Implicit Neural Representations as Continuous Coordinate Mappings

83 concepts79 papers1 proposal

For decades, digital media and computational physics have relied on discrete representations of data. Images are stored as grids of pixels, 3D shapes as voxel arrays or polygonal meshes, and physical simulations as values locked to a discrete computational mesh. While intuitive, this grid-locked paradigm introduces fundamental limitations: memory consumption scales poorly with resolution, and interpolating between discrete points introduces artifacts and approximation errors.

Implicit Neural Representations (INRs)—often referred to as coordinate-based neural fields—fundamentally rethink this approach. Instead of storing data as a discrete array of samples, an INR parameterizes a signal as a continuous mathematical function. This function is approximated by a neural network, typically a Multi-Layer Perceptron (MLP), which maps a continuous spatial or temporal coordinate directly to its corresponding attribute value (such as color, density, or physical velocity).

Discrete Representation (Voxel Grid/Pixels):
[Index (i, j, k)] ---> [Lookup Table] ---> [Value]

Implicit Neural Representation (Neural Field):
[Coordinate (x, y, z)] ---> [ MLP (θ) ] ---> [Value]

By shifting the problem from signal processing transforms to network parameter optimization, INRs decouple the resolution of the data from the underlying storage mechanism.

Shifting from Grids to Functions

To understand why this shift is powerful, consider the challenge of 3D object reconstruction. In a comprehensive roadmap of deep learning-based 3D reconstruction, Xian-Feng Han et al. observe in Image-Based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era that modeling geometry as a continuous mathematical field allows for virtually infinite resolution. Instead of allocating memory for empty space in a 3D bounding box, the network weights act as a highly compressed prior capable of representing complex topologies by querying the field at any arbitrary coordinate.

This capability is particularly transformative when data is inherently continuous but captured through sparse or noisy observations. For example, in cryo-electron microscopy (cryo-EM), researchers attempt to reconstruct 3D protein structures from noisy 2D projections. Ellen D. Zhong et al. demonstrate in Reconstructing continuous distributions of 3D protein structure from cryo-EM images that mapping coordinates and a latent state to density values bypasses discrete voxel grids entirely. This allows for the smooth interpolation of structural configurations, enabling scientists to capture the full spectrum of a protein's structural flexibility.

Subsequent work by Axel Levy et al. in CryoDRGN-AI: Neural ab initio reconstruction of challenging cryo-EM and cryo-ET datasets and Ellen D. Zhong et al. in CryoDRGN2: Ab initio neural reconstruction of 3D protein structures from real cryo-EM images confirms that this continuous formulation avoids the memory overhead of discrete grids while resolving fine structural details from scratch.

Why It Works: Resolution Independence and Differentiability

The intuition behind the success of INRs rests on two core pillars: resolution independence and analytical differentiability.

Resolution Independence

Because the input to an INR is a continuous vector (e.g., $\mathbf{x} \in \mathbb{R}^3$), the representation can be queried at an arbitrary level of granularity. You are no longer constrained by the resolution at which the data was originally sampled or trained.

In generative modeling, Emilien Dupont leverages this in Generative Models as Distributions of Functions to build a framework (GASP) that treats images, 3D shapes, and climate data as continuous stochastic processes. By treating data points as neural fields, the model can synthesize data at any resolution or on complex manifolds, such as spheres, bypassing the discretization constraints that plague standard convolutional architectures.

Analytical Differentiability

Because the coordinate mapping is parameterized by a neural network, the entire representation is differentiable with respect to its input coordinates via standard backpropagation. This means we can compute exact, analytical spatial or temporal derivatives of the represented signal without relying on error-prone finite-difference approximations on a grid.

This property makes INRs an exceptional tool for solving partial differential equations (PDEs) and physics-informed problems:

Overcoming the Discretization Curse in Physical Systems

Traditional numerical simulations suffer from the "curse of discretization"—to get higher accuracy, you must scale the grid density, which causes memory and computational requirements to explode. INRs offer a mesh-free alternative.

In Continuous PDE Dynamics Forecasting with Implicit Neural Representations, Yuan Yin et al. combine INRs with Latent ODEs to forecast physical systems continuously across space and time. Because the spatial fields are represented as continuous functions, the model can perform inference on irregular meshes or at higher resolutions than those seen during training.

Similarly, Peter Yichen Chen et al. introduce CROM in CROM: Continuous Reduced-Order Modeling of PDEs Using Implicit Neural Representations, demonstrating that learning a low-dimensional manifold of continuous vector fields rather than their discretizations allows models to capture complex physics while remaining completely agnostic to the underlying mesh resolution.

Data Compression as Model Compression

If a neural network can perfectly fit a continuous signal by mapping coordinates to values, then the weights of that network effectively contain the signal itself. This realization opens up a new paradigm: data compression via model compression.

Traditional Compression (JPEG):
[Pixels] ---> [DCT / Quantization] ---> [Bitstream]

INR Compression (COIN):
[Coordinates] ---> [ Overfitted MLP (θ) ] ---> [Target Colors]
                               |
                        (Store Weights θ)

Emilien Dupont et al. proposed this concept in COIN: COmpression with Implicit Neural representations by overfitting a small MLP to an individual image. Instead of storing pixel arrays, they store the network's weights. This approach shifts the compression task from traditional signal transforms to neural network parameter optimization.

To make this practical across diverse data modalities, Yannick Strümpler et al. show in Implicit Neural Representations for Image Compression that combining meta-learning with refined architectures allows INRs to compete effectively with established standards like JPEG2000.

Addressing the High-Frequency Challenge

Despite their elegance, standard MLPs face a fundamental mathematical hurdle when acting as coordinate mappings: spectral bias. Deep networks naturally bias toward learning low-frequency components first, meaning a standard MLP mapping $(x,y) \to \text{color}$ will yield blurry images, struggling to capture sharp edges and fine textures.

To resolve this, modern INRs employ coordinate remapping strategies, such as mapping inputs into a Fourier-variant space or utilizing hybrid, learnable data structures. In DINER: Disorder-Invariant Implicit Neural Representation, Hao Zhu et al. address this by prepending a learnable hash-table to a standard MLP backbone. By re-arranging signal coordinates into a smoother latent space before passing them to the network, they allow compact networks to represent high-frequency signals with exceptional accuracy and speed.

By treating the world not as a collection of discrete blocks, but as a continuous field of coordinates, Implicit Neural Representations provide a unified, resolution-agnostic framework that bridges the gap between discrete digital computation and the continuous physical world.

Key Foundational Work

SHARP 2020: The 1st Shape Recovery from Partial Textured 3D Scans Challenge Results (2020) — contributed 2 concept elements to this direction.
Learning Signal-Agnostic Manifolds of Neural Fields (2021) — contributed 2 concept elements to this direction.
IntraTomo: Self-supervised Learning-based Tomography via Sinogram Synthesis and Prediction (2021) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_neural_representation_coordinat_96e3aabe,
  title = {implicit neural representation (coordinate-based neural representation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-neural-representation-coordinate-based-neural-representation/rdr_96e3aabee1b0411c80e79e7935744844},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_implicit_neural_representation_coordinat_96e3aabe_v1,
  title = {implicit neural representation (coordinate-based neural representation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-neural-representation-coordinate-based-neural-representation/rdr_96e3aabee1b0411c80e79e7935744844/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake