wang et al. proposed HSDF in HSDF: Hybrid Sign and Distance Field for Modeling Surfaces with Arbitrary Topologies. HSDF decouples geometry and orientation by predicting a continuous distance field and a separate sign field simultaneously, fusing them via gradient-based optimization to allow high-fidelity meshing of thin clothes and open leaves.
- Multi-Pole Fields: In 3PSDF: Three-Pole Signed Distance Function for Learning Surfaces with Arbitrary Topologies, the author (Weikai Chen) introduced a "Three-Pole" formulation. By adding a third
NULLlabel alongside the traditional "inside" and "outside" classifications, the network can explicitly flag open boundaries, maintaining direct compatibility with standard industry meshing pipelines. - Vector Fields: Moving away from scalar fields entirely, the author of Neural Vector Fields for Implicit Surface Representation and Inference (Ajad Chhatkuli) proposed mapping space directly to unit vectors that point toward the nearest surface. This representation inherently encodes surface normals and uses a specialized "flux density" metric to stabilize directional information exactly where UDF gradients typically degrade.
By rethinking how implicit fields represent boundaries—whether through gradient analysis, global shell projections, or multi-state field formulations—these methods allow modern graphics pipelines to reconstruct and generate highly complex, non-watertight real-world structures with the same mathematical elegance once reserved for solid, watertight shapes.
Key Foundational Work
SuperUDF: Self-supervised UDF Estimation for Surface Reconstruction (2023) — contributed 2 concept elements to this direction.
FullFormer: Generating Shapes Inside Shapes (2023) — contributed 2 concept elements to this direction.
NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering (2023) — contributed 2 concept elements to this direction.