- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes RaFL, a random filter loss for image reconstruction. It combines convolutional, Fourier, and wavelet features with adaptive loss weighting, and studies filter selection, resampling frequency, and initialization.
Critical strengths
The loss requires no pretrained features or additional inference computation. Evaluation across natural, remote-sensing, and medical images, with eight random seeds per configuration, provides useful evidence of transferability and training stability.
Critical weaknesses
- The comparison with pretrained perceptual losses is missing. Although avoiding domain-specific pretrained features is a central motivation, the VGG19 baseline uses random weights.
- The design choices are mainly evaluated on SIDD using DnCNN. It remains unclear whether the same filter combination and initialization are preferable for other tasks and network architectures.
- The improvement in perceptual quality is relatively limited. On SIDD, RaFL performs slightly worse than the pixel loss on both LPIPS and DISTS, and using only convolution and wavelet filters gives better results than the full combination. These results do not sufficiently support the motivation of improving texture preservation.
- The actual training cost is substantially higher despite the low FLOPs. On SIDD, RaFL takes 17.8 ms per iteration, compared with 3.3 ms for the pixel loss and 5.4 ms for untrained VGG19. Even with Adam alone, it takes 12.1 ms.
Questions for authors
See Weaknesses.
4 comments
Response to Reviewer vsvm (2/4): CW2 transfer to other tasks and architecture
CW2. Do the filter-family and initialization choices transfer to other tasks and architectures?
Yes, they transfer as a PSNR- and stability-oriented default. We have repeated the filter-family and initialization ablations on another task and architecture, and tested the default on two more architectures.
(1) WorldView-2 with SFINet (App. A.7, Table S10, which also lists SSIM, ERGAS and SAM):
| Configuration | SIDD, DnCNN: PSNR ↑ (Table 1) | WorldView-2, SFINet: PSNR ↑ |
|---|---|---|
| Default: all three families, Gaussian (RaFL) | 40.28 ± 0.13 | 39.62 ± 0.08 |
| Filter families (Gaussian initialization) | ||
| Fourier + Wavelet | 39.28 ± 1.36 | 39.64 ± 0.15 |
| Fourier only | 39.92 ± 0.32 | 39.47 ± 0.13 |
| Convolution + Fourier | 39.98 ± 0.33 | 39.41 ± 0.13 |
| Convolution + Wavelet | 39.88 ± 0.75 | 39.28 ± 0.13 |
| Convolution only | 39.75 ± 0.17 | 38.20 ± 0.20 |
| Wavelet only | 32.26 ± 3.13 | 34.74 ± 4.15 |
| Initialization (all three families) | ||
| Kaiming normal | 39.92 ± 0.60 | 39.57 ± 0.12 |
| Xavier uniform | 39.98 ± 0.41 | 39.60 ± 0.18 |
| Sparse | 39.91 ± 0.31 | 39.59 ± 0.12 |
On WorldView-2 the default ties with Fourier + Wavelet for the highest PSNR, outperforms the other five subsets, and has the lowest standard deviation across seed (0.08 dB), as on SIDD. The four initializations lie within 0.05 dB of each other, and Gaussian initialization again has the highest mean PSNR and the lowest standard deviation across seed.
(2) Other backbones on SIDD (App. A.7, Table S11):
| Backbone (SIDD) | Loss | PSNR ↑ | SSIM ↑ | LPIPS ↓ | DISTS ↓ |
|---|---|---|---|---|---|
| DnCNN (Table 2) | pixel | 39.64 ± 1.00 | 0.9551 ± 0.0011 | 0.2478 ± 0.0056 | 0.1830 ± 0.0057 |
| RaFL | 40.28 ± 0.13 | 0.9557 ± 0.0012 | 0.2488 ± 0.0036 | 0.1847 ± 0.0034 | |
| U-Net (new) | pixel | 35.21 ± 0.65 | 0.9079 ± 0.0022 | 0.3406 ± 0.0075 | 0.2414 ± 0.0031 |
| RaFL | 35.45 ± 0.89 | 0.9208 ± 0.0074 | 0.3090 ± 0.0071 | 0.2253 ± 0.0025 | |
| DRUNet (new) | pixel | 42.02 ± 0.03 | 0.9669 ± 0.0001 | 0.2215 ± 0.0009 | 0.1873 ± 0.0007 |
| RaFL | 42.06 ± 0.05 | 0.9669 ± 0.0002 | 0.2197 ± 0.0007 | 0.1863 ± 0.0007 |
On both backbones the default matches pixel in PSNR and achieves lower LPIPS and DISTS, clearly with U-Net and slightly with DRUNet.
Accordingly, we have revised the manuscript to present the SIDD-selected configuration as a practical PSNR- and stability-oriented default, added both tables to App. A.7, and updated Sec. 5 to state that the design study now covers two datasets (SIDD and WorldView-2).
Response to Reviewer vsvm (3/4): CW3 perceptual gains
CW3. Perceptual gains and the texture-preservation motivation.
We thank the reviewer for this careful observation. We agree that SIDD with DnCNN does not show a perceptual advantage for RaFL. There, RaFL and the pixel loss do not differ significantly on LPIPS or DISTS. On this setting, RaFL's benefit is instead in PSNR and stability: +0.64 dB on average, with a much smaller standard deviation (0.13 vs. 1.00 dB). Similarly, adding the Fourier family to Convolution + Wavelet changes LPIPS and DISTS little but adds 0.40 dB PSNR and cuts the standard deviation from 0.75 to 0.13 dB (Table 1).
The perceptual gains are clear on other backbones and tasks:
- U-Net and DRUNet on SIDD (CW2): lower LPIPS and DISTS than the pixel loss, at matched PSNR.
- LOL (Table 2): the best LPIPS (0.2568 vs. 0.3055 for pixel ) and DISTS (0.1995 vs. 0.2273) among methods without pretrained features, and SSIM 0.8035 vs. 0.7799.
- WorldView-2 (Table 2): better than all baselines on all four metrics.
- OASIS3 (Table 3): better than the pixel loss and uPL in PSNR, SSIM and LPIPS.
We agree that the original framing overstated texture preservation as the main benefit on SIDD. In the revised Abstract and Introduction, we now frame the selected configuration around PSNR and cross-seed stability, and present the perceptual gains as task- and backbone-dependent.
Response to Reviewer vsvm (4/4): CW4 training cost
CW4. Training cost.
We thank the reviewer for highlighting the gap between theoretical FLOPs and wall-clock training time. We have revised the Abstract to state "requiring 182× fewer loss FLOPs than a VGG19 perceptual loss", so that it refers to loss FLOPs rather than training cost.
Profiling our submitted code showed that the 12.1–17.8 ms step time on SIDD was caused by CUDA kernel-launch overhead from sequential small-filter evaluations rather than arithmetic volume. Keeping the backbone and eager execution unchanged, we refactored the loss by batching the filter evaluations and reusing intermediate output gradients. This mathematically equivalent implementation cuts the step time by more than half:
| Loss on SIDD (DnCNN) | Loss GFLOPs | Submitted ms/it (former Table 4) | Refactored ms/it (Table S12) | Ratio vs. pixel |
|---|---|---|---|---|
| RaFL (MAdam) | 1.12 | 17.8 ms | 8.3 ms (2.1× faster) | 2.52× (was 5.4×) |
| RaFL (Adam only) | 1.12 | 12.1 ms | 4.6 ms (2.6× faster) | 1.39× (was 3.7×) |
| pixel | 0.00 | 3.3 ms | 3.3 ms | 1.00× |
| uPL, SimpleNet | 10.12 | 3.6 ms | 3.6 ms | 1.09× |
| CDC | 5.00 | 4.2 ms | 4.2 ms | 1.27× |
| uPL, VGG19 | 203.80 | 5.4 ms | 5.4 ms | 1.64× |
After refactoring, RaFL with Adam only (4.6 ms) is faster than the untrained VGG19 loss (5.4 ms), and RaFL with MAdam (8.3 ms) takes 1.5× as long, down from 3.3×. RaFL also reaches 1.10 dB higher PSNR than the untrained VGG19 loss on SIDD (40.28 vs. 39.18 dB, Table 2) with 182× fewer loss FLOPs.
On OASIS3 (SwinUNETR), RaFL with Adam only is within 6% of the pixel loss, and most of the remaining overhead comes from the MAdam preconditioner:
| Loss on OASIS3 (SwinUNETR) | Loss GFLOPs | Submitted ms/it (former Table 4) | Refactored ms/it (Table S12) | Ratio vs. pixel |
|---|---|---|---|---|
| RaFL (MAdam) | 11.8 | 296.0 ms | 265.2 ms | 2.07× (was 2.3×) |
| RaFL (Adam only) | 11.8 | 166.6 ms | 135.6 ms | 1.06× (was 1.3×) |
| pixel | 0.00 | 128.4 ms | 128.1 ms | 1.00× |
| uPL, VGG19 | 1821.0 | 138.4 ms | 138.4 ms | 1.08× |
We have updated Sec. 4.4, Sec. 5, App. C.3, and Table S12 (formerly Table 4), and we will include the refactored loss in the code release.
We hope our responses have addressed your concerns, and we would appreciate it if you would take them into account in your final evaluation. We are happy to clarify any remaining points.
Response to Reviewer vsvm (1/4): CW1 pretrained perceptual loss
Response to Reviewer vsvm (R3)
We thank Reviewer vsvm for the careful review and for recognizing that RaFL is simple, transferable, and stable. Please find the following answers to your concerns:
CW1. Comparison with a pretrained perceptual loss.
We have added a pretrained perceptual loss baseline. We trained pixel ℓ1 + pretrained VGG19 (Lℓ1+LVGG19 with ImageNet-pretrained VGG19 features) on SIDD and LOL, the only tasks with 3-channel RGB images where such features are directly applicable, using the same eight seeds (42–49).
Compared with pixel ℓ1 + pretrained VGG19, RaFL achieves higher PSNR on both tasks (+1.00 dB on SIDD, +0.43 dB on LOL). Adding the pretrained VGG19 term lowers PSNR below pixel ℓ1 alone (-0.36 dB on SIDD, -0.42 dB on LOL). The pretrained term does achieve lower LPIPS and DISTS on both tasks and higher SSIM on LOL, which is expected since LPIPS and DISTS are themselves computed from ImageNet-pretrained features. In contrast, RaFL requires no pretrained weights, and uses 182× fewer FLOPs in the loss and applies directly to multispectral and 3D data, whereas pretrained VGG19 features are limited to 2D RGB images. We have added this baseline to Table 2 and discussed these results in Sec. 4.3.