Implicit neural representations have transformed 3D computer vision and graphics. By representing geometry as the continuous level set of a neural network, methods can reconstruct intricate shapes without the memory constraints of traditional voxels or the fixed topologies of meshes. However, the dominant paradigms—Occupancy Networks and Signed Distance Functions (SDFs)—suffer from a fundamental topological limitation: they require shapes to be "watertight."
A surface is watertight if it perfectly partitions 3D space into a distinct "inside" and "outside." As noted by Tiago Novello in a survey of Geometric implicit neural representations for signed distance functions, this mathematical assumption breaks down when encountering real-world geometries like open clothing, thin sheets, internal walls, or raw, fragmented sensor scans. To model these arbitrary topologies, researchers have shifted toward Unsigned Distance Fields (UDFs).
The Limitation of the Sign
To understand why traditional implicit functions fail on open surfaces, consider a simple 3D model of a t-shirt or an open jacket. If you sample a point in space near the sleeve, is it inside or outside the garment? Because the fabric is a thin shell with open boundaries, a path can be drawn from the "interior" to the "exterior" without ever piercing the surface.
When an SDF or occupancy network is forced to represent such a shape, it must invent a closed volume. This mathematical coercion introduces severe geometric artifacts, such as artificial "webbing" that seals openings, or erratic holes where the network struggles to resolve the inside-outside ambiguity.
Signed Distance Field (SDF) Unsigned Distance Field (UDF)
Outside (+) Proximity (d > 0)
─────────────────── ───────────────────
─── Surface (0) ─── ─── Surface (0) ───
─────────────────── ───────────────────
Inside (-) Proximity (d > 0)
(Requires Closed Volume) (Supports Open Shells)
Unsigned Distance Fields bypass this requirement entirely. As introduced by Julian Chibane et al. in Neural Unsigned Distance Fields for Implicit Function Learning, a UDF maps any spatial coordinate $x$ to its absolute, unsigned distance to the closest surface point:
$$f(x) = \min_{p \in \mathcal{S}} \|x - p\|$$
Because $f(x) \geq 0$ everywhere, the surface is defined simply as the zero-level set $\mathcal{S} = \{x \mid f(x) = 0\}$. By discarding the sign, UDFs can natively represent open boundaries, infinitely thin sheets, and complex nested or multi-layered structures.
The Gradient Discontinuity Problem
While UDFs solve the topological constraints of SDFs, they introduce a severe optimization challenge known as the gradient discontinuity problem.
In an SDF, the field transitions smoothly from negative to positive across the surface. The gradient $\nabla f(x)$ is well-behaved and points consistently along the surface normal. In a UDF, however, the distance values decrease as you approach the surface from either side, hit exactly zero at the surface, and immediately begin to increase again.
SDF Distance Profile UDF Distance Profile
| |
| / \ /
______|_/______ ______ _______\_/__ ______
/| |/
/ | |
This creates a sharp, V-shaped valley at the zero-level set. Mathematically, the function is non-differentiable at the surface; the gradient vector abruptly flips direction by 180 degrees. For neural networks trained via backpropagation, this discontinuity makes optimization near the surface highly unstable, as gradients vanish or conflict exactly where precision matters most.
To stabilize training, researchers have developed specialized geometric priors. For example, Fang Zhao et al. introduced Learning Anchored Unsigned Distance Functions with Gradient Direction Alignment for Single-view Garment Reconstruction, which explicitly aligns the network's gradient directions during training to resolve the valley instability.
Taking a more radical approach, Ajad Chhatkuli et al. proposed Neural Vector Fields for Implicit Surface Representation and Inference. Instead of predicting a scalar distance, their network maps spatial coordinates directly to a unit vector field pointing toward the nearest surface point. This formulation avoids the scalar UDF gradient discontinuity entirely while inherently encoding surface normals.
Differentiable Rendering without Signs
To learn UDFs directly from multi-view images without 3D supervision, the field has had to reinvent neural volume rendering. Standard volume rendering pipelines, like NeRF or VolSDF, map signed distances or occupancies to ray termination probabilities. Replicating this with a purely positive UDF requires a different mathematical formulation.
The authors of NeuralUDF: Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces with Arbitrary Topologies solved this by constructing a symmetric joint probability density function. Instead of looking for a sign change along a camera ray, the rendering engine tracks the UDF values and evaluates a weight function that peaks sharply when the UDF value reaches zero. Similarly, Yu-Tao Liu et al. developed NeUDF: Leaning Neural Unsigned Distance Fields with Volume Rendering, which uses an unbiased weighting function to ensure that thin-walled structures like sleeves or leaves can be reconstructed from images without producing hollow volumetric artifacts.
Recent work has further enhanced the speed and accuracy of these rendering pipelines. Wenyuan Zhang et al. in Learning Unsigned Distance Functions from Multi-view Images with Volume Rendering Priors replaced handcrafted rendering equations with a pre-trained neural network that learns the physics of mapping unsigned distances to pixels. To accelerate optimization, Shujuan Li introduced GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting, which fits flat 2D Gaussian primitives onto surfaces to provide fast, explicit geometric supervision for a continuous continuous UDF field.
Surface Extraction: Beyond Marching Cubes
Once a neural network has learned a UDF, converting that continuous field into a discrete polygon mesh for downstream applications is non-trivial. The classic Marching Cubes algorithm relies on identifying edges where voxel vertices exhibit a sign change (moving from inside to outside). Because UDF values are strictly non-negative, standard Marching Cubes cannot detect the surface.
To extract open surfaces from UDFs, researchers have devised sign-agnostic meshing strategies:
- Intersection Detection: In Neural Dual Contouring, Zhiqin Chen et al. bypassed sign-based logic entirely. Their sign-agnostic variant (UNDC) trains a network to predict edge-crossing flags and vertex positions directly from the field, allowing the algorithm to snap vertices onto thin sheets and open boundaries.
- Off-Surface Optimization: Congyi Zhang et al. developed Surface Extraction from Neural Unsigned Distance Fields (DualMesh-UDF), which uses an off-surface sampling strategy and iterative optimization to project a dual mesh onto the zero-level set without introducing artificial holes.
- Geometry-Guided Marching: In GeoUDF: Surface Reconstruction from 3D Point Clouds via Geometry-guided Distance Representation, Junhui Hou and Siyu Ren introduced an Edge-based Marching Cube algorithm designed specifically to navigate the geometric valleys of unsigned fields.
Applications and Generative Modeling
The topological freedom of UDFs has unlocked capabilities across several challenging domains in 3D vision:
Garment and Human Modeling
Because clothing is inherently composed of open, thin surfaces, it has become the primary benchmark for UDFs. Early frameworks like SMPLicit: Topology-Aware Generative Model for Clothed People by Enric Corona et al. demonstrated that UDFs could represent diverse clothing topologies—from sleeveless tops to open jackets—on top of parametric body models. Subsequent works like DrapeNet: Garment Generation and Self-Supervised Draping by Luca De Luigi et al. leveraged UDFs for physics-based draping without fixed mesh templates.
Real-World Scene and Object Reconstruction
Raw data from LiDAR or depth sensors rarely forms a perfect closed manifold. Ajad Chhatkuli et al. demonstrated in Self-supervised Shape Completion via Involution and Implicit Correspondences that UDFs are fundamentally superior to SDFs for completing raw, partial scans. This flexibility is critical when reconstructing complex human-object interactions (CHORE by Xianghui Xie et al.), sparse scenes with hidden surfaces (3DFIRES by Linyi Jin et al.), or open-boundary structures from single-view inputs (NU-MCC by Stefan Lionar et al. and IPoD by Yushuang Wu et al.). Remarkably, the formulation can even scale down to 1D structures, as shown in 3D Neural Edge Reconstruction by Lei Li and Zehao Yu, which uses UDFs to represent wireframe curves that lack any volumetric mass.
Generative Diffusion Models
UDFs have also been successfully integrated into latent diffusion pipelines to synthesize brand-new 3D shapes. Junsheng Zhou et al. developed Learning Consistency-Aware Unsigned Distance Functions Progressively from Raw Point Clouds to train UDFs directly on unorganized data, and later expanded this into UDiFF: Generating Conditional Unsigned Distance Fields with Optimal Wavelet Diffusion to generate open-surface objects from text or image prompts.
Concurrently, Zhengming Yu et al. introduced Surf-D: Generating High-Quality Surfaces of Arbitrary Topologies Using Diffusion Models, showing that a single UDF-based diffusion architecture can seamlessly generate both solid shapes and open structures with low memory overhead. These generative UDF fields underpin modern text-to-3D garment pipelines, such as WordRobe by Astitva Srivastava et al. and Design2Cloth by Jiali Zheng and Stefanos Zafeiriou, producing simulation-ready open meshes.
Nested Geometry Generation (e.g., FullFormer)
┌─────────────────────────────────────────┐
│ Outer Shell (UDF = 0) │
│ ┌───────────────────────────────────┐ │
│ │ Inner Structure (UDF = 0) │ │
│ └───────────────────────────────────┘ │
└─────────────────────────────────────────┘
(Allowed because UDFs evaluate absolute
proximity without interior constraints)
Finally, by discarding the inside-outside constraint, UDFs allow generative models to create nested surfaces—shapes inside other shapes. As demonstrated by Tejaswini Medi et al. in FullFormer: Generating Shapes Inside Shapes, a UDF can model a car's exterior shell and its internal cabin structures (like seats and steering wheels) simultaneously within a single continuous field. By decoupling implicit representations from the strictures of solid volumes, unsigned distance fields provide a mathematically rigorous path toward representing the full, unconstrained complexity of the physical world.
Key Foundational Work
Neural Unsigned Distance Fields for Implicit Function Learning (2020) — contributed 2 concept elements to this direction.
SMPLicit: Topology-Aware Generative Model for Clothed People (2021) — contributed 1 concept elements to this direction.
Learning Anchored Unsigned Distance Functions with Gradient Direction Alignment for Single-view Garment Reconstruction (2021) — contributed 1 concept elements to this direction.