acceptodds
Under review as a conference paper at ICLR 2027

Inverse Problem Solving by Walking the Sphere

Abstract

Latent-space inverse problem solving searches a generator's latent space for a code whose output matches the measurement, making data fit the only objective. With GAN, VAE, and normalizing-flow priors, this recovers signals well, but the set of codes the generator decodes reliably is known only implicitly, so an iterate that drifts away from it can only be pulled back by a penalty that competes with the data fit. Sphere encoders make this set explicit: their latents lie on a hypersphere by construction, and a single normalization returns any iterate to it. We solve inverse problems by projected gradient descent on this sphere, a method we call Walk the Sphere (WTS), and introduce a variant, WTS+Loop, that periodically re-encodes the iterate. Against five diffusion-based solvers tuned under the same budget, WTS or WTS+Loop ranks first, outright or tied, in 13 of 18 settings spanning ImageNet-256 and FFHQ-256 restoration and three scientific problems from InverseBench. The solver needs only three hyperparameters and 300 gradient steps, taking about two seconds per image on FFHQ-256.

Primary area:
generative models
Keywords:
Inverse Problems, Sphere Encoders, Latent-Space Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_v5ZAOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a mechanism for solving inverse problems based on sphere encoders. Although the method is very simple, experimental results on FFHQ, ImageNet super-resolution, and additional scientific imaging tasks demonstrate the effectiveness of the proposed method, WTS.

Critical strengths

Building on sphere encoders, the paper proposes a simple yet effective method for solving inverse problems. While diffusion-model-based inverse problem solvers are currently considered mainstream in the field, WTS revisits latent-based solvers and achieves good quantitative results. One particular strength is its runtime compared with other methods. Although additional experiments are needed to further validate WTS’s efficiency (see [W3, W4]), this non-diffusion-based framework demonstrates the potential of WTS as a strong, efficient solution.

Critical weaknesses

[W1] The paper presents results primarily on the FFHQ dataset and lacks evaluations of other ImageNet tasks beyond super-resolution. Since there are other common tasks besides super-resolution (e.g., box inpainting, phase retrieval, HDR, and deblurring), evaluations on these tasks are needed to validate WTS’s effectiveness. The same applies to the FFHQ experiments, which focus on only three tasks. These evaluations should be included before presenting results on the scientific imaging benchmark to support WTS’s general applicability.

[W2] The baselines primarily consist of representative diffusion-model-based methods. Comparisons with additional methods should be included (e.g., DAPS++ [1] and DDiff [2]). There are also many recent flow-model-based methods, so comparisons with these would be helpful.

[W3] The paper makes a fair comparison using a fixed five-hour budget for hyperparameter search. However, a clearer way to present the results might be to show Pareto curves while varying hyperparameters such as the number of iterations (e.g., as in DAPS). This could also better demonstrate WTS’s strengths.

[W4] Gradient-based methods typically require substantial VRAM for computation. To assess WTS’s actual efficiency, comparisons should include other computational resources, such as peak memory usage, in addition to runtime. The paper does not present this information. Furthermore, the paper does not discuss techniques for alleviating this computational burden (see Q1). I believe these aspects require further study for the work to be complete.

Minor weaknesses that may not affect acceptance

  • Line 99 mentions sequential Monte Carlo methods without a citation. Adding a citation, possibly [3], along with a brief comparison would improve the discussion.
  • Algorithm 1 may be clearer in pseudocode format than as code.
  • Figure 1 presents the workflow intuitively, but its legibility could be improved—for example, by using example images instead of placeholders and reformatting the diagram.
  • In line 191, consider representing noise_scaler as a mathematical symbol for consistency.
  • In Figure 6, WTS+Loop’s PSNR appears to decrease quickly at shorter observation times, making PnP-DM the top-performing method at 3%. Separate visualizations may obscure this comparison. A single line chart, with difficulty on the x-axis and PSNR on the y-axis, could present the results more clearly. The same applies to Figure 5.
  • The Figure 6 caption, “EHT-Imaging is the strongest additional method on this problem,” is somewhat vague. Consider changing it to “EHT-Imaging is the strongest among the additional methods on this problem.”
  • Figure 7 is difficult to follow, and the intensity-based ranking visualization may not be effective. Consider presenting the results in a unified visualization or a table.

[1] Chen, Hao, Renzheng Zhang, and Scott S. Howard. "DAPS++: Rethinking diffusion inverse problems with decoupled posterior annealing." ECCV 2026.

[2] Kim, Minseo, Axel Levy, and Gordon Wetzstein. "Dual ascent diffusion for inverse problems." CVPR 2026.

[3] Dou, Zehao, and Yang Song. "Diffusion posterior sampling for linear inverse problem solving: A filtering perspective." ICLR 2024

Questions for authors

[Q1] Although gradient-based backpropagation is an intuitive solution for solving Eq. (6), it incurs VRAM costs. Did the authors study alternatives for alleviating this issue, such as gradient-free inversion [4]? If WTS is well designed to operate on spherical encoders with numerical stability, gradient-free inversion may be a more effective solution to this problem.

[4] Hong, Seongmin, et al. "Gradient-free decoder inversion in latent diffusion models." NeurIPS 2024

8 comments

We thank the reviewer for the detailed and constructive review. We used the rebuttal period to run the additional experiments you suggested. We added 11 new settings across FFHQ and ImageNet, including box inpainting, phase retrieval, HDR and deblurring. We also added 5 new baselines, including DAPS++, DDiff, FlowDPS, and PnP-Flow. We will revise the paper to include all the new results from our rebuttal experiments.


[W1] More tasks on FFHQ and ImageNet

We added 11 new tasks:

  • FFHQ: nonlinear motion deblur (noise-free), Gaussian deblur (noisy) at σ=6 and σ=12, motion deblur (noisy), HDR, phase retrieval, and box inpainting at two box sizes. These come on top of the paper's SR, Gaussian deblur (noise-free) and random inpainting.
  • ImageNet: Gaussian deblur (noise-free), random inpainting 95%, and HDR, on top of SR.

FFHQ-256: Nonlinear motion deblur (noise-free). We use the nonlinear camera-shake blur of DPS (Chung et al., 2022).

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS30.490.8670.112
WTS+Loop30.470.8680.113
RED-diff27.240.7590.234
DiffPIR27.060.7320.188
DAPS26.830.6810.254
DAPS++26.630.7330.194
SP³26.310.7160.248
PnP-DM26.240.7350.238
DDiff25.600.7100.241

FFHQ-256: Gaussian deblur σ=6 (noisy) Gaussian deblur with the typical 0.05 additive Gaussian measurement noise.

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS+Loop26.210.7290.254
WTS26.020.7180.263
RED-diff25.460.7140.306
PnP-DM25.410.7070.247
DDiff25.290.7190.285
SP³25.220.6890.345
DAPS25.020.6680.275
DAPS++24.510.6160.270
DiffPIR23.310.6000.272
DPS22.890.5630.284

FFHQ-256: Gaussian deblur σ=12 (noisy) Gaussian deblur with the typical 0.05 additive Gaussian measurement noise.

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS23.460.6470.308
WTS+Loop23.410.6500.324
RED-diff22.880.6400.378
SP³22.600.6160.386
PnP-DM22.590.6250.320
DAPS22.440.5910.328
DDiff22.220.6340.368
DAPS++21.720.5830.322
DiffPIR20.560.5040.340
DPS20.300.4910.347

FFHQ-256: Phase retrieval

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS+Loop28.080.7830.203
DAPS26.700.7130.284
DDiff23.720.6990.327
WTS22.540.6180.361
DAPS++21.820.6060.354
DiffPIR18.840.5270.396
DPS14.150.3690.560
PnP-DM12.240.3730.659
RED-diff11.610.3360.749

FFHQ-256: Box inpainting, small boxes (128–148 px)

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS+Loop23.280.8050.143
DAPS22.870.8390.103
RED-diff22.740.8350.133
WTS22.310.7700.181
PnP-DM22.270.7800.180
DPS22.170.7710.165
DiffPIR21.580.7740.155

FFHQ-256: Box inpainting, large boxes (148–181 px)

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS+Loop20.360.7180.215
RED-diff19.820.7780.180
DAPS19.800.7300.187
PnP-DM19.600.7110.240
WTS19.590.7040.234
DiffPIR19.410.6500.285
DPS19.130.7020.208

FFHQ-256: Motion deblur (noisy)

MethodPSNR ↑SSIM ↑LPIPS ↓
RED-diff29.030.8230.186
WTS+Loop28.760.7990.166
DDiff28.610.8110.218
PnP-DM28.020.7780.180
WTS27.720.7720.209
DAPS++27.640.7670.200
DAPS27.350.7430.226
DPS25.930.6730.230
DiffPIR25.390.6500.248

FFHQ-256: HDR

MethodPSNR ↑SSIM ↑LPIPS ↓
RED-diff28.140.8740.129
WTS+Loop27.490.8410.158
WTS27.240.8420.166
DAPS++27.160.8470.143
DAPS27.150.8340.155
DDiff26.620.8660.178
DiffPIR26.460.8120.165
PnP-DM26.030.8010.214
DPS21.780.5950.329

ImageNet-256: Gaussian deblur σ=6 (noise-free)

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS (sphere-L-FT)23.390.6090.362
WTS (sphere-XL-FT)23.250.6020.371
WTS+Loop (sphere-XL-FT)23.120.5950.370
RED-diff22.870.5830.385
WTS+Loop (sphere-L-FT)22.820.5830.399
DiffPIR21.060.5050.417
DAPS20.300.2760.505
DPS19.280.4000.374

ImageNet-256: Random inpainting 95%

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS+Loop (sphere-L-FT)22.460.6410.274
WTS (sphere-L-FT)22.230.6320.287
WTS+Loop (sphere-XL-FT)22.010.6380.273
WTS (sphere-XL-FT)21.640.6090.312
DPS21.180.5570.339
RED-diff20.990.5760.376
DiffPIR20.660.5280.476
DAPS20.320.5250.453

ImageNet-256: HDR

MethodPSNR ↑SSIM ↑LPIPS ↓
WTS (sphere-XL-FT)25.570.8130.199
WTS (sphere-L-FT)25.400.8030.211
DAPS25.080.8230.177
WTS+Loop (sphere-L-FT)24.930.7950.216
WTS+Loop (sphere-XL-FT)24.840.7970.208
RED-diff24.690.8120.193
DiffPIR22.120.7400.226

Across the 11 new tasks we are first on PSNR in 9 and second in the other 2, behind RED-diff both times. We are best on LPIPS in 6. Our largest margins are on nonlinear motion deblur (+3.3 dB over the best baseline), phase retrieval (+1.4 dB) and ImageNet random inpainting 95% (+1.3 dB).

The settings where we do not lead are the most generative settings, HDR and box inpainting. There, large parts of the image are not measured at all (saturated regions, or a box of up to 181×181 pixels), so every method is essentially forced to guess the missing content, and the result depends more heavily on the generative prior. RED-diff leads us on FFHQ HDR (28.14 vs 27.49 dB), and on box inpainting DAPS and RED-diff have better SSIM and LPIPS, although we keep the best PSNR. A stronger sphere autoencoder may close this gap.

Initialization, and phase retrieval. WTS starts from an estimate of the solution computed from the measurement alone, which it encodes onto the sphere. On nearly every task this is a simple pseudo-inverse: the degraded image for deblurring and super-resolution, the masked image for inpainting, and the rescaled measurement for HDR. Phase retrieval is highly non-convex and has no such pseudo-inverse, so we generate several candidate initialisations and let them compete, much like a particle method in order to find a good initialization. The exact steps of the initialization are the following:

  • Classical HIO phase retrieval (Fienup) is run from 32 random starts on the measured magnitudes, and the start with the lowest measurement residual is kept.
  • Its 4 sign/flip ambiguities, each encoded once directly and twice with latent noise, give 12 candidates; the 8 with the lowest measurement loss are each run for 30 steps.
  • The candidate with the lowest measurement loss is then run for the full 300 steps.

Every method gets exactly one attempt per image. Runtime includes this initialization.

MethodPSNR ↑Seconds per image (incl. warm start)
WTS+Loop28.084.97
DAPS26.701.45
DDiff23.722.89
WTS22.544.77
DAPS++21.821.37
DiffPIR18.842.33
DPS14.155.99
PnP-DM12.2435.58
RED-diff11.612.90

Even with this more expensive initialization we are not the slowest method (DPS and PnP-DM take longer), and we obtain the best reconstruction by 1.4 dB.

We will add all of these results to the paper and, following the reviewer's suggestion, present the image results before the scientific benchmarks.

[W2] Additional baselines

All baselines we added were tuned for 5h following our standard protocol. Due to the limited rebuttal time, we ran SP³, some Flow baselines, and all of the methods for noisy Gaussian deblur and nonlinear motion deblur tasks on B300 GPUs. Therefore SP³, and some Flow baselines enjoy a slight advantage, since they likely received more tuning trials than they would have on B200 GPUs.

We added both baselines you named, DAPS++ and DDiff, using the same ADM prior as the other diffusion based solvers. DAPS++ has no public code, so we implemented it ourselves from the paper. We also added FlowDPS, and PnP-Flow, to accommodate your request for Flow Matching baselines. No public pixel-space unconditional FFHQ-256 flow-matching model exists, so we trained one with the same training data, split, architecture and number of training steps (1M) as the ADM diffusion prior.

We also added SP³, concurrent work that also uses a sphere encoder prior. We ran its public code on our FFHQ sphere autoencoder; its relation to our method is discussed in our reply to the public comment regarding SP³.

PSNR (dB) ↑, FFHQ-256

TaskWTS+LoopWTSDDiffDAPS++PnP-FlowFlowDPSSP³
SR ×826.9927.0026.0125.6326.7326.3725.91
SR ×1623.1623.4322.8022.2823.0622.4522.51
Random inpainting 95%25.9225.6524.4822.4925.4124.4423.59
Random inpainting 98%23.9523.2321.5120.7021.2521.1821.18
Gaussian deblur σ=6 (noise-free)27.1427.2025.4224.2025.2923.7930.71
Gaussian deblur σ=12 (noise-free)24.0124.2622.0821.9622.1221.1535.46
Gaussian deblur σ=6 (noisy)26.2126.0225.2924.51––25.22
Gaussian deblur σ=12 (noisy)23.4123.4622.2221.72––22.60
Nonlinear motion deblur (noise-free)30.4730.4925.6026.63––26.31
Motion deblur (noisy)28.7627.7228.6127.64–––
HDR27.4927.2426.6227.16–––
Phase retrieval28.0822.5423.7221.82–––

LPIPS ↓, FFHQ-256

TaskWTS+LoopWTSDDiffDAPS++PnP-FlowFlowDPSSP³
SR ×80.1940.1940.2660.2200.2570.2450.257
SR ×160.2970.2890.3530.3050.3690.3560.372
Random inpainting 95%0.1610.1720.2830.2880.2310.2070.298
Random inpainting 98%0.2080.2370.3310.3410.3700.2920.361
Gaussian deblur σ=6 (noise-free)0.2060.1950.2810.2750.3070.2870.126
Gaussian deblur σ=12 (noise-free)0.3040.3170.3520.3190.4050.3900.068
Gaussian deblur σ=6 (noisy)0.2540.2630.2850.270––0.345
Gaussian deblur σ=12 (noisy)0.3240.3080.3680.322––0.386
Nonlinear motion deblur (noise-free)0.1130.1120.2410.194––0.248
Motion deblur (noisy)0.1660.2090.2180.200–––
HDR0.1580.1660.1780.143–––
Phase retrieval0.2030.3610.3270.354–––

"–": not run.

Why noise-free Gaussian deblurring favours SP³. Without noise, a Gaussian blur can be largely undone from the measurement alone. To show this, we ran SP³'s exact data step on its own, with no prior, tuning its weight on validation. This inversion alone works extremely well for noise-free Gaussian deblurring, and SP³ improves upon it further. Since the measurement alone already recovers most of the image here, we added the noisy version of it, which is common in the literature. By adding Gaussian noise (σ=0.05) to the measurement, the problem is no longer solvable with the data step alone. Both the no prior inversion and SP³ fall below RED-diff and WTS. More broadly, across super-resolution, random inpainting, noisy Gaussian deblurring and nonlinear motion deblurring, SP³ is middle of the pack on PSNR and below most baselines on LPIPS.

Gaussian deblurNo prior (exact inversion)WTSWTS+LoopBest diffusion baseline (RED-diff)SP³
σ=6, noise-free29.8027.2027.1427.0530.71
σ=6, noisy24.0026.0226.2125.4625.22
σ=12, noise-free33.2424.2624.0124.5835.46
σ=12, noisy21.6423.4623.4122.8822.60
Reviewer_wUKgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce Walk the Sphere (WTS), a latent-space inverse problem solver that runs projected gradient descent on the hyperspherical latent space of a sphere autoencoder, where a single normalization keeps every iterate on the decoder's training domain. In addition, an extended version of WTS, WTS+Loop, periodically re-encodes the iterate. The idea is simple yet clever, and across 18 settings on benchmarks at 256×256 resolution, the authors report both reconstruction quality and speed.

Critical strengths

1- The paper identifies a core reason why classical latent-space solvers fell behind diffusion models: their latent manifold is only implicitly defined. To address this, it introduces a geometric formulation with "free" Riemannian gradients, using sphere autoencoders whose valid codes lie on the explicit hypersphere Sdd−1\mathbb{S}_{\sqrt{d}}^{d-1}. Placing the normalization inside the forward pass also means that backpropagation yields the tangent-space gradient without an explicit projection step.

2- The method is fast compared to related work and is schedule-free. A solve takes about 2 s per image on the FFHQ dataset, without needing a noise schedule during inference.

3- It is simple and somewhat principled, in the sense that it is motivated by the dynamics and geometry of the latent space.

Critical weaknesses

1- Lack of generalization to higher resolutions: the authors only provide results at 64×64, 128x128 and 256×256 resolution and on multi-coil MRI; no higher-resolution results are provided. If speed is a concern, then higher resolutions would offer more room for improvement, which is where the method would be most useful.

2- I think the figures, including the methodology figure, and even most of the text are heavily LLM-generated (e.g., line 53, line 185); some sentences read as non-human and overly formal.

3- I think there is also an asymmetry in measuring the efficiency bottleneck, as WTS requires training, while some baselines can already be adapted to the same problem. One way to make the comparison fairer is to give those baselines the same wall-clock time by increasing their number of steps, which would result in a clearer and fairer comparison.

4- There is another source of unfairness in the evaluation: all solvers are tuned to maximize validation PSNR, which naturally favors WTS, since PSNR-only tuning favors point estimators. WTS returns a single deterministic estimate, and PSNR rewards smooth, average-looking reconstructions, whereas diffusion samplers generate sharp but stochastic details that pixel-wise metrics penalize.

Questions for authors

No questions.

6 comments

We thank the reviewer for their time and feedback on our work. We used the rebuttal period to run additional experiments you suggested. Beyond your suggestions, we added many more results, which are in our response to reviewer v5ZA. In brief, we added 11 new settings and 5 new baselines, as well as a variety of analyses. We will revise the paper to include all the new results from our rebuttal experiments.


W1: Higher resolutions

The resolutions in the paper are:

ProblemResolution
Black hole imaging64×64
Inverse scattering128×128
FFHQ and ImageNet256×256
Multi-coil MRI320×320

We evaluate at the above resolutions because they are the standard resolutions for these benchmarks. InverseBench's image tasks and the diffusion-solver literature we compare against all use 256×256 resolution. Further, the released sphere encoders are trained at 256. No publicly released sphere encoder exists at higher resolution, and we do not have the compute to train one, not to mention a matched diffusion prior for a fair comparison. At 256, our method leads on reconstruction quality and is among the fastest methods.


W2: Writing and figures

Our AI-use statement is in the main text at the end of page 9. We are revising the text for clarity and readability, so your specific pointers are helpful feedback. Thank you! We will improve lines 53 and 185, as well as anything else in the draft that can use improvement.

Regarding the figures:

  • Method figure: a first draft was generated with an LLM but it was reworked by a human in Figma.
  • Results figures: these are Matplotlib. The plotting code was mostly written with coding assistants, but every figure was fully specified by the authors and iterated on until we found it clear.

If any figure is confusing, we would gladly take that feedback and update it.


W3: Training asymmetry and equal wall-clock

It sounds like there may be a misunderstanding here. WTS is not retrained per problem. Every method in the paper requires a trained prior, but none of them train a prior specific to a task. We train one sphere autoencoder per dataset and use that same model for every task on that dataset, exactly as the diffusion baselines use one diffusion prior per dataset. On FFHQ, where we trained both priors ourselves on the same data and hardware, the diffusion prior is in fact much more expensive to train.

PriorDataTraining cost (GPU-hours)
Sphere autoencoder (WTS, WTS+Loop)FFHQ-256282
ADM diffusion prior (DPS, DAPS, RED-diff, DiffPIR, PnP-DM, ...)FFHQ-256667

If anything, the training-compute asymmetry benefits the diffusion baselines. Every model we trained, and its cost, is listed in the appendix.

Equal wall-clock / more steps for the baselines. We ran exactly the experiment you suggested. We varied the number of steps for WTS, WTS+Loop, DPS, and RED-diff. For each step count we run a full 5h tuning. DAPS, DiffPIR and PnP-DM already tune their step count as part of their 5h tuning, so we report their best step count found. Below is time per image, and PSNR for each method on two tasks.

Method, stepsTime per imageSR ×8Inpaint 95%
WTS, 500.30 s26.5325.08
WTS, 1000.60 s26.8125.45
WTS, 3001.78 s27.0025.65
WTS, 6003.55 s26.9025.34
WTS+Loop, 500.30–0.35 s26.5225.95
WTS+Loop, 1000.60–0.64 s26.8326.18
WTS+Loop, 3001.78–1.80 s26.9925.92
WTS+Loop, 6003.56–3.60 s26.4625.72
RED-diff, 2500.72 s26.0222.11
RED-diff, 5001.43 s26.0923.32
RED-diff, 10002.90 s26.1823.93
RED-diff, 20005.74 s26.1824.58
DPS, 2501.44 s23.3723.80
DPS, 5002.86 s23.7623.94
DPS, 10005.79 s23.9524.09
DPS, 200011.5 s23.9424.15
DAPS, 151/36 (SR), 200/19 (inpaint)¹0.58 s (SR), 0.66 s (inpaint)25.6721.95
DiffPIR, 200 (SR), 800 (inpaint)0.58 s (SR), 2.31 s (inpaint)24.0823.32
PnP-DM, 174/31 (SR), 106/50 (inpaint)¹10.29 s (SR), 8.80 s (inpaint)25.4819.08

¹ Outer/inner steps, for methods with a nested loop (i.e. DAPS and PnP-DM).

WTS with only 50 steps (0.3 s per image) beats every diffusion baseline at every step count (all the way up to 2000 steps) on both tasks. Our pareto curves dominate those of the baselines. WTS with 50 steps requires 19–38× less time per image than DPS and RED-diff at 2000 steps.

The seed-to-seed standard deviation of PSNR is below 0.05 dB for every method (3 seeds on SR ×8 and inpaint 95%), and rankings do not change. More steps for the baselines does not close the gap. Nor does more tuning. For the slowest-tuning baselines, we extended tuning to 10 h, double our budget, evaluated the new best configuration on the test set, and compared it with our 5 h result. Test PSNR (dB):

BaselineTaskBaseline, 5 h → 10 hOurs, 5 h
PnP-DMFFHQ SR ×8 / ×1625.48 → 25.65 / 22.03 → 22.2627.00 / 23.43
PnP-DMFFHQ Gaussian deblur σ=6 / σ=12 (noise-free)25.55 → 25.72 / 22.71 → 22.9227.20 / 24.26
PnP-DMFFHQ inpaint 95% / 98%19.08 → 20.90 / 13.41 → 14.9825.92 / 23.95
PnP-DMImageNet SR ×4 / ×823.51 → 24.53 / 19.31 → 19.3126.38 / 23.27
PSLD (U-ViT-L)ImageNet SR ×4 / ×823.47 → 24.33 / 20.95 → 21.6326.38 / 23.27
PSLD (U-ViT-H)ImageNet SR ×4 / ×823.20 → 23.65 / 21.14 → 21.1426.38 / 23.27
DPSImageNet SR ×820.96 → 21.0223.27
DiffPIRImageNet SR ×822.39 → 22.4123.27
LatentDAPS (3 priors), PSLD (LDM-4), DPSImageNet SR ×4 / ×8no change

With twice our tuning budget, no extended baseline overtakes our 5 h result on any task.


W4: PSNR-only tuning and point estimates

Thank you for the valuable suggestion! We address this with the following experiments and analysis.

(A) Re-selecting every method by LPIPS. For each method, we took the configuration with the best validation LPIPS (instead of PSNR) from our existing tuning trials and evaluated it on the test set. The values in the table below are LPIPS (lower is better).

TaskOursNext best
FFHQ SR ×80.1840.207 (DAPS)
FFHQ SR ×160.2850.291 (DAPS)
FFHQ Gaussian deblur σ=6 (noise-free)0.1930.225 (RED-diff)
FFHQ Gaussian deblur σ=12 (noise-free)0.2850.294 (RED-diff)
FFHQ inpaint 95%0.1630.203 (DPS)
FFHQ inpaint 98%0.2020.270 (DPS)
ImageNet SR ×40.1760.188 (DPS)
ImageNet SR ×80.3210.313 (DPS)

When tuned for a perceptual metric instead of PSNR, we remain the best method among the paper's baselines on 7 of 8 tasks.

(B) Are our reconstructions realistic? KID. As the reviewer notes, diffusion samplers generate sharp but stochastic details that pixel-wise metrics penalize. Avoiding exactly these details is the goal of WTS. Details that look plausible but are not supported by the measurement are hallucinations, and for most inverse problems a faithful reconstruction matters more than a sharp one. We therefore do not aim to maximize realism. However, to address the reviewers concern that our reconstructions are unrealistic, we compute KID which measures how close the distribution of a method's reconstructions is to the distribution of real images.

We compare each of our two variants, WTS and WTS+Loop, with every baseline, using a paired bootstrap over the test images (1000 resamples). We call a method better only when the 95% confidence interval of the difference excludes zero; otherwise we count it as a tie. The table below reports KID ×10³ against the test set; cells show WTS / WTS+Loop.

TaskOur KIDBaselines we beatTiesBaselines with lower KID (their KID)
FFHQ SR ×89.6 / 9.34 / 41 / 2DiffPIR 3.2; DAPS++ 8.9 (WTS only)
FFHQ SR ×1627.8 / 24.04 / 40 / 0DPS 9.6, DiffPIR 7.3, DAPS++ 17.7
FFHQ Gaussian deblur σ=6 (noise-free)8.3 / 12.06 / 41 / 1DAPS 9.7, DiffPIR 8.5 (WTS+Loop only)
FFHQ Gaussian deblur σ=12 (noise-free)31.5 / 27.43 / 30 / 1DPS 10.5, DAPS 17.6, DiffPIR 4.3; DAPS++ 25.6 (WTS only)
FFHQ inpaint 95%4.4 / 9.46 / 41 / 2DiffPIR 4.3 (WTS+Loop only)
FFHQ inpaint 98%4.5 / 5.56 / 60 / 0DiffPIR 2.8
FFHQ HDR13.1 / 11.40 / 12 / 2DAPS++ 1.0, RED-diff 4.3, DAPS 5.3, DiffPIR 5.6; PnP-DM 10.1 (WTS only)
FFHQ motion deblur (noisy)11.0 / 5.65 / 61 / 0PnP-DM 2.4
ImageNet SR ×41.5 / 4.610 / 100 / 0DPS ≈0
ImageNet SR ×813.3 / 22.010 / 80 / 1DPS 3.5; LatentDAPS (U-ViT-H) 16.4 (WTS+Loop only)
ImageNet Gaussian deblur σ=6 (noise-free)26.5 / 21.82 / 20 / 0DPS 7.0, DiffPIR 14.3
ImageNet inpaint 95%13.2 / 7.43 / 30 / 0DPS 4.9
FFHQ Gaussian deblur σ=6 (noisy)18.3 / 15.73 / 31 / 2DAPS 13.8, DPS 10.2, DiffPIR 7.5; DAPS++ 14.5 (WTS only)
FFHQ Gaussian deblur σ=12 (noisy)21.6 / 33.44 / 30 / 1DAPS++ 18.3, DAPS 17.6, DPS 9.9, DiffPIR 14.2
FFHQ nonlinear motion deblur (noise-free)11.2 / 11.56 / 60 / 0DAPS++ 4.1
Total72 / 677 / 1230 / 30

Our reconstructions are as realistic as, or more realistic than, the baselines' in 79 of 109 comparisons, for both WTS and WTS+Loop.

In nearly all cases where a baseline has lower KID, our reconstructions are better as measured by PSNR. The exceptions are RED-diff on HDR for WTS+Loop, and RED-diff on HDR and PnP-DM on noisy motion deblur for WTS. This is the typical perception–distortion trade-off. Against the baselines that compete with us on PSNR (RED-diff, PnP-DM, DAPS, DDiff), our KID is better in 23 of 24 comparisons for WTS and 22 of 24 for WTS+Loop on the 6 paper tasks. So our PSNR advantage does not come from smoother, average-looking images. We will add PSNR-vs-KID perception–distortion plots to the paper.

Reviewer_Pba4OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper we introduce Walk the Sphere (WTS), an inverse-problem solver that reconstructs signals from observations by optimizing the latent representation of a pretrained sphere autoencoder. It keeps that representation on a sphere while fitting the observed measurements and WTS+Loop is a variant in which during optimization it periodically decodes and re-encodes the representation.

The evaluation includes natural-image restoration, inverse scattering, black hole imaging and MRI. Image-restoration methods share a tuning budget while scientific comparisons use published results from InverseBench. Experiments are also performed to study how latent compression and arrangement of observations affect reconstruction.

The results show that WTS can compete with diffusion-based methods and an ablation shows the benefit of enforcing the sphere constraint. Performance varies across tasks: MRI remains challenging, sometimes WTS+Loop is worse than WTS, large contiguous missing regions are hard to reconstruct.

The practical contribution is a simple fast approach to reconstruction from a learned generative representation. The main limitations are whether the sphere constraint will produce plausible outputs (or not), how the method differs from earlier constrained latent-search methods and how much of the improvement comes from training the prior, constraint or re-encoding.

Critical strengths
  • The paper makes a strong practical case for an inverse solver with a sphere-autoencoder prior. The image experiments compare priors trained on the same data and solvers given the same tuning budget. WTS achieves competitive or better reconstruction on super-resolution and random inpainting with fast inference.
  • An ablation with the same decoder shows that enforcing the sphere constraint during reconstruction improves reconstruction quality. Adding projection always improves reconstruction quality so this part of the method is directly supported by evidence.
  • The evaluation is done on natural images and scientific problems but openly reports limitations (e.g., MRI performance worse than WTS, cases where WTS+Loop performs worse than WTS, difficult to reconstruct contiguous missing regions). The compression and observation-layout analyses explain when latent inversion works and why the prior’s reconstruction quality by itself does not predict success.
Critical weaknesses
  • Keeping the latent representation on a sphere does not guarantee a plausible reconstruction. The paper admits that some points on the sphere decode poorly so we cannot make the stronger claim that projection keeps the solution on the data manifold. Re-encoding an image generated by a generator also needs evidence that it moves the solution towards more plausible outputs.

  • The novelty and source of the gains need to be separated. PULSE already used spherical latent search, sphere encoder and basic re-encoding step are from previous work as well. The fixed-decoder ablation confirms constraint but comparisons between prior types change their training losses too. Stronger constrained-latent baselines would clarify what this method adds.

  • The headline claims about equal tuning budgets and runtime are not restricted to the controlled comparisons; those conditions apply only to image experiments; the scientific comparisons use published InverseBench results under a different protocol without comparable timings. The overall claims should reflect that distinction.

Questions for authors
  • What does the sphere constraint guarantee besides controlling the latent norm? What evidence is there that re-encoding moves the solution toward more plausible reconstructions?

  • How does WTS compare with a conventional prior using spherical projection or soft norm penalty on the same decoder? How was the constraint disabled in ablation and separate effects of the training losses of the prior from sphere constraint?

  • Which re-encoding schedules were used in practice and how do they affect inference cost? When re-encoding is disabled, what fraction of the difference from WTS comes from retuning other parameters?

4 comments

We thank the reviewer for a careful and constructive review, which motivated additional analysis of our method including the role of the sphere constraint and the re-encoding looping. Below are the results from new experiments we ran during rebuttal answering your questions. Please also see our response to reviewer v5ZA for the new tasks and baselines added during the rebuttal period. We are working on updating our draft accordingly.


W1 / Q1: What does the sphere constraint guarantee, and does re-encoding move toward plausible reconstructions?

We agree with the reviewer that the sphere projection only guarantees that every iterate stays within the set of codes the decoder was trained on and does not necessarily mean it will map to a plausible image. We will revise the text accordingly. The sphere projection does make reconstructions more realistic.

(A) Same decoder, with and without the projection. We replay the paper's FFHQ configurations on 200 test images. "No projection" starts from the same encoding and runs Adam on the raw latent with no normalisation. We report KID (×10³, lower is more realistic) in the table below.

TaskWTSNo projectionFinal latent norm without projection
SR ×86.310.414× the sphere radius
SR ×1624.634.581×
Gaussian deblur σ=6 (noise-free)5.710.45×
Gaussian deblur σ=12 (noise-free)28.230.57×
Inpaint 95%2.334.8136×
Inpaint 98%2.136.5182×

Without the projection the latent drifts 5–182× off the sphere. It fits the measurements better (its final measurement loss is lower than WTS's on all 6 tasks), but reconstructions are less realistic (KID is worse on 5 of 6 tasks; Gaussian deblur σ=12 the gap is within noise). So the sphere projection serves as regularization to our prior, preventing the solver from overfitting the measurement with unrealistic images.

(B) Does re-encoding make outputs more plausible? We ablate the loop by taking WTS+Loop's tuned configuration and turning off the re-encoding loop. This also removes the latent noise which is only injected at re-encodings. Below is the KID in the 9 natural image settings where tuning chose to re-encode, with and without the loop.

SettingLoop onLoop offOn − off (95% CI)
FFHQ inpaint 95%9.487.8−78.4 [−83.0, −73.4]
FFHQ inpaint 98%5.520.6−15.1 [−17.9, −12.3]
FFHQ Gaussian deblur σ=6 (noise-free)12.011.2+0.8 [−0.1, +1.7]
ImageNet SR ×4 (sphere-L)52.950.3+2.6 [+1.9, +3.3]
ImageNet SR ×8 (sphere-L)36.252.4−16.2 [−17.6, −14.9]
ImageNet SR ×8 (sphere-XL)25.632.7−7.2 [−9.1, −5.1]
ImageNet SR ×8 (sphere-L-FT)31.430.1+1.2 [−1.0, +3.6]
ImageNet SR ×4 (sphere-XL-FT)4.38.4−4.1 [−4.8, −3.4]
ImageNet SR ×8 (sphere-XL-FT)22.027.2−5.2 [−6.3, −4.2]

Re-encoding significantly lowers KID on 6 of 9 settings, by up to 9× (inpaint 95%: 87.8 → 9.4). Of the other three settings, one is a slight loss and two are ties. Intervals are 95% confidence intervals from a paired bootstrap over the test images (1000 resamples). This isolates the loop within WTS+Loop's tuned configuration. WTS tuned from scratch without the loop often reaches even lower KID (see our response to wUKg, W4); the loop's main benefit is higher PSNR on inpainting and the scientific problems.

W2 / Q2 (first part): Novelty, source of the gains, and constrained-latent baselines

Novelty. PULSE projects onto a sphere because their GAN was trained on Gaussian latents which concentrate on the sphere. However, the sphere is only a subset of the generators training distribution, a portion of this distribution is inaccessible once the latent is constrained to the sphere. We use a sphere encoder so the decoder is trained only on latents from the sphere. We added encoder initialisation and the re-encoding loop in addition to projected gradient descent. We will revise the paper to more clearer attribute credit to prior work including PULSE and Sphere Encoders.

Constrained-latent baselines. We ran the reviewers suggested ablation for on all 6 FFHQ settings following our 5h tuning protocol.

SR ×8SR ×16Gaussian deblur σ=6 (noise-free)Gaussian deblur σ=12 (noise-free)Inpaint 95%Inpaint 98%
Sphere AE, hard projection (WTS)27.0023.4327.2024.2625.6523.23
Sphere AE, soft norm penalty27.0323.3627.2024.0825.8923.94
Sphere AE, no norm control26.6222.5726.8223.9624.4721.73
Conventional VAE + spherical projection25.4322.0526.1223.0824.2721.93
Conventional VAE, no projection25.4022.0326.0623.0122.2119.57
  • Soft norm penalty on our decoder: it matches the hard projection within 0.2 dB on four tasks and is ahead on inpainting (+0.2 / +0.7 dB).
  • Spherical projection in a conventional VAE latent space: it helps on inpainting (+2.1 / +2.4 dB), but does not win in any setting. The VAE stays 1–2 dB below the sphere autoencoder on every task.
  • No norm control at all: this costs 0.3–1.5 dB.

So the benefit comes from both training the decoder on the sphere, and keeping the iterate on or near it. The hard projection is the parameter-free way to do the latter - there is no penalty weight to tune. As noted in the paper, the VAE reconstructs the test split within 0.5 dB of the sphere autoencoder but trails WTS by 1–4 dB as a solver. The gap is therefore not explained by decoder fidelity.


Q2 (second part): How was the constraint disabled in the ablation?

The "Sphere + GD" rung, uses the same sphere autoencoder checkpoint and starts from the same sphere-normalised latent as the "WTS" rung. The "Sphere + GD" rung then runs Adam, with no projection of the latent anywhere. It is tuned with the same 5 h protocol. The step from Sphere + GD to WTS therefore only adds the projection.


Q3 (first part): Re-encoding schedules and inference cost

The re-encoding frequency is tuned as a fraction of the 300 steps, from {0, 1/2, 1/4, 1/8, 1/16}. A frequency of 1/k gives k−1 evenly spaced re-encodings: 1, 3, 7 or 15. Each re-encoding costs one extra decode and one encode. Across the settings, tuning WTS+Loop chose:

Re-encodings per solveNumber of settings
06 (FFHQ SR ×8, SR ×16, Gaussian deblur σ=12 (noise-free); ImageNet sphere-XL SR ×4, sphere-L-FT SR ×4; inverse scattering 360 receivers)
13
36
74
155 (FFHQ inpaint 98%; inverse scattering 180 and 60 receivers; black hole 3% and 100%)

More underdetermined inverse problems (sparse inpainting, sparse scattering, black hole) choose more frequent re-encoding; depending more heavily on the generative prior.

Cost. Measured on the same GPU at the same step count, WTS+Loop's wall-clock time per image is −1.1% to +3.9% relative to WTS; the negative values are timing noise. Even at 15 re-encodings the overhead is at most about 3%. For example, FFHQ inpaint 98% takes 1.82s for WTS+Loop vs 1.77s for WTS, and inverse scattering at 180 receivers takes 4.02s for WTS+Loop vs 3.93s for WTS. The overhead is small because 15 decode–encode passes are negligible compared to 300 decoder forward and backward passes.

Q3 (second part): With re-encoding off, how much of the gap is retuning?

We took WTS+Loop's tuned configuration and switched re-encoding off without retuning anything.

TaskWTS+LoopSame settings, loop offWTS (retuned, no loop)
FFHQ inpaint 95%25.9220.4225.65
FFHQ inpaint 98%23.9522.4923.23
FFHQ Gaussian deblur σ=6 (noise-free)27.1427.0427.20
ImageNet SR ×4 (sphere-XL-FT)25.7424.8726.38
Inverse scattering, 60 receivers34.0531.1333.11
MRI ×428.2123.0126.74
Black hole, 100%27.1524.8726.56

Turning the loop off always costs PSNR, and retuning without the loop always recovers at least part of that loss. This is because WTS+Loop's learning rate and Adam moments are adjusted to the loop. We find that in the settings where the measurement constrains the image the least (inpainting, sparse scattering, MRI, black hole), the loop makes a real contribution.


W3: Scope of the equal-budget and runtime claims

This is a good catch! Thank you for pointing this out. The equal tuning budget and the runtime comparisons apply to the image tasks, where we tune and time every method ourselves on the same hardware (some baselines added during the rebuttal period were tuned on B300s instead of the B200s which were used for the submission results, this benefits the baselines giving them an edge over us). For the scientific problems, the baseline numbers are InverseBench's published results under their own tuning protocol. We will scope the claims accordingly in the paper.

Thank you again for your thoughtful review. We hope our rebuttal experiments have answered your questions. If you have any remaining questions we are happy to respond to those as well.

Comments on the paper

5 comments
~Sean_Man1OpenReview ↗

Related Work: SP³ (NeurIPS 2026)

We would like to bring a closely related paper to the attention of the authors and reviewers: SP³: Spherical Priors for Plug-and-Play Restoration, accepted to NeurIPS 2026, first publicly available on arXiv on June 15, 2026 (https://arxiv.org/abs/2606.16396), with the camera-ready version available on our project page (https://man-sean.github.io/sp3-website/). For transparency, I am a coauthor of SP³.

Relationship between the formulations, methods, and review discussion:

  • Spherical PGD (S-PGD vs. WTS): SP³ also explores Sphere Encoders (Yue et al., 2026) as generative priors for inverse problems. In particular, §3.1 (Eq. 5) formulates measurement fitting under a spherical latent constraint and introduces Spherical Projected Gradient Descent (S-PGD) as a baseline that projects the data-fidelity gradient onto the sphere's tangent space and retracts the updated latent onto the sphere. This shares the core objective and optimization strategy of WTS (§3, Eq. 6), while differing in optimization details such as WTS's encoder warm-start and Adam updates. Regarding Reviewer Pba4's question on soft norm penalties vs. hard projection on the same decoder, SP³ (§3.1, Eq. 6, Table 1, and Fig. 2) also compares against a soft spherical norm penalty baseline (S-GD) on the same Sphere Encoder.
  • Role of the encoder and noisy re-encoding (§3.1 to §3.2 vs. WTS+Loop): SP³ (§3.1 to §3.2) discusses the limitation of optimizing solely through the decoder and analyzes the noisy spherical encode-decode cycle P(x)=D(S(E(x),σ))P(x) = D(S(E(x), \sigma)) as a manifold projection operator. In connection with Reviewer Pba4's question on whether re-encoding moves iterates toward more plausible reconstructions, SP³ (Fig. 3) provides empirical analysis showing that repeated noisy SE projections pull degraded inputs toward the natural image manifold (lowering both KID and LPIPS), that injecting latent noise (σ>0\sigma > 0) stabilizes repeated re-encoding, and that standard VAEs diverge under the same loop. This closely relates to the motivation and mechanism of WTS+Loop (§3.2, Eq. 7).
  • Gradient-free PnP vs. backpropagation (SP³ vs. WTS / WTS+Loop): In connection with Reviewer v5ZA's question (W4, Q1) on gradient-free alternatives to backpropagating through the decoder, SP³ (§3.3, Alg. 1) embeds P(x)P(x) inside Half-Quadratic Splitting (HQS), alternating the noisy encode-decode projection with a closed-form image-space proximal step. Avoiding backpropagation through the generative model yields a gradient-free solver with anytime restoration and empirical convergence, running in 0.28 s per 256x256 image for 20 steps (and 0.016 s for 1 step) on a single L40S GPU, compared to 1.17 s for S-PGD (Table 1). Additionally, related to Reviewer wUKg's comment (W4) on PSNR vs. perceptual quality, SP³ (§4, Fig. 5; §5) evaluates Sphere Encoder priors in the perception-distortion plane (KID vs. LPIPS and PSNR).
  • NeurIPS 2026 camera-ready additions: While the points above appear in the June 2026 arXiv preprint, the camera-ready version (PDF) includes several additional analyses relevant to the discussion: (1) extensions to non-linear inverse problems (§4, Table 2, Fig. 9), including JPEG artifact removal and non-linear motion deblurring without backpropagating through the SE prior; (2) trajectory-local contraction and α\alpha-averagedness experiments (Appendix D.3), examining the local contractiveness of PP and the full SP³ update along restoration trajectories and the stabilizing role of latent noise (σ>0\sigma > 0); and (3) ablations on replacing SE with VAEs, OOD generalization, and multi-seed averaging to trade perceptual sharpness for higher PSNR (Appendix B).

The present submission investigates a broader set of tasks and domains, including ImageNet-256 and scientific inverse problems from InverseBench, as well as the periodic re-encoding schedule in WTS+Loop. We believe discussing SP³ would help readers contextualize the shared spherical-prior formulation and noisy re-encoding mechanism alongside the differences in solver design and experimental scope.

Our project page, camera-ready PDF, and open-source code are available at https://man-sean.github.io/sp3-website/ for reference. We hope this is helpful to the authors and reviewers, and we would be happy to clarify any technical details.

Thank you for brining this to light, I will read it.

Thank you for the pointer to SP³, and for your summary of the methods. We will cite SP³ in the related work of our revision and discuss how it relates to our method.

Concurrent work. The two papers were developed independently, both building on the sphere encoder of Yue et al. (2026). SP³'s peer-reviewed publication at NeurIPS 2026 came after the July 17, 2026 cutoff in the ICLR 2027 reviewer guidelines, so it is contemporaneous work under those guidelines.

Substantive differences.

  • Projected gradient descent on the sphere. SP³ includes S-PGD as a baseline for one task (AFHQ-Cat deblurring). We study it as a solver in its own right: with an encoder warm start, Adam updates and an optional re-encoding schedule. In a controlled comparison of solvers, it is a leading method in terms of reconstruction performance on a variety of image (FFHQ-256 & ImageNet-256) and scientific (InverseBench) tasks.
  • Solver design and goals. SP³ places the noisy encode–decode cycle inside a gradient-free half-quadratic splitting scheme, emphasising speed and perceptual quality. We optimize the latent directly against the measurement by backpropagating through the decoder, aiming for the most accurate single reconstruction.
  • Soft versus hard constraints. SP³ reports a soft spherical norm penalty (S-GD) well below S-PGD on AFHQ-Cat deblurring. In our setting, a soft norm penalty on our decoder, tuned under the same 5h budget, matches the hard projection within 0.2 dB on four of six FFHQ tasks and is slightly ahead on inpainting.
  • Scope. We evaluate on 29 settings across natural and scientific imaging, including nonlinear operators and physical forward models.

Benchmarking SP³. We use SP³'s public code, and the same FFHQ sphere autoencoder checkpoint as our methods, to tune and evaluate SP³. We use the same tuning protocol as for every method we ran: 5h on the validation split. Due to the rebuttal timeline, SP³ was tuned mostly on B300s, which are faster than the B200s our method used on the original six tasks.

PSNR (dB) ↑

TaskWTSWTS+LoopSP³
SR ×827.0026.9925.91
SR ×1623.4323.1622.51
Random inpainting 95%25.6525.9223.59
Random inpainting 98%23.2323.9521.18
Gaussian deblur σ=6 (noise-free)27.2027.1430.71
Gaussian deblur σ=12 (noise-free)24.2624.0135.46
Gaussian deblur σ=6 (noisy)26.0226.2125.22
Gaussian deblur σ=12 (noisy)23.4623.4122.60
Nonlinear motion deblur (noise-free)30.4930.4726.31

SSIM ↑

TaskWTSWTS+LoopSP³
SR ×80.7600.7590.730
SR ×160.6610.6490.629
Random inpainting 95%0.7710.7760.700
Random inpainting 98%0.7030.7210.614
Gaussian deblur σ=6 (noise-free)0.7580.7580.852
Gaussian deblur σ=12 (noise-free)0.6590.6600.921
Gaussian deblur σ=6 (noisy)0.7180.7290.689
Gaussian deblur σ=12 (noisy)0.6470.6500.616
Nonlinear motion deblur (noise-free)0.8670.8680.716

LPIPS ↓

TaskWTSWTS+LoopSP³
SR ×80.1940.1940.257
SR ×160.2890.2970.372
Random inpainting 95%0.1720.1610.298
Random inpainting 98%0.2370.2080.361
Gaussian deblur σ=6 (noise-free)0.1950.2060.126
Gaussian deblur σ=12 (noise-free)0.3170.3040.068
Gaussian deblur σ=6 (noisy)0.2630.2540.345
Gaussian deblur σ=12 (noisy)0.3080.3240.386
Nonlinear motion deblur (noise-free)0.1120.1130.248

Outside noise-free Gaussian deblurring, both of our variants beat SP³ on every task and every metric.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.