A computer-vision method called WAVE had lower reported RMSE—the paper’s error score—than SPFNet at the 32× setting on two benchmark datasets. On RGB-D-D, the scores were 3.77 for WAVE and 3.97 for SPFNet; on NYU_v2, they were 7.90 and 8.06. In both comparisons, the lower RMSE was better.
The finding comes from an arXiv preprint, version 1, dated 26 August 2026. The paper asks whether reversing the usual fine-to-coarse guidance hierarchy could support coarse-to-fine guided depth super-resolution and reduce misleading effects from RGB guidance.
Rebuilding depth from coarse structure
WAVE uses ML-DWT, a multilevel wavelet transform, and consumes the RGB-guide sub-bands in reverse generation order. It starts with the coarsest structure and proceeds towards finer detail.
The method separates structure and detail processing, applies semantic gating to high-frequency bands and uses invertible coupling for multimodal fusion. It keeps DINOv3 frozen and uses a lightweight, task-specific token-adaptation module instead of fine-tuning the backbone or adding an auxiliary semantic loss.
The evaluation used HYPERSIM and NYU_v2 training protocols, with RMSE reported at 8×, 16× and 32× across RGBD-D, TOFDSR, NYU_v2, Middlebury, Lu and DIML. The optimization used Adam with an initial learning rate of 1e-4.
The scale changed the picture
Under HYPERSIM training, WAVE’s reported RMSE at 8× and 16× was 2.13 and 3.17 on RGBD-D; 3.49 and 5.33 on TOFDSR; 4.61 and 7.16 on NYU_v2; 2.28 and 3.52 on Middlebury; and 2.35 and 4.86 on Lu. The averages were 2.97 and 4.81.
Under NYU_v2 training, the 8× and 16× values, in the paper’s listed order, were 1.58 and 2.39 on RGBD-D; 1.62 and 2.50 on DIML; 2.50 and 4.60 on NYU_v2; 1.56 and 2.82 on Middlebury; and 1.40 and 3.20 on Lu.
The authors interpret the largest gains as occurring at higher upsampling factors. They acknowledge smaller gains in the saturated in-domain low-scale regime and dependence on a frozen foundation model.
At 16×, the paper reports sharper object boundaries, cleaner surfaces, fewer texture-copying artifacts and less boundary bleeding for WAVE than for competing methods. These findings came from qualitative comparisons, without a formal observer or perceptual analysis.
Tests of the moving parts
The ablation test used WAVE trained under the NYU_v2 protocol at 16× and evaluated it on 560 TOFDSR samples unseen during training and treated as out-of-distribution. Full WAVE recorded an RMSE of 4.57 cm, compared with 4.79, 4.63, 4.63, 4.65, 4.69, 4.60, 4.62 and 4.71 cm for WAVE-a through WAVE-h. Every listed ablation had a higher RMSE than the full method.
In a separate robustness test, 16× models were evaluated after the RGB guide was translated by 1 to 8 pixels. WAVE remained the lowest-scoring method or tied with the best RMSE across all four test sets, and it did not show the sharper large-shift degradation described for C2PD.
Replacing DINOv3 with SAM-2 or ResNet-50 produced comparable performance, with differences by dataset. DINOv3 was best on RGB-D-D, Middlebury and Lu, while SAM-2 was marginally better on DIML and NYU_v2.
A benchmark result with defined limits
At 448 × 448 single-batch inference, WAVE used 39.69 M trainable parameters and 21.60 M non-trainable parameters, with 1736.76 G FLOPs and 1.09 GB of peak GPU memory. The same figures were reported at both 8× and 16× under that measurement setup.
The study is an arXiv preprint, and the supplied analysis reports no confidence intervals or inferential uncertainty estimates for the benchmark results. The reported margins are comparisons under the stated protocols, rather than estimates accompanied by a measured range of variation.
Paper data and sources
Original title: WAVE: Reversing the Guidance Hierarchy for Coarse-to-Fine Guided Depth Super-Resolution
Authors: Tayyab Nasir, Daochang Liu, Ajmal Mian
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text