Structure-Constrained Super-Resolution Using an Information-Guided Signal

A ×4 luminance super-resolution network conditioned on an explicit local structural map, computed deterministically from the low-resolution input. It is evaluated on the public DIV2K validation set (images 0801–0900) after training on the disjoint DIV2K training set. On structural ROIs, the model improves on classical interpolation (median, paired Wilcoxon test, n = 100, p < 0.001):- **over Lanczos:** +0.559 dB PSNR-Y;- **over bicubic:** +0.799 dB. The paper reports a clean negative result: **the gain cannot be attributed to the structural channel.**- With the channel set to zero at an identical training budget, the model does slightly better.- Other window sizes never make the channel beat its absence.- An earlier three-seed ablation reached the same verdict. The published model was also under-trained: longer training gains +0.304 dB, the largest effect in the study. The method is not competitive in quality with deep baselines. What remains is a compact, deterministic and publicly verifiable pipeline. **Version notes (rev3)** - The ablation showing that the structural channel brings no gain is kept and placed at the centre of the abstract.- Statements contradicted by companion papers 17 and 20 were corrected: the measure is not "beyond entropy", and the robustness figures are attributed correctly.- The "no regression" claim was corrected: the model falls 0.637 dB below Lanczos outside the favourable degradation regime.- Results are reported on a public corpus.- A duplicated reference was removed.