acceptodds
Under review as a conference paper at ICLR 2027

Fidelity-Preserved Diffusion-Based Image Super-Resolution with Guidance

Abstract

While diffusion-based image super‑resolution models have demonstrated remarkable performance in generating visually appealing images, they are prone to hallucination, often producing image details that appear deceptively realistic but are, in fact, spurious. We attribute this to the probabilistic generative nature of diffusion models, yet we contend that generation can be guided. To this end, this paper proposes a conditioning strategy that guides the diffusion process to preserve the fidelity with the given low-resolution (LR) observation, and consequently prevents hallucinatory detail artifacts. Specifically, we derive a theoretical formulation, enabling a unified conditional posterior score to guide the diffusion generation process, including both the training of the denoising autoencoder and the inference of super-resolved images. During training, the LR observation serves as a condition input to guide the denoising autoencoder, while the noise predicted by the autoencoder is used to calculate the consistency loss against the high-resolution (HR) image, steering the training to be well-aligned with the generation of super-resolved images. During inference, where the HR ground truth is unavailable, we instead pair the LR observation with itself. One serves as an input to the trained autoencoder, whereas the other is utilized to compute the residual against the inferred super-resolved image, thereby guiding the inference process. Experimental results demonstrate that our method effectively suppresses hallucinations, achieving significantly more faithful super-resolved results on digits, text characters, and texture details compared to competing methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
diffusion model, image fidelity, image super-resolution

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aJKpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper adds explicit guidance to diffusion-based super-resolution during both training and inference to improve content fidelity. It combines HR-supervised gradient guidance during training with LR observation-consistency guidance during sampling, implemented on StableSR. Evaluation uses bicubic-generated LR images from RealSR and DRealSR. Quantitative comparisons, visual examples, and ablations report improved reconstruction under this known-degradation setting.

Critical strengths

On the two test sets synthesized using bicubic degradation, the full method improves PSNR over StableSR by 6.184 and 6.571 dB, respectively, while also improving LPIPS. Relative to inference guidance alone, adding training guidance provides further PSNR gains of 0.393 and 0.383 dB, respectively, and further reduces LPIPS. Some text examples are also closer to the ground truth. These are meaningful empirical benefits.

Critical weaknesses

(1) About the central posterior interpretation and the new mechanisms

At inference time, the explicit condition is a deterministic transformation of the existing LR observation, y~=ϕt(y)\tilde y=\phi_t(y). Consequently,

p(zt∣y,ϕt(y))=p(zt∣y).p(z_t\mid y,\phi_t(y))=p(z_t\mid y).

Conditioning again on this transformation therefore introduces no new posterior information. The additional nonzero guidance in Eq. (13) may serve as a data-consistency correction to an approximate conditional model, but it does not follow simply from conditioning twice on the same observation.

Moreover, DifIISR (DifIISR: A Diffusion Model with Gradient Guidance for Infrared Image Super-Resolution, CVPR25) already injects loss gradients into noise predictions during training and uses a closely corresponding normalized log-amplitude spectral loss. FGPS (Frequency-Guided Posterior Sampling for Diffusion-Based Image Restoration, ICCV25) already introduces guidance based on observation residuals filtered by a time-varying low-pass operator. The paper needs to establish the specific technical contribution of combining its training and inference stages; describing the combination as a unified framework is not sufficient.

(2) About evaluation under a known bicubic degradation

The paper states that test LR images are generated by bicubic downsampling the HR images of RealSR and DRealSR, with the same known operator used during inference. However, performance under real imaging degradations remains untested.

The comparison lacks fidelity-oriented baselines matched to bicubic degradation, such as classical 4x SwinIR, as well as controls using ordinary fine-tuning with the same budget, direct spectral-loss supervision, and existing guidance methods. The independent benefit of the proposed mechanism is therefore difficult to identify. The text examples are encouraging, but systematic content-error statistics are absent. LR consistency cannot exclude incorrect HR details in the null space of the downsampling operator, and no-reference scores exceeding those of the ground truth do not substitute for an evaluation of hallucinations.

Questions for authors

How is the nonzero guidance term in Eq. (13) justified given that y~=ϕt(y)\tilde y=\phi_t(y) adds no information beyond the existing condition yy?

Are the fidelity conclusions intended to be limited to the known bicubic setting evaluated in Section 4.1?

What specific technical contribution distinguishes the proposed guidance from the training guidance in DifIISR and the inference guidance in FGPS?

Under the current bicubic test protocol, how does the method compare with classical ×4 SwinIR in reconstruction fidelity?

Reviewer_ZdxXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary
  1. This paper tries to solve the hallucination problem in diffusion-based image super resolution. Their argument is that the hallucination of diffusion-based methods are mainly in the high-frequency details that are not supported by the low-resolution input. One example is the characters in the Figure 1.
  2. They add explicit fidelity guidance in addition to the implicit conditions like the original low-resolution image.
  3. They inject the visual perception loss into the estimated score function for denoising. And instead of calculating the perception loss on the spatial domain, they calculate it in the frequency-domain. In particular, the distance between the input and the downgraded intermediate diffusion prediction.
  4. Built on the stablesr, the proposed method is better on PSNR, SSIM, and LPIPS compared with other methods.
Critical strengths
  1. The idea is simple and easy-to-understand. one extra visual perception loss is added to the noise prediction.
  2. Instead of only doing the rectification during the inference, they do it in both training and inference stages,
  3. In table 1, it shows that the proposed method is good at full-reference IQA metrics like PSNR, SSIM, LPIPS.
  4. Ablation study shows the importance of guidance in both training and inference stages.
  5. Examples in the Figures show good results on high-frequency details.
Critical weaknesses
  1. The most important weakness that I want to mention is: this method is not blind. During both training and inference, they need the the downgrading operator A, which is unknown for real-world cases.
  2. The main idea is applying the training-free guidance of some existing diffusion posterior/data-consistency guidance to the training stage. But my another concerns is that training does not help too much based on the ablation study, 25.01vs 25.59. Adding the guidance during the inference significantly helps the PSNR, 25.01 vs 30.81 , but this is not new.
  3. Why frequency-domain consistency instead of pixel-domain consistency? Any experimental results?
  4. Any user study about the hallucination? In my opinion, there is no perfect algorithm to evaluate the hallucination yet.
Questions for authors
  1. Why do you generate LR test images using bicubic downsampling instead of using the original LR images from RealSR/DRealSR?
  2. What will happen if I do not know A or the estimation of A is wrong?
  3. Could you provide a direct ablation between pixel-domain guidance and frequency-domain guidance?
  4. Can you evaluate hallucination more directly, for example using OCR accuracy or digit recognition accuracy on the text/number examples? Or user study?
Reviewer_dCNROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper adds explicit frequency-domain consistency guidance to a conditional diffusion super-resolution model during training and inference. Training uses HR spectral information, while inference applies a time-dependent low-pass consistency term against the LR observation. Experiments use RealSR and DRealSR HR images with synthetic bicubic LR inputs, with additional DIV2K results in the appendix.
Critical strengths
  1. Table 1 reports PSNR gains of 2.395 dB on RealSR and 1.799 dB on DRealSR over LDM, the strongest PSNR baseline in that table. The method also achieves the best SSIM and LPIPS there. The character and digit examples show why the fidelity gains matter.
  2. Table 2 separates training and inference guidance. Inference guidance accounts for most of the improvement, while training adds approximately 0.39 dB on both datasets. This ablation helps explain the contribution of each stage.
Critical weaknesses
  1. The posterior derivation does not justify the inference rule as written. Algorithm 2 sets tilde_y = phi_t(y), so p(z_t|y,phi_t(y)) = p(z_t|y). Equation (9) replaces p(tilde_y|x_0,z_t,y) with p(tilde_y|x_0), an independence assumption that is generally invalid when tilde_y reuses y. The empirical gains support the usefulness of the added gradient, but its posterior interpretation needs revision.
  2. The experiments do not isolate the benefit of frequency-domain guidance from ordinary data consistency using the known degradation operator. A comparison with pixel-domain or standard data-consistency guidance on the same backbone would address this directly. Section 4.1 uses bicubic LR inputs generated from HR crops, and the DIV2K appendix follows the same setting, so the results support a narrower fidelity claim than robustness to captured degradations. I lean toward rejection because the theoretical interpretation and the evidence for the specific guidance design remain incomplete.
Questions for authors
  1. What probabilistic model supports Equation (9) when tilde_y = phi_t(y)? Please explain how the deterministic reuse of y yields the additional posterior-score term, or revise its interpretation.
  2. How does frequency-domain guidance compare with pixel-domain or standard data-consistency guidance under the same backbone and compute budget? Please also state whether any evaluation uses the original captured LR observations.
  3. Please specify the eta_t and phi_t schedules, whether training differentiates through the guidance gradient, and the runtime and memory overhead. Algorithm 2 samples noise in step 4 but never uses it; does the implementation use a deterministic update?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.