Quillwave: Sparse-Sensor Reconstruction of Synthetic Ripple Fields

Ada Placeholder, Ben Example, Cleo Sample

Example University (fictional), Department of Imaginary Sciences

FCSV 2026

Paper Code
Linear interpolation of 64 sensors beside the Quillwave reconstruction of the same synthetic ripple field.

Abstract

Quillwave reconstructs a full 256×256 ripple field from 64 point sensors. Ripple fields are procedurally generated sums of interfering circular waves: cheap to synthesise, but hard to interpolate, because the wave fronts between sensors are lost by smooth interpolation. Quillwave encodes the sensor readings into a 128-dimensional latent field code and decodes it with a coordinate-based field decoder. On RippleBench-Synthetic, a benchmark of 500 synthetic test fields generated for this work, Quillwave reaches 29.46 dB PSNR and 0.887 SSIM, 5.41 dB above radial-basis interpolation (the best baseline), which is a 71.2% lower mean squared error. It is also 1.36× faster than radial-basis interpolation at inference. All results are on synthetic fields only.

Video

Read the transcript (the video has on-screen captions and no speech) · Download the video (MP4)

Method

Pipeline diagram: 64 point sensors feed a sensor encoder, which produces a 128-dimensional latent field code; a field decoder, also given a pixel position, outputs a value and a confidence for the 256 by 256 field.
Figure 1. Quillwave pipeline: the 64 point sensors, the sensor encoder, the 128-dimensional latent field code, the field decoder, and the reconstructed 256×256 field.

The input is 64 point sensors on a regular 8×8 grid. A sensor encoder maps the readings to a 128-dimensional latent field code. A coordinate-based field decoder then predicts the field value and a confidence at every pixel, so the output resolution is not tied to the training resolution (all results here use the benchmark resolution). Quillwave-Lite is the ablation without the latent code: its decoder is conditioned on the raw readings. The full model has 1.2M parameters and was trained on 2000 synthetic fields for 200 epochs.

Results

Quillwave reaches 29.46 dB PSNR and 0.887 SSIM on RippleBench-Synthetic with 64 sensors. That is 5.41 dB above radial-basis interpolation, the best baseline, a 71.2% reduction in mean squared error, and it runs 1.36× faster than that baseline (4.7 ms against 6.4 ms), though slower than linear interpolation (0.8 ms). Quillwave-Lite reaches 27.82 dB, so the latent field code accounts for part of the gain. Quillwave is the best method at every sensor count from 16 to 128.

Table 1. Reconstruction quality on RippleBench-Synthetic (64 sensors, 256×256 fields, mean over 5 training runs). Higher PSNR/SSIM is better; lower runtime is better.

MethodPSNR (dB)SSIMRuntime (ms)
Linear interpolation21.370.6120.8
Radial-basis interpolation24.050.7016.4
Quillwave-Lite (ablation)27.820.8343.1
Quillwave (ours)29.460.8874.7

Table 2. PSNR (dB) on RippleBench-Synthetic as the number of sensors changes (256×256 fields).

Method16 sensors32 sensors64 sensors128 sensors
Linear interpolation16.9219.1021.3723.88
Radial-basis interpolation18.4421.2624.0526.71
Quillwave-Lite (ablation)21.0324.6727.8230.15
Quillwave (ours)22.5826.3129.4631.94
Six panels: PSNR and SSIM against number of sensors, runtime per field, absolute error histogram, per-scene PSNR of Quillwave against radial-basis interpolation, and validation PSNR during training.
Figure 2. Results on RippleBench-Synthetic. Top row: (a) PSNR and (b) SSIM against the number of sensors (16, 32, 64, 128), and (c) runtime per field. Bottom row: (d) histogram of absolute per-pixel error, (e) per-scene PSNR of Quillwave against radial-basis interpolation (points above the diagonal favour Quillwave), and (f) validation PSNR during training.
Four test scenes in rows; columns show the sensor map, linear interpolation, radial-basis interpolation, Quillwave, and ground truth. Quillwave's wave fronts closely match the ground truth, while the interpolations are blurred.
Figure 3. Qualitative comparison on four test scenes. Columns: sensor map, linear interpolation, radial-basis interpolation, Quillwave (ours), ground truth. The wave fronts in the Quillwave column stay sharp.

Comparison Clips

Test scene 2: linear interpolation of the 64 sensors beside the Quillwave reconstruction.
Test scene 3: linear interpolation of the 64 sensors beside the Quillwave reconstruction.

Poster

Open the poster (PDF)

Download poster (PDF, 420 KB)

Limitations

  • Baselines are not compute-matched. Quillwave is trained for 3.5 GPU-hours, while the interpolation baselines need no training at all.
  • Synthetic fields only. Every result is on RippleBench-Synthetic; no real sensor data was used, and behaviour on real measurements is unknown.
  • Confidence is not calibration. The per-pixel confidence map that Quillwave outputs has not been checked against observed error rates, so it should not be read as a calibrated probability.

BibTeX

@inproceedings{placeholder2026quillwave,
  title     = {Quillwave: Sparse-Sensor Reconstruction of Synthetic Ripple Fields},
  author    = {Ada Placeholder and Ben Example and Cleo Sample},
  booktitle = {Proc. FCSV 2026 (Fictional Conference on Synthetic Vision)},
  year      = {2026},
  url       = {https://quillwave.example.org/},
  note      = {Fictional entry for testing}
}