acceptodds
Withdrawn from ICLR 2027

Single Image Super-Resolution via Bivariate À Trous Wavelet Diffusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The effectiveness of super-resolution (SR) models hinges on their ability to recover high-frequency structure without introducing artifacts. Diffusion-based approaches have recently advanced the state of the art in SR. However, most diffusion-based SR pipelines operate purely in the spatial domain, which may yield high-frequency details that are not well supported by the underlying low-resolution evidence. On the other hand, unlike supervised SR models that may inject dataset-specific textures, single-image SR relies primarily on internal image statistics and can therefore be less prone to dataset-driven hallucinations; nevertheless, ambiguity in the LR observation can still lead to inconsistent high-frequency details. To tackle this problem, we introduce BATDiff, an unsupervised Bivariate À trous Wavelet Diffusion model designed to provide structured cross-scale guidance during the generative process. BATDiff employs an à Trous wavelet transform that constructs an undecimated multiscale representation in which high-frequency components are progressively revealed while the full spatial resolution is preserved. As the core inference mechanism, BATDiff includes a bivariate cross-scale module that models parent–child dependencies between adjacent scales. It improves high-frequency coherence and reduces mismatch artifacts in diffusion-based SR. Experiments on standard benchmarks demonstrate that BATDiff produces more structurally consistent reconstructions than existing diffusion and non-diffusion baselines, achieving improvements in fidelity and perceptual quality.

Primary area:
generative models
Keywords:
Super-resolution, diffusion models, wavelet transforms

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gM3xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Authors propose to perform a multi-level wavelet decomposition of LR image, and uses a diffusion technique to better estimate the residuals and then adds it to the bicubic interpolated image to get the superresolution done.

Critical strengths

It is a good thing to explore the use of wavelets for SR (albeit not a new technique -- see comments later). It also obviates the need for "paired" LR-HR training data. Experiments are adequate and results are a bit better than SOTA.

Critical weaknesses

The key step in the manuscript is eq (7) that tries to utilize a conditional probability relationship between various layers (scales) of wavelet representation. I have a major concern for this -- if the lower scale image generation is done through a proper choice of wavelet, then the residual (after LPF) should be orthogonal to the LR image and if that is true, there is not a conditional probability as the conditioning falls off, ie, p(x/y) = p(x). How do you justify this? Your diffusion process then becomes unconditioned. How does it then converge?

You have done several ablation studies regarding various (hyper) parameters of the learning process, but not the most important part related to the wavelet decomposition! How does the reconstruction change if I use different types of wavelets (you may use different Daubechis (sp?) wavelets and then Haar and other wavelets). You must do an ablation on that.

Although the manuscript does refer to the most recent relevant citations, but it is very poor in terms of refering to distant past -- as if it does not exist! There have been several works on single frame super resolution, use of wavelet or contourlet for image SR without and LR-HR paired training (off hand, I remember C.V. Jiji's work in around 2004-5, pl also check P. Milanfar, etc). They did precise what you did except, of course, the ML part.

Questions for authors

See above.

Reviewer_KgMzOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

BATDiff learns a diffusion model from a single bicubic-upsampled LR image using an undecimated à trous wavelet representation. It combines time-aligned parent-scale conditioning with LR-consistency updates to guide multiscale reconstruction. Experiments cover standard, blind, and real-world SR benchmarks, with ablations comparing time-aligned conditioning against clean-parent and no-parent variants.

Critical strengths

On Urban100 at ×4, Table 5 reports a PSNR of 28.53 dB with time-aligned parent conditioning, compared with 27.46 dB without parent conditioning and 28.14 dB with a clean parent condition; LPIPS also improves. This supports the hypothesis that the parent's diffusion-time state can affect restoration quality and is the paper's most valuable empirical finding.

Critical weaknesses

(1) Substantive gaps in the algorithm definition and the explanation of its restoration mechanism

Under Eq. (4), w(1)w^{(1)} represents the finest-scale details, but Eq. (6) adds this component first, contrary to the claimed coarse-to-fine order. Algorithm 1 also directly assigns the clean output of the parent scale to xT(s)x_T^{(s)} at the next scale, whereas training inputs in Eq. (14) follow the standard forward noising process. The joint noise-generation procedure for the parent and child scales is unspecified.

Furthermore, every clean training target is constructed from the bicubic reference, and the final target is exactly U(y)U(y). This does not prove that the model can only reproduce bicubic interpolation. However, the authors need to explain how the network's receptive field, training-sample construction, and sampling mechanism prevent direct memorization of the reference and enable additional restoration capability. The current description does not establish which algorithm produces the reported results or whether these inconsistencies are merely errors in the presentation.

(2) About the contributions and SOTA claims

Single-image internal learning, a shared scale-conditioned denoiser, and multiscale diffusion already appear in SinDDM. Progressive frequency-restoration targets constructed through wavelet decomposition already appear in FDDiff. The candidate contribution should therefore be defined more specifically as the proposed time-aligned parent conditioning within an à trous representation, and evaluated against directly related internal-learning methods under comparable budgets.

In the experiments, the observation operator DD used in blind and real-world settings, the metric channels and border-cropping rules, and the sources of baseline results are insufficiently specified. Some claims of best performance also contradict the tables: the Set5 PSNR of 32.89 is below SwinFIR's 33.08, and the DRealSR NIQE of 6.1326 is higher, hence worse, than PASD's 5.8595. These issues prevent the current evidence from establishing the broad SOTA claims.

Questions for authors

How is initialization from a clean parent-scale output in Algorithm 1 reconciled with the noisy training distribution in Eq. (14)?

How much improvement does the learned diffusion model provide over bicubic interpolation followed by the same LR-consistency updates?

How is the observation operator DD obtained for the blind and real-world experiments?

Under matched degradation information and comparable optimization budgets, how does the method compare with SinDDM?

Reviewer_ARyuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BATDiff, a single-image super-resolution framework that combines an undecimated à trous wavelet transform with bivariate conditional diffusion. The method constructs multi-scale reconstruction targets from bicubic-upsampled LR images and trains a shared DDPM through image-specific optimization without external paired LR–HR data. During inference, the denoising process incorporates time-aligned parent–child dependencies across adjacent wavelet scales, while an LR-consistency constraint is applied to maintain fidelity to the observed input. Experiments on standard and real-world super-resolution benchmarks report improvements in reconstruction quality and structural consistency over several existing methods.

Critical strengths

1. Clear Presentation and Organization. The paper is well-organized and clearly written, with a logical progression from the motivation and problem formulation to the proposed methodology and experimental evaluation. The overall framework is easy to follow, and the illustrations help explain the proposed approach.

Critical weaknesses

1. Insufficient Explanation of Performance Gains

The scale-specific training targets are entirely derived from bicubic-upsampled LR images (Eqs. 4–6), without access to genuine HR high-frequency information. Nevertheless, BATDiff achieves substantial improvements, such as 28.53 dB vs. 24.69 dB for ZSSR on Urban100 (×4) in Table 1. The manuscript lacks a convincing explanation of how such significant gains arise from these limited training signals, and fails to isolate the contributions of the diffusion prior, cross-scale conditioning, and data consistency.

2. Limited Novelty and Incomplete Related Work

The proposed method largely follows a conventional conditional DDPM formulation, with wavelet-derived cross-scale information introduced as additional conditioning rather than fundamentally modifying the diffusion process. Moreover, closely related CVPR 2026 studies, such as HDW-SR, which explores wavelet-guided diffusion for SR, and FiDeSR, which investigates frequency-aware one-step diffusion SR, are neither adequately discussed nor compared. These omissions weaken the novelty claims and make it difficult to distinguish BATDiff from recent frequency-guided diffusion approaches.

3. High Computational Cost and Limited Practicality

As reported in Table 3, BATDiff requires 120,000 image-specific optimization steps (approximately 9 minutes per image on an NVIDIA B200 GPU), followed by 100-step reverse diffusion sampling. Since optimization must be repeated for every new image, the actual end-to-end computational cost is substantial, particularly compared with recent one-step diffusion-based SR approaches. The manuscript lacks sufficient efficiency–performance analysis to justify this expensive per-image optimization or demonstrate its advantages over more efficient reconstruction alternatives.

Questions for authors

Q1. Since all scale-specific training targets are derived from the bicubic-upsampled LR image (Eqs. 4–6), what enables the proposed framework to recover high-frequency details that are absent from the reference image? In particular, how can the authors explain the substantial 3.84 dB PSNR improvement over ZSSR on Urban100 (×4) in Table 1?

Q2. The proposed bivariate mechanism essentially introduces cross-scale information as an additional condition to a standard DDPM. What specifically justifies the use of iterative diffusion rather than a direct cross-scale predictor, such as an autoregressive Transformer? Is there any theoretical or empirical evidence demonstrating that diffusion is necessary or particularly beneficial for the proposed wavelet-based reconstruction?

Q3. In Algorithm 1, each scale is sampled sequentially, while the denoiser requires the time-aligned parent state xt(s−1)x_t^{(s-1)} at every reverse step. How exactly are these parent states obtained during inference? Furthermore, how is initializing xT(s)x_T^{(s)} with the previously denoised estimate x^0(s−1)\hat{x}_0^{(s-1)} reconciled with the forward diffusion distribution used during training?

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

Related papers

Discussion (0)

Sign in to comment.