SuperF

Abstract

Methodology

We introduce SuperF, a test-time optimization approach that leverages Implicit Neural Representations (INRs) to achieve super-resolution without the need for high-resolution training data. SuperF shares a single INR across multiple low-resolution frames, jointly optimizing the neural network alongside the sub-pixel alignment of each frame. By parameterizing these alignments as affine transformations and optimizing on a super-sampled coordinate grid, our method effectively reconstructs the underlying high-resolution signal. Furthermore, SuperF can estimate uncertainty maps to robustly handle noise, such as ignoring pixels obscured by clouds in satellite imagery.

Figure Animation

Figure 1: Overview of the SuperF method, showing the shared implicit neural representation, frame-specific alignments, and super-resolved reconstruction.

Animation of the SuperF method, illustrating joint alignment, the shared neural field, and super-resolved output.

Robustness: Handling Real-World Noise

Real-world imagery is rarely perfect. Satellite observations are frequently obstructed by clouds or shadows, while handheld bursts often contain moving objects or lighting shifts. Standard super-resolution methods can be easily confused by these inconsistencies, resulting in glitchy artifacts in the final image. SuperF does not just reconstruct the image; it also learns to estimate its own confidence via pixel-wise uncertainty maps that identify which parts of an image are reliable and which are noisy. During optimization, the model assigns high uncertainty to outliers—such as a cloud that appears in only one satellite frame—and effectively ignores or strongly down-weights these corrupted pixels. This allows SuperF to extract the consistent, high-quality underlying structure of the scene, even when individual input frames are messy or partially occluded.

‹ ›

Four examples showing LR frames (top row) and the corresponding estimated uncertainty maps (bottom row) for images with occlusions and inconsistencies. The uncertainty maps highlight cloudy or inconsistent pixels, enabling the GNLL loss to downweight these during optimization.

Results

SuperF consistently outperforms test-time optimization baselines across multiple datasets, delivering sharper detail and lower perceptual error.

Quantitative Results

Real Satellite Data

Results on real satellite imagery from Sentinel-2 time series. These reconstructions were generated using the SuperF demo application.

‹

LR Sequence

LR Sequence

LR Sequence

LR Sequence

LR Sequence

›

‹

›

Qualitative Comparison

Results on synthetic data for both satellite images and ground-level bursts.

‹ ›

Upsampling Factor: ×2

Qualitative comparison of SuperF reconstructions against baseline methods.

Demo App

Try SuperF in your browser: choose a location and time period, and inspect the super-resolved render directly.

Citation

@misc{jyhne2025superfneuralimplicitfields,
                title={SuperF: Neural Implicit Fields for Multi-Image Super-Resolution}, 
                author={Sander Riisøen Jyhne and Christian Igel and Morten Goodwin and Per-Arne Andersen and Serge Belongie and Nico Lang},
                year={2025},
                eprint={2512.09115},
                archivePrefix={arXiv},
                primaryClass={cs.CV},
                url={https://arxiv.org/abs/2512.09115}, 
          }

Acknowledgments

This work was supported in part by the Pioneer Centre for AI, DNRF grant number P1 and by the Global Wetland Center (grant number NNF23OC0081089) from Novo Nordisk Foundation.