Implicit Neural Representation-Based Continuous Single Image Super Resolution: An Empirical Study

TL;DR

This study empirically benchmarks implicit neural representation (INR)–based continuous single-image super-resolution (ASSR) methods under a unified, reproducible framework. It analyzes how training configurations, objective designs, and scaling laws affect performance across diverse IQA metrics, datasets, and encoders, revealing marginal gains from newer architectures and substantial sensitivity to training choices. A hybrid pixel-gradient loss is proposed to enhance texture fidelity while preserving edges, achieving consistent perceptual improvements. The work validates scaling trends (more parameters, longer training, and greater data diversity yield gains with diminishing returns) and provides a public benchmarking framework to guide future INR-based ASSR research and fair comparisons.

Abstract

Implicit neural representation (INR) has become the standard approach for arbitrary-scale image super-resolution (ASSR). To date, no empirical study has systematically examined the effectiveness of existing methods, nor investigated the effects of different training recipes, such as scaling laws, objective design, and optimization strategies. A rigorous empirical analysis is essential not only for benchmarking performance and revealing true gains but also for establishing the current state of ASSR, identifying saturation limits, and highlighting promising directions. We fill this gap by comparing existing techniques across diverse settings and presenting aggregated performance results on multiple image quality metrics. We contribute a unified framework and code repository to facilitate reproducible comparisons. Furthermore, we investigate the impact of carefully controlled training configurations on perceptual image quality and examine a new loss function that penalizes intensity variations while preserving edges, textures, and finer details during training. We conclude the following key insights that have been previously overlooked: (1) Recent, more complex INR methods provide only marginal improvements over earlier methods. (2) Model performance is strongly correlated to training configurations, a factor overlooked in prior works. (3) The proposed loss enhances texture fidelity across architectures, emphasizing the role of objective design for targeted perceptual gains. (4) Scaling laws apply to INR-based ASSR, confirming predictable gains with increased model complexity and data diversity.

Figures (13)

• Figure 1: Comparison of best results achieved by different models on the DIV2K dataset across different scaling factors. The x-axis corresponds to the scaling factors, while the y-axis shows the image quality assessment (IQA) values for the different techniques. On all the scales, only marginal performance differences are observed across architectural variations.
• Figure 2: Models ranked based on aggregated performance across datasets, scales, and seven IQA metrics using Borda count aggregation, where lower ranks indicate better overall performance. Ranking variations highlight the sensitivity of different techniques to training configurations. The final rank for each method is shown on the far right, and red downward arrows indicate a drop in rank compared to previously reported results jiang2025hiif.
• Figure 3: Vertical axis, in each case, is the change in the IQA metric due to a change in training configuration. a) Replacing the multi-step learning rate scheduler with SGDR improves performance across all IQA metrics for all techniques, with HIIF showing the most significant gain. b) Increasing the training patch size from $48^{2}$ to $64^{2}$ positively impacts performance across all IQA metrics. c) Increasing model complexity by switching the encoder from EDSR (1.5M parameters) to RDN (21.9M parameters) yields improvements consistent with model scaling trends. d) Extending training from 100 to 150 epochs shows performance gains consistent with training cost scaling trends. e) Expanding the training random scale distribution range from 1–4 to 1–6 improves performance, consistent with scaling trends related to data volume and diversity.
• Figure 4: Models ranked per IQA metric based on aggregated performance across datasets, scales, and settings using Borda count aggregation. Variations in ranking highlight the sensitivity of different techniques to specific perceptual metrics. The red downward arrows indicate a decrease in rank relative to previously reported results.
• Figure 5: Relative change in texture-sensitive IQA metrics when replacing the L1 loss with L1-Gram and Hybrid-Gradient losses. Overall, the Hybrid-Gradient loss shows consistent performance improvements across architectures.