Integrated Positional Encoding (IPE)

• IPE is an area-aware positional encoding that aggregates frequency features over a pixel's spatial extent to mitigate aliasing.
• IPE improves arbitrary-scale image super-resolution by fusing spatial location with support size, preserving fine details while reducing artifacts.
• IPE integrated into LIIF architectures enhances PSNR/SSIM performance and eliminates checkerboard artifacts across multiple benchmarks.

Integrated Positional Encoding (IPE) is a variant of traditional positional encoding schemes for coordinate-based neural implicit representations. Unlike standard point-sampled positional encodings, which represent a coordinate as a single point (typically the center of a pixel), IPE explicitly aggregates frequency-encoded information over a spatial region—accounting for pixel area or cell extent. This area-aware design addresses aliasing at high frequencies and resolution loss at lower output scales, providing principled fusion of spatial location and support size in applications such as arbitrary-scale image super-resolution (Liu et al., 2021).

1. Motivation and Conceptual Foundation

IPE extends traditional multi-frequency positional encodings by recognizing that, in discrete images, a pixel represents the integral of the underlying continuous signal over a cell, rather than its value at the center. In arbitrary-scale super-resolution, this mismatch leads to artifacts: center-based PE (point-wise encoding) causes high frequencies to alias in small pixels and fails to capture fine detail in large-scale outputs. IPE remedies this by encoding each coordinate not as a single location, but as the expected value of the frequency features over the pixel's spatial footprint (Liu et al., 2021).

2. Mathematical Definition

Let $c \in \mathbb{R}^2$ denote the 2D center coordinate of a pixel and $r=(r_W, r_H)$ denote its half-widths along the $x$ and $y$ axes. Traditional positional encoding for a point $x$ with $L$ frequency bands is: $$\gamma(\mathbf{x}) = [\sin(2^0 \mathbf{x}), \cos(2^0 \mathbf{x}), ..., \sin(2^{L-1} \mathbf{x}), \cos(2^{L-1} \mathbf{x})]$$ IPE replaces this with the area-integrated encoding: $$\hat{\gamma}(c, r) = \mathbb{E}_{\mathbf{x} \in \text{pixel}(c, r)} [\gamma(\mathbf{x})]$$ For a sine component with angular frequency $\omega$, the integral over the region yields: $$\mathbb{E}_{(x, y)}[\sin{(\omega x)}] = \sin(\omega c_x)\, \mathrm{sinc}(\omega r_W)$$ where $r=(r_W, r_H)$0. Analogous results hold for the cosine terms. The final IPE vector for a 2D pixel, stacking frequency bands for both $r=(r_W, r_H)$1 and $r=(r_W, r_H)$2, is: $r=(r_W, r_H)$3 where each term is a 2D vector when $r=(r_W, r_H)$4.

3. Distinction from Traditional Positional Encoding

The central distinction between PE and IPE is summarized as follows:

Aspect	Traditional PE	Integrated PE (IPE)
Sampling	At pixel or query center	Integrated over pixel area
High-frequency	Prone to aliasing at small scales	Suppressed/anti-aliased by sinc modulation
High-frequency limit	Always retained, regardless of pixel size	Recovered only when upscaling, i.e., for small cells
Location dependence	Ignores scale/area	Jointly encodes center and spatial extent

In all cases, IPE ensures that low-frequency content is preserved, while high-frequency bands whose wavelength is finer than pixel size are suppressed, effectively anti-aliasing the encoding. For large-scale super-resolution (small $r=(r_W, r_H)$5), the full high-frequency content is restored as $r=(r_W, r_H)$6 (Liu et al., 2021).

4. Integration in Neural Architectures

IPE has been primarily applied within the Local Implicit Image Function (LIIF) super-resolution architecture. In original LIIF, the decoder MLP is fed with a concatenation of the local feature vector and offset from the feature map center (optionally with pixel size). IPE-LIIF replaces this input with the local feature and its IPE-encoded offset and area: $r=(r_W, r_H)$7 where $r=(r_W, r_H)$8 is the nearest local feature vector, $r=(r_W, r_H)$9 is the relative coordinate, and $x$0 is the upsampling factor. The IPE encoding is computed for each query and combined with bilinear interpolation over four nearest feature positions. This respects the area each high-resolution query corresponds to in the continuous domain (Liu et al., 2021).

5. Empirical Evaluation

Experiments demonstrate that IPE-LIIF yields consistent quantitative improvements over center-based LIIF on multiple super-resolution benchmarks:

• On DIV2K validation (EDSR backbone):
  • Scale $x$1: LIIF = 28.94 dB / 0.8962; IPE-LIIF = 29.04 dB / 0.8979 (PSNR/SSIM)
  • Scale $x$2: LIIF = 26.75 dB / 0.8477; IPE-LIIF = 26.79 dB / 0.8483
• With RDN backbone, gains are more pronounced at very high upsampling factors (e.g., $x$3, $x$4).
• On Set5, Set14, B100, Urban100, PSNR gains of 0.03–0.1 dB at large scales are typical; SSIM also improves.

Qualitatively, IPE-LIIF eliminates checkerboard and aliasing artifacts that manifest with center-based encodings, especially on fine, repeating textures or stripes (Liu et al., 2021).

6. Generalization and Applicability

IPE can be substituted into any coordinate-based implicit neural representation where each query represents a region rather than an exact point. Demonstrated applications beyond LIIF include MetaSR and super-resolution models supporting continuous scale factors. Consistent gains over all upsampling ratios $x$5, both in-distribution and out-of-distribution, were observed when the encoding scheme is adopted as a drop-in replacement (Liu et al., 2021).

Potential extensions include:

• Neural radiance fields (anti-aliased rendering)
• Implicit 3D shape functions (sampling voxel averages)
• Video super-resolution (spatio-temporal pixel aggregation)
• Finite-element PDE solvers (basis functions with spatial support)

IPE is particularly relevant where the query’s physical interpretation is inherently extended or aggregate (e.g., volumetric, areal, or temporal) rather than pointwise.

7. Limitations and Future Directions

IPE-LIIF remains a regression architecture; it does not synthesize new plausible high-frequency textures at extreme upsampling rates, unlike generative adversarial frameworks. Additional training time (10–20% overhead) arises from the evaluation of multi-frequency sinc terms. Future directions include combining IPE with generative decoders, adaptive bandwidth selection, or learned anisotropic pixel footprints for non-rectangular sampling kernels (Liu et al., 2021).

A plausible implication is that as coordinate-based neural fields become increasingly prevalent for tasks involving real-valued or composite queries, area-aware encodings like IPE may serve as a general architectural primitive for mitigating aliasing and improving spectral fidelity without reliance on purely point-based representations.