A test-time optimization approach for multi-image super-resolution (MISR) that leverages coordinate-based neural networks (implicit neural representations). SuperF reconstructs high-resolution imagery from multiple shifted low-resolution frames without requiring high-resolution training data.

# Clone the repository
git clone https://github.com/sjyhne/superf.git
cd superf

# Install dependencies (recommended: use a virtual environment)
pip install -e .

# Or install with development tools
pip install -e ".[dev]"

# Basic training with default settings (4x upsampling)
python optimize.py --dataset satburst_synth --sample_id sample_1 --df 4 --iters 1000

# With uncertainty estimation
python optimize.py --dataset satburst_synth --sample_id sample_1 --df 4 --iters 2000 \
    --use_gnll --variance_reg 0.001

# Alternative model (NIR)
python optimize.py --model nir --input_projection fourier_10 --df 4 --iters 2000

A single fully specified run using every argument (synthetic data; ensure data/sample_1 exists, e.g. via create_data_from_single_image.py):

python optimize.py \
  --dataset satburst_synth \
  --sample_id sample_1 \
  --df 4 \
  --scale_factor 4 \
  --lr_shift 1.0 \
  --num_samples 16 \
  --aug none \
  --model mlp \
  --network_depth 4 \
  --network_hidden_dim 256 \
  --projection_dim 256 \
  --input_projection fourier_10 \
  --fourier_scale 10.0 \
  --use_gnll \
  --use_separate_ud \
  --variance_reg 0.001 \
  --variance_smooth_reg 0.0001 \
  --visualize_variance \
  --iters 2000 \
  --learning_rate 2e-3 \
  --weight_decay 0.05 \
  --seed 6 \
  --device 0

Optional flags you can add when needed: --no_base_frame, --no_direct_param_T, --use_color_shift, --no_variance_viz. For multi-sample runs add --multi_sample and --output_folder <dir>.

# Basic WorldStrat training
python optimize.py --dataset worldstrat --root_worldstrat ~/data/worldstrat_kaggle/ \
    --area_name "Landcover-1295513" --df 4 --num_samples 8 --iters 10000

# With affine transformation (translation + rotation)
python optimize.py --dataset worldstrat --root_worldstrat ~/data/worldstrat_kaggle/ \
    --area_name "UNHCR-SYRs008164" --df 4 --num_samples 8 --iters 10000 --rotation True

If your data is in a FORCE datacube (Landsat/Sentinel-2 Level 2 or 3, organised by tiles like X0134_Y0097), you can extract RGB images from one tile and then run super-resolution on that folder.

Step 1 — Extract
extract_force_datacube.py finds reflectance GeoTIFFs (BOA or TOA) in the given tile, reads the first three bands as RGB, scales to 0–255, and writes PNGs plus a short manifest. Each date/sensor becomes one image; multiple images in the folder act as the multi-frame input for SuperF.

# Required: path to datacube root and tile ID
python extract_force_datacube.py --datacube /path/to/level2 --tile X0134_Y0097 --output_dir force_extract

# Optional: limit to a date range and cap the number of images
python extract_force_datacube.py --datacube /path/to/level2 --tile X0134_Y0097 --output_dir force_extract \
    --date_from 20200101 --date_to 20201231 --max_files 16

# Optional: only certain sensors (e.g. Sentinel-2A and Landsat 8)
python extract_force_datacube.py --datacube /path/to/level2 --tile X0134_Y0097 --output_dir force_extract \
    --sensors SEN2A LND08 --max_files 12

Step 2 — Super-resolution
Point optimize_against_folder.py at the extracted folder. It will use all PNGs in that folder as LR frames and output a super-resolved result.

python optimize_against_folder.py --input_folder force_extract --output_folder sr_results --df 2

Useful options for extract_force_datacube.py:

Run python extract_force_datacube.py --help for the full list.

Create synthetic multi-frame data from a single image:

python create_data_from_single_image.py --input_image path/to/image.png --output_dir data/sample_1

SuperF learns sub-pixel transformations from multiple low-resolution observations to reconstruct a high-resolution image:

1. Input: Multiple LR frames with unknown sub-pixel shifts
2. Optimization: Jointly optimize an INR and per-frame affine transformations
3. Uncertainty: Learn pixel-wise confidence to handle clouds, shadows, and moving objects
4. Output: Super-resolved image at the target resolution

For synthetic experiments, training data is generated as follows:

1. Start with an HR image (e.g., 256x256)
2. Upsample 2x using bilinear interpolation (512x512)
3. Apply 16 random sub-pixel translations (-6 to +6 pixels)
4. Create HR-LR pairs by downsampling each translated version
5. Track all transformation parameters for evaluation

After training, results are organized as:

results/
└── {dataset}/
    └── df{factor}_shift{shift}_samples{n}/
        └── {sample_id}/
            └── {model}_{projection}_{iters}/
                ├── comparison.png
                ├── final_training_curves.png
                ├── final_translation_vis.png
                ├── metrics.json
                ├── output_prediction.png
                ├── hr_ground_truth.png
                └── lr_input.png

The project includes visualization utilities in viz_utils.py:

from viz_utils import plot_training_curves, visualize_translations, create_model_comparison_grid

# Plot training curves
plot_training_curves(history, save_path='training_curves.png')

# Visualize predicted vs ground truth translations
visualize_translations(pred_dx, pred_dy, target_dx, target_dy, save_path='translations.png')

# Compare multiple models
create_model_comparison_grid(base_dir='results', save_dir='comparison')

If you find this work useful, please cite:

@article{jyhne2025superf,
  title={SuperF: Neural Implicit Fields for Multi-Image Super-Resolution},
  author={Jyhne, Sander Riisøen and Igel, Christian and Goodwin, Morten and
          Andersen, Per-Arne and Belongie, Serge and Lang, Nico},
  journal={arXiv preprint arXiv:2512.09115},
  year={2025}
}

• Sander Riisøen Jyhne - University of Agder
• Christian Igel - University of Copenhagen
• Morten Goodwin - University of Agder
• Per-Arne Andersen - University of Agder
• Serge Belongie - University of Copenhagen
• Nico Lang - University of Copenhagen

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

This project is licensed under the MIT License - see the LICENSE file for details.