Implicit Neural Representation Decoder

• Implicit Neural Representation (INR) decoders are neural architectures that map continuous coordinates directly to signal values, enabling compact and continuous interpolation.
• They employ advanced encoding methods such as RFF-cosine, Fourier features, and wavelet activations to capture high-frequency details and optimize bandwidth usage.
• Empirical benchmarks demonstrate that modular INR designs enhance PSNR and reduce BD-Rate, with applications in image compression, scientific imaging, video processing, and segmentation.

Implicit Neural Representation (INR) Decoders

Implicit neural representation (INR) decoders are neural architectures that map continuous coordinate inputs directly to signal values (e.g., RGB color, density, or semantic class), serving as decoder modules for tasks such as image/video compression, scientific signal modeling, and 3D scene reconstruction. INR decoders differ fundamentally from explicit representations (e.g., pixel grids) by leveraging continuous coordinate-based MLPs or hybrid architectures, enabling continuous signal interpolation, parameter sharing, and compact memory footprints.

1. Core Mathematical Design of INR Decoders

The canonical INR decoder is a multilayer perceptron (MLP) that receives as input a coordinate $v$ (e.g., $(x, y)$ for images or $(x, y, t)$ for video) and outputs the corresponding signal value $f_\theta(v)$. Variations arise in input encoding, network depth/width, activation functions, and how additional latent variables or global context are provided.

A general mathematical formulation is: $f_\theta(v) = \text{MLP}(\gamma(v))$ where $\gamma(\cdot)$ denotes a positional or spectral encoding mapping the coordinate into a high-dimensional embedding. The choice of $\gamma$ and the structure of $f_\theta$ are central to decoder performance and application.

Damodaran et al. (Damodaran et al., 2023) propose a “richer” random Fourier-feature (RFF-cosine) encoding,

$$\phi(v) = [\sqrt{2}\cos(2\pi w_1^T v + b_1), \ldots, \sqrt{2}\cos(2\pi w_{2D}^T v + b_{2D})]^T$$

with $w_i \sim \mathcal{N}(0, \sigma^2 I)$, $(x, y)$0, which doubles the number of independent frequency bases for fixed embedding size, compared to classic sin–cos splits. This leads to lower kernel-approximation error, improved rate-distortion behavior, and superior high-frequency reconstruction, especially when embedding budgets are small.

Typical decoder backbones may use activation functions such as $(x, y)$1 (“SIREN” style), piecewise quadratic (Zhou et al., 20 Aug 2025), Gabor wavelet activations (Roddenberry et al., 2023), or ReLU. The output layer is generally a linear map to the desired signal dimension (e.g., $(x, y)$2 for RGB).

2. Integration of Input Encoding and Frequency Coverage

High-frequency detail and anti-aliasing capacity in INR decoders are largely determined by the positional encoding and first-layer transformations:

• Fourier Feature Mapping: $(x, y)$3 encodes $(x, y)$4 with $(x, y)$5 frequency bases via sinusoids; in low-budget regimes, bandwidth is often underestimated.
• RFF-Cosine Encoding: By mapping each frequency sample to a separate embedding slot with random phase, the representation can double spectral richness without increasing memory (Damodaran et al., 2023). This drastically reduces bandwidth distortion (BD-Rate) and enables finer details at low bitrates.
• Learned or Multi-resolution Encodings: Hybrid approaches with hash-grid encoders, as in NeRF-style models and WSI-INR for medical images (Wu et al., 4 Mar 2026), support arbitrary output resolutions and adapt to local spatial complexity.
• Wavelet Encodings: Gabor and complex wavelet activations enable simultaneous spatial and spectral localization, better edge preservation, and sharp singularity modeling (Roddenberry et al., 2023).

Empirical findings demonstrate that careful input encoding is essential: RFF-cosine embeddings permit small MLPs to reach the fidelity of much larger classic decoders at fixed parameter counts, with BD-Rate reductions up to 98% at embedding size 8 and still 10% at size 64 (Damodaran et al., 2023).

3. Decoder Architecture: Depth, Width, Nonlinearity, and Adaptation

INR decoder depth and activation function selection directly affect signal fitting power and speed–quality tradeoffs:

• Shallow Decoders with Pretrained Encoders: STRAINER (Vyas et al., 2024) splits the network into a shared multi-layer “encoder” and minimal per-signal decoder (typically a single sin-activated layer). Pretraining encoder layers on a dataset permits extreme transfer: at test time, a randomly initialized decoder rapidly specializes, yielding 7–10 dB higher PSNR and up to 3× faster convergence versus training the full network from scratch.
• Activation Function Choices:
  • Sinusoidal (SIREN): Exhibits high-frequency expressivity but high hardware cost (Zhou et al., 20 Aug 2025).
  • Piecewise Quadratic: QuadINR’s quadratic activation packs rich Fourier harmonics, can be implemented in exact two-stage pipelines for hardware efficiency, and outperforms sinusoidal in both PSNR and area/power consumption (up to 97% savings) (Zhou et al., 20 Aug 2025).
  • Complex Wavelets: Decoders with Gabor/morlet wavelet activations and band-pass support can resolve sharp features, enabling split modeling of low and high frequencies (Roddenberry et al., 2023).
  • Affine-Parameterized Activations: INCODE (Kazerouni et al., 2023) employs a harmonizer network to adapt amplitude, frequency, phase, and bias per layer/sample, regulated by learned priors, achieving faster convergence and higher fidelity across modalities.

Table: INR Decoder Backbone Choices

Decoder Variant	Activation	Encoding	Notable Benefit
SIREN	sin	Fourier	High-freq, easily tuned
QuadINR	quadratic	linear	Hardware optimal
WIRE/Gabor	wavelet	affine template	Spatial-freq localization
INCODE	param. sine	harmonizer	Adaptivity, multi-modal

4. Modulation and Conditional/Generalizable Decoding

Emerging INR decoders leverage weight modulation, hypernetworks, and cross-attention to encode data-dependent and generalizable representations:

• Latent Modulation and Hypernetwork Decoding: Latent-INR (Maiya et al., 2024) and HUVR (Gwilliam et al., 20 Jan 2026) use per-sample (or per-frame) latent codes, converted by small hypernetworks to produce weight modulations for a shared base MLP. In video, a sequence of low-dimensional temporal latents, mapped via hypernetwork to low-rank per-layer deltas, enables per-frame adaptation, interpolation, and semantic alignment, with minimal parameter overhead.
• Selective Latent Aggregation: Generalizable INR decoders (Lee et al., 2023) use transformer-encoded tokens and spatial cross-attention to form locality-aware modulations for each input coordinate. Multi-band, coarse-to-fine stacking enables spectral adaptation at each network depth, achieving large PSNR advances (up to +8 dB on ImageNette).
• Unified Vision Embedding: HUVR’s patch-wise decoders are modulated by transformer-compressed tokens; distillation from large-scale recognition models ensures that the learned INR embedding is both generative and discriminatively powerful (Gwilliam et al., 20 Jan 2026).

5. Quantization, Compression, and Hardware Considerations

Efficient INR decoder deployment, especially for compressed representation, requires extreme parameter efficiency, low computation, and hardware alignment:

• Quantization-Aware Optimization: RQAT-INR (Damodaran et al., 2023) adopts range-aware quantization and explicit entropy modeling for parameter bits. Regularization anchors the decoded representation to a fixed-precision solution via a distillation term, achieving 41% bit-rate reduction vs. prior INR codecs, with per-pixel decoding MAC count an order of magnitude lower than VAE-based methods.
• Resource-Optimal Activation Functions: QuadINR (Zhou et al., 20 Aug 2025) achieves 97% reductions in LUT, DSP, and power vs. trigonometric or Gaussian AFs, while improving accuracy. Pipeline architectures for activations facilitate systolic-style hardware mapping.
• Consistent Entropy Models: Video INR codecs with conditional decoders (e.g., SNeRV Boost (Zhang et al., 2024)) use a network-free Gaussian model for entropy minimization, replaced with a mixture of L1, MS-SSIM, and frequency-domain losses to jointly preserve edges and compressibility. Convergence is up to 8× faster, and performance matches or beats leading hand-engineered codecs at a fraction of resource cost.

6. Application-Specific INR Decoder Instantiations

Advanced applications motivate bespoke decoder designs:

• Scientific Imaging: QSMnet-INR (Cai et al., 10 Dec 2025) integrates a SIREN-based decoder to complete ill-posed kernel regions in k-space, guided by a physics-informed loss. The INR module provides explicit strong priors on spatial smoothness, greatly suppressing artifacts (+32% reduction in HFEN), and is critical for structural stability in single-orientation QSM.
• Scalable High-Dimensional Modeling: F-INR (Vemuri et al., 27 Mar 2025) decomposes high-dimensional signals via CP/TT/Tucker tensor formats, each axis modeled by a separate subnetwork. This axis-wise design achieves 10–100× speedups (video) while improving PSNR, supporting plug-and-play with any INR “backend” (e.g., SIREN, WIRE, or hash-grid). Ranks tune fidelity/cost trade-off.
• Pathology and Segmentation: WSI-INR (Wu et al., 4 Mar 2026) fuses multi-resolution hash-grid coordinate encodings with a dual-branch CNN+MLP decoder, enabling continuous, resolution-robust segmentation directly on gigapixel images; segmentation Dice is preserved even after 4× downsampling, where classical patchwise UNet pipelines fail catastrophically.

7. Empirical Benchmarks and Practical Guidelines

INR decoder advances are reflected in both quantitative metrics and operational recipes:

• Rate–Distortion and Fidelity: RFF-cosine positional encoding (Damodaran et al., 2023) yields up to 2 dB PSNR boosts and 98% BD-Rate cuts at low mapping sizes.
• Transferability and Adaptation: STRAINER’s split encoder/decoder achieves +10 dB PSNR at onset and up to 13 dB higher in-domain PSNR at convergence than vanilla SIREN (Vyas et al., 2024).
• Modular Guidance: Practical recommendations include maximizing shared encoder depth for transfer, selecting activation function by frequency requirements and hardware, and controlling tensor rank/backbone choice in F-INR for scaling to desired problem size and fidelity (Vemuri et al., 27 Mar 2025).

The table below summarizes key factors in INR decoder deployment:

Application	Encoding	Modulation	Notable Gains
Image Compression	RFF-cosine	None	+2 dB PSNR, 98% BD-Rate cut
General INR Transfer	SIREN	Shallow decoder	+10 dB early PSNR, 3× speed
Scientific Imaging	SIREN	Physics-informed	+32% lower HFEN, +0.07 SSIM
Video	Hash/MLP	Hypernet/Cond.	+1.9 dB, 8× faster conv.
Pathology Segment.	Hash-grid	Dual-branch, CNN	+26% Dice at low resolution

A plausible implication is that, as INR decoders become more modular and modality-aware, they will continue to displace classical decoders in compression, scientific modeling, and semantic-rich tasks, contingent on further advances in encoding, modulation, and efficient hardware integration.