acceptodds
Under review as a conference paper at ICLR 2027

LPNSR: Learnable Noise Prediction for Diffusion-Based Image Super-Resolution

Abstract

Diffusion-based image super-resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) observations. A key property of diffusion models is that, once the starting point of the reverse chain and the denoising network are fixed, the quality of the generated image is dictated by the noise maps sampled at the intermediate steps, which are drawn from an unconstrained standard Gaussian distribution in conventional pipelines. This property raises a natural question: does there exist a noise sampling distribution better than the standard Gaussian that improves the quality of the generated images? To this end, we use a parameterized deep neural network to predict the mean and the variance of the Gaussian noise sampling distribution at each intermediate step, and design two training schemes, one supervised by the quality of the final generated image and the other by aligning each reverse step with the forward posterior. Experiments show that, compared with the standard Gaussian distribution, the learnable noise sampling distribution improves the quality of the generated images.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image super-resolution, Diffusion models, Noise sampling distribution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bhqHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LPNSR, a learnable noise prediction framework built on ResShift, a four-step residual-shifting diffusion model for image SR. The authors observe that, when the initial latent and pretrained denoiser are fixed, the intermediate noise samples determine the variation in the final reconstruction. Instead of sampling these noise maps from a standard Gaussian distribution, LPNSR introduces an additional network that predicts a conditional Gaussian distribution at each intermediate reverse step. Specifically, the paper proposes two training schemes: 1) LPNSR-A: The reverse diffusion chain is unrolled, and the noise predictor is optimized using latent-space L2, LPIPS, and adversarial losses on the final reconstruction. 2) LPNSR-B: The network predicts the residual between the ground-truth latent and the frozen denoiser's estimate. The authors derive the optimal noise mean from per-step KL minimization and train the residual predictor using L2 and LPIPS losses, while keeping the noise variance fixed.

Critical strengths
  1. Clear formulation and implementation: The paper is well organized, and the proposed modification to ResShift is straightforward. The pretrained denoiser remains frozen, and the noise predictor is incorporated into the existing reverse update through a clearly specified parameterization.

  2. Useful ablations: The authors investigate sensitivity to inference seeds (Table 2), loss functions (Table 4), deterministic versus stochastic prediction (Table 8), initialization versus intermediate noise prediction (Table 10), and inference efficiency (Table 7). The inclusion of computational overhead is particularly useful for assessing the method's practical value.

  3. Perceptual performance: LPNSR-A achieves the best perceptual metrics among the compared methods, without relying on large-scale text-to-image pretraining.

Critical weaknesses
  1. Novelty: The novelty discussion would also benefit from a more careful distinction from prior work* on learned reverse-process variances, learned diffusion samplers, and adversarially trained reverse transitions. The distinction between learning a noise distribution and adding a residual-correction network is insufficiently established. The proposed formulation is presented as learning the distribution of intermediate diffusion noise. However, the interpretation of this contribution differs substantially between the two training schemes.
  • Bridging the Distribution Gap to Harness Pretrained Diffusion Priors for Super-Resolution, ICLR 2026
  • Arbitrary-steps Image Super-resolution via Diffusion Inversion, CVPR 2025
  1. The comparisons do not adequately control for additional model capacity, training data, and loss functions: LPNSR increases the parameter count from 173.92M to approximately 292.52M, corresponding to a 70% increase. It also requires approximately 30% more FLOPs and 37% greater inference latency. These are substantial additional resources, yet the main comparisons do not include a baseline with a matched parameter or computational budget.

  2. Fairness: The noise predictor is trained on approximately one million images collected from LSDIR, DIV2K, Flickr2K, and filtered web images, using objectives that include LPIPS and, for LPNSR-A, adversarial supervision. These training conditions are not matched to those of the frozen ResShift baseline. A particularly important missing comparison is to fine-tune the original ResShift denoiser, or train a larger denoiser, using the same data, objectives, and computational budget. Such experiments would help determine whether LPNSR provides benefits beyond conventional perceptual fine-tuning or increased network capacity. Although Table 9 includes a capacity-matched experiment, it uses considerably smaller models trained from scratch. It therefore does not fully address the capacity and training-budget confounds in the main experiments. The authors also acknowledge that residual correction produces larger gains at this smaller scale, where the base denoiser is less accurately fitted.

  3. Limited evaluation of generalization and practical applicability: The experiments are restricted to a single diffusion backbone, ResShift, with four reverse steps. The method is not evaluated on alternative diffusion architectures or longer sampling schedules, despite the broader framing of learnable intermediate noise distributions. The evaluation also lacks real-world SR benchmarks such as RealSR or DRealSR.

Questions for authors
  1. Ground-truth perceptual quality: Can the authors report no-reference perceptual metrics for the HR images in Table 1 to substantiate the claim that LPNSR-B approaches the perceptual ceiling of the ground truth?
3 comments
AuthorsOpenReview ↗

Response to Reviewer bhqH

We thank the reviewer for the constructive comments and we address the questions below point by point.

Response for Q1:

To substantiate this claim, we compute the three no-reference metrics on the HR ground-truth images and compare them with the outputs of Origin and LPNSR-B:

DatasetMethodMUSIQ↑MANIQA↑CLIPIQA↑
LSDIRHR (ground truth)71.76660.54120.6978
Origin67.47410.43630.6445
LPNSR-B68.38110.45820.6722
ImageNetHR (ground truth)68.71580.45790.6619
Origin65.53880.39280.6238
LPNSR-B66.63610.40950.6476

On both datasets the ground truth scores highest on all three metrics, and LPNSR-B lies strictly between Origin and the ground truth in every case: the perceptual quality of LPNSR-B moves toward the ceiling set by the ground truth.

Response for weakness:

  1. Modifying the sampling distribution of the noise injected at an intermediate step is, by construction, modifying the mean and the variance of the reverse process itself: since the reverse step reads xt−1=μθ+Σθ zt−1x_{t-1}=\mu_\theta+\sqrt{\Sigma_\theta}\thinspace z_{t-1}, a noise distribution N(μω,σω2I)\mathcal{N}(\mu_\omega,\sigma_\omega^2 I) shifts the reverse mean to μθ+Σθ μω\mu_\theta+\sqrt{\Sigma_\theta}\thinspace\mu_\omega and scales its variance by σω2\sigma_\omega^2. LPNSR-B therefore amounts to replacing N(0,I)\mathcal{N}(0,I) with N(αtηtΣθfω,I)\mathcal{N}\left(\frac{\alpha_t}{\eta_t\sqrt{\Sigma_\theta}}f_\omega, I\right).

Distinction from InvSR and DM-SR. Both methods attempt to construct a diffusion starting point that improves the quality of the generated image. Our target is instead the noise injection at the intermediate diffusion steps. InvSR already considered a predictor for the intermediate steps (InvSR-Int), but its training scheme brought no gain; our contribution is to provide effective ways of training such a predictor.

  1. For matched parameter and computational budgets, Table 9 in Appendix A.5 trains two models from scratch with comparable capacity (48.19M vs 50.16M) and the same total budget of 300k iterations. In this setting the residual-correction scheme improves PSNR by 0.39 dB over the monolithic model, more than the 0.10 dB it gains at full scale, which indicates that the benefit comes from the formulation rather than from the additional parameters. We understand the concern about the inference latency and the additional parameters. We would point out that our focus and contribution is the training of an intermediate noise predictor rather than these costs, and that reducing the overhead by optimizing the capacity and architecture of the predictor is an interesting direction for future work.

  2. Under the same training setting and budget, we fine-tune the ResShift backbone. The results are as follows:

MethodPSNR↑SSIM↑LPIPS↓DISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
Origin (frozen backbone)20.700.52810.22490.142067.47410.43630.6445
Backbone fine-tuning20.610.52390.22150.141367.73480.43970.6511
LPNSR-A19.770.51400.20990.133671.79040.56180.7389
LPNSR-B20.800.53450.21970.139468.38110.45820.6722

Fine-tuning decreases fidelity slightly and increases perceptual quality slightly, but the change is small: the ResShift backbone is already well trained on natural images, so further fine-tuning can hardly yield a noticeable benefit. We would also point out that the gain of LPNSR-A over Origin at full scale (MUSIQ +4.32) is far larger than what fine-tuning provides, so it cannot be attributed to compensating an under-fitted base denoiser.

  1. We have added evaluations on two real-world datasets, RealSR and ScreenSR, together with additional qualitative comparisons, in Appendices A.7 and A.8; on both datasets the proposed method still brings performance gains. We agree that validating this training scheme for the noise predictor on other diffusion frameworks and on longer sampling schedules is a valuable direction, which we leave to future work.

We would be grateful to know whether our response has addressed your concerns. Your feedback is important for us to further improve the paper, and we remain available for any further clarification during the discussion period.

As the discussion period closes on November 18 (AoE), we would like to check whether our response above has addressed your concerns. We have not yet received further comments from you, and any feedback before the discussion closes would be very helpful; we remain available to provide clarification immediately.

Reviewer_Dhc4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LPNSR, a learnable noise prediction framework for diffusion-based image super-resolution. Instead of sampling intermediate noise from a fixed standard Gaussian distribution, LPNSR learns a conditional noise distribution while keeping the pretrained denoiser frozen. Two training strategies are introduced: LPNSR-A uses end-to-end supervision to improve perceptual quality, while LPNSR-B employs KL-based mean correction to enhance reconstruction fidelity. Experiments on LSDIR and ImageNet demonstrate improvements across multiple image quality metrics, supported by ablation studies and distribution alignment analysis.

Critical strengths

1. Clear Motivation and a Novel Research Perspective

This paper investigates the impact of intermediate noise on the final generation quality during the reverse diffusion process. Unlike existing methods that primarily focus on optimizing denoising networks, conditioning information, or initialization noise, this work proposes learning conditional noise distributions at intermediate sampling steps to improve super-resolution performance. This perspective offers a certain degree of novelty, and the motivation is intuitively supported by qualitative comparisons using different random seeds.

2. Well-Designed Methodology with Two Complementary Training Strategies

The paper introduces two training strategies: end-to-end supervised training (LPNSR-A) and mean-correction training (LPNSR-B). The former improves perceptual quality by directly supervising the final generated results, while the latter derives a residual-based mean correction strategy by minimizing the KL divergence between the model reverse distribution and the forward posterior. These two strategies emphasize perceptual quality and reconstruction fidelity, respectively, providing a theoretically motivated and well-structured framework.

3. Comprehensive Ablation Studies and Detailed Analysis

The paper provides relatively comprehensive ablation studies, including analyses of different loss functions, random seed stability, step-wise KL divergence, and inference efficiency. In addition, the appendix compares deterministic noise prediction with distribution prediction, initialization noise learning with intermediate noise learning, and investigates the effect of residual correction through capacity-matched experiments. These analyses help explain the behavior of the proposed training strategies and strengthen the credibility of the experimental findings.

Critical weaknesses

1. Limited Performance Gains Despite Increased Computational Cost

Although the proposed method achieves improvements across multiple evaluation metrics, the gains are relatively modest, particularly for LPNSR-B. As shown in Table 1, LPNSR-B improves PSNR by only 0.10 dB on LSDIR and 0.20 dB on ImageNet compared to ResShift. However, Table 7 shows that the model parameters increase from 173.92M to 292.51M, while inference latency increases by approximately 36%. Therefore, whether these performance gains justify the additional computational overhead remains questionable.

2. Limited Evaluation Across Datasets

The experimental evaluation is mainly conducted on LSDIR and ImageNet, which limits the evidence for the generalization capability of the proposed method. Additional experiments on representative benchmarks, such as DIV2K and RealSR, would strengthen the evaluation, particularly under different image distributions and real-world degradation scenarios. Since DIV2K is already included in the training data, any additional evaluation should ensure a strict separation between training and testing samples.

Questions for authors

1. Trade-off Between Performance Gains and Computational Cost

According to Table 1, LPNSR-B improves PSNR by only 0.10 dB and 0.20 dB on LSDIR and ImageNet, respectively, compared to ResShift. However, Table 7 shows that the number of parameters and inference latency increase by approximately 68% and 36%, respectively. Could the authors provide performance comparisons under comparable computational budgets to further demonstrate the practical advantages of the proposed method?

2. Generalization Across Different Datasets

The quantitative evaluation is mainly conducted on LSDIR and ImageNet. Have the authors evaluated the proposed method on other representative super-resolution datasets, such as DIV2K and RealSR? If so, could the corresponding results be provided? In particular, evaluation on RealSR would help demonstrate the generalization capability of the proposed method under real-world degradation scenarios. For DIV2K, a strict separation between training and testing data should be ensured.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Dhc4

We thank the reviewer for the constructive comments and we address the questions below point by point.

Response for Q1:

Table 9 in Appendix A.5 provides exactly this comparison under comparable computational budgets. The two models are trained from scratch with comparable capacity (48.19M vs 50.16M) and the same total budget: one is a single network trained directly, the other a small denoising network followed by a small residual predictor. Under this setting the residual-correction scheme improves PSNR by 0.39 dB over the monolithic model, more than the 0.10 dB it gains over Origin at full scale. This indicates that the gain brought by residual correction exceeds that brought by simply increasing the parameter count.

Response for Q2:

We have additionally evaluated the method on two real-world datasets, RealSR and ScreenSR. The quantitative results are reported in Table 11 of Appendix A.7, and additional qualitative comparisons in Appendix A.8. On both datasets, the proposed method still yields clear performance gains.

Response for weakness:

  1. We understand the concern: the inference overhead is not negligible. We would note that the main contribution of this work is a new way of training an intermediate noise predictor, together with evidence of its effectiveness, rather than a latency-optimal design. Table 9 in Appendix A.5 further shows that under matched capacity and an identical total budget, the residual-correction formulation yields a larger gain than it does at full scale (0.39 dB vs 0.10 dB), indicating that the benefit comes from the formulation rather than from the extra parameters. Reducing the overhead by optimizing the capacity and architecture of the noise predictor is an interesting direction for future work.
  2. Further evaluations on real-world datasets, together with visual comparisons, have been added to Appendices A.7 and A.8 of the revised PDF.

The authors have adequately addressed my concerns, and I am willing to maintain my original score.

AuthorsOpenReview ↗

Response to Reviewer Dhc4

Thank you for taking the time to review our work and for your valuable comments. Please feel free to raise any further questions during the discussion period — we would be glad to discuss them.

Reviewer_GHb9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper learns the noise distributions at intermediate sampling steps while keeping the ResShift denoising network frozen. LPNSR-A unrolls the sampling chain and learns the noise mean and variance using losses on the final reconstruction. LPNSR-B corrects the reverse-process mean by predicting the denoising residual while retaining the original variance. The experiments indicate that LPNSR-A emphasizes perceptual quality, whereas LPNSR-B offers more balanced quality improvements.

Critical strengths
  1. The research question is clear, and the two training schemes have well-defined optimization objectives. The connection between mean correction and residual prediction helps explain the method.
  2. Results on two datasets support the effectiveness of the method, and the paper reports the trade-off between perceptual quality and fidelity. Experiments across multiple inference random seeds indicate that the results do not depend on a particular sampling seed.
  3. The ablations cover loss functions, deterministic noise, model capacity, and noise injection locations, and the paper reports actual computational costs. These experiments address relevant design choices.
Critical weaknesses
  1. The novelty is limited. InvSR already predicts Gaussian distribution parameters and explores intermediate noise prediction through InvSR-Int. The paper needs to explain more clearly the distinct contributions of its conditioning inputs and training schemes. LPNSR-B is also mathematically equivalent to residual correction on top of a frozen denoiser, further limiting its methodological novelty.
  2. Table 8 provides evidence in favor of distribution prediction over deterministic noise prediction, but does not disentangle the effects of learned variance, fixed small variance, and stochastic regularization during training. The independent benefit of learning input-dependent variance remains unclear.
  3. The per-sample optimality derivation is sound, but it conditions on the ground-truth image. Its implications for choosing unit noise variance in a practical conditional predictor require further clarification; this derivation alone does not establish the optimality of the complete sampling process.
Questions for authors
  1. Compared with InvSR-Int, what are the respective contributions of conditioning on the current state and denoising prediction, and of training through the full sampling chain? Can controlled experiments clarify these differences?
  2. Does learned variance outperform a fixed small variance or a variance that depends only on the timestep? For the same trained model, what happens if the stochastic term is removed only during inference?
  3. How does LPNSR-B's choice of unit noise variance extend from the per-sample analysis to practical conditional prediction? At t=1t=1, does the final output include residual correction? Please clarify the inference rule.
  4. In the full-scale setting, does the advantage persist over backbone fine-tuning or alternative parameter allocation schemes using the same data and training budget? How is the training budget allocated between the two stages in Table 9?
  5. How does the method perform on real-world datasets such as RealSR and RealLR200? Could the authors provide quantitative evaluations appropriate for each dataset, along with visual results, to assess generalization to unknown real-world degradations?

If the authors adequately address my concerns, I will raise my rating.

6 comments
AuthorsOpenReview ↗

Response to Reviewer GHb9 (cont., Q1-Q3)

We thank the reviewer for the constructive comments and address them point by point below. All the additional experiments reported in this response are conducted on the LSDIR dataset.

Response for Q1:

To isolate the contribution of each input, we ablate the input conditioning of the noise predictor, using the same training configuration and budget as LPNSR-A . The results are given below:

Configurationxtx_tx0′x_0'PSNR↑SSIM↑LPIPS↓DISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
Origin (standard Gaussian)––20.700.52810.22490.142067.47410.43630.6445
w/o both inputs✗✗20.280.52340.22250.140367.75150.46940.6677
w/ xtx_t only✓✗19.980.51840.21310.136770.04380.52940.7029
w/ x0′x_0' only✗✓20.050.52120.21820.138969.41560.50070.6803
LPNSR-A (both inputs)✓✓19.770.51400.20990.133671.79040.56180.7389

Under our training schemes, when both inputs are removed at the same time, the noise predictor still leads to a performance change, but its magnitude is limited. Keeping either xtx_t or x0′x_0' alone leads to a substantially larger change, and the direction is the same as that of LPNSR-A: fidelity decreases while perceptual quality increases.

To further isolate the effect of the training scheme, we design an additional training strategy following InvSR-Int, in which the noise predictor still takes xtx_t, x0′x_0', y0y_0 and tt as input, and is trained under the same training budget:

For each iteration we sample t∼Uniform{2,3,4}t \sim \mathrm{Uniform}\lbrace 2,3,4 \rbrace together with ϵ∼N(0,I)\epsilon \sim \mathcal{N}(0,I), and construct the state xt=x0+ηt(y0−x0)+κηt ϵ.x_t = x_0 + \eta_t(y_0-x_0) + \kappa\sqrt{\eta_t}\thinspace\epsilon .

Given the estimate x0′=fθ(xt,y0,t)x_0' = f_\theta(x_t,y_0,t), the predictor outputs z=gω(xt,x0′,y0,t)z = g_\omega(x_t,x_0',y_0,t), and a single reverse step gives

xt−1=ηt−1ηtxt+αtηtx0′+κηt−1αtηt z.x_{t-1} = \frac{\eta_{t-1}}{\eta_t}x_t + \frac{\alpha_t}{\eta_t}x_0' + \kappa\sqrt{\frac{\eta_{t-1}\alpha_t}{\eta_t}}\thinspace z .

We then re-estimate the clean latent as x^0=fθ(xt−1,y0,t−1)\hat x_0 = f_\theta(x_{t-1},y_0,t-1) and update ω\omega only, with the LPNSR-A loss evaluated on x^0\hat x_0. The results are as follows:

ConfigurationPSNR↑SSIM↑LPIPS↓DISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
Origin (standard Gaussian)20.700.52810.22490.142067.47410.43630.6445
Single-step training (InvSR-Int style)20.250.52030.22070.139369.07180.47160.6738
LPNSR-A (full-chain training)19.770.51400.20990.133671.79040.56180.7389

This scheme does change the performance, but the magnitude is limited compared with the training scheme of LPNSR-A. This is consistent with the result reported for InvSR-Int.

Response for Q2:

To answer this question, on the same trained LPNSR-A model we replace σω\sigma_\omega with the fixed values 00, 0.010.01 and 11, and additionally set the other inputs to zero when the model predicts σω\sigma_\omega, so that the predicted variance depends only on tt. The results are as follows:

ConfigurationPSNR↑SSIM↑LPIPS↓DISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
LPNSR-A19.770.51400.20990.133671.79040.56180.7389
σω=0\sigma_\omega = 019.770.51400.20980.133671.79110.56180.7391
σω=0.01\sigma_\omega = 0.0119.770.51400.20990.133671.79310.56180.7389
σω=1\sigma_\omega = 119.000.48370.22810.158970.06840.40560.6106
only tt as input19.760.51380.20980.133571.79210.56200.7402

Since the predicted σω\sigma_\omega is already small, replacing it with zero or with a fixed small value barely changes the results. Setting it to a large value such as 11, by contrast, degrades all the metrics considerably, which also gives an intuitive explanation for why the predicted σω\sigma_\omega is small. Making the predicted σω\sigma_\omega depend only on tt likewise has no noticeable effect.

Response for Q3:

The per-step KL decomposes into a mean term and a variance term, DKL=12Δ⊤Σθ−1Δ+12[tr(Σω−I)−ln⁡det⁡Σω]D_{KL} = \frac{1}{2}\Delta^{\top}\Sigma_\theta^{-1}\Delta + \frac{1}{2}\left[\mathrm{tr}(\Sigma_\omega-I) - \ln\det\Sigma_\omega\right]. The mean term depends on the sample, since Δ\Delta is the deviation between the posterior mean and the model mean and thus involves x0x_0; the variance term depends only on Σω\Sigma_\omega and is minimised at Σω=I\Sigma_\omega = I for every sample, because ResShift satisfies Σ~t=Σθ\tilde\Sigma_t = \Sigma_\theta at every step. The sample-dependent part of the per-step optimum is therefore confined to the mean, which is exactly what fωf_\omega predicts, while the unit variance holds for any input and is fixed at inference.

At t=1t=1 the residual correction is also included.

AuthorsOpenReview ↗

esponse to Reviewer GHb9 (cont., Q4-Q5)

We continue our response in this second comment and address Q4 and Q5 below.

Response for Q4:

To test whether the same improvement can instead be obtained by adapting the backbone itself, we fine-tune ResShift backbone on the same data with the same training budget, The results are as follows:

MethodPSNR↑SSIM↑LPIPS↓DISTS↓MUSIQ↑MANIQA↑CLIPIQA↑
Origin (frozen backbone)20.700.52810.22490.142067.47410.43630.6445
backbone fine-tuning20.610.52390.22150.141367.73480.43970.6511
LPNSR-A19.770.51400.20990.133671.79040.56180.7389
LPNSR-B20.800.53450.21970.139468.38110.45820.6722

Fine-tuning decreases fidelity slightly and increases perceptual quality slightly, but the change is small: the ResShift backbone is already well trained on natural images, so further fine-tuning can hardly yield a noticeable benefit.

In Table 9 both models are trained with the same total budget of 300k iterations. The monolithic model is trained for 300k iterations in a single stage, whereas the split model is trained for 150k iterations to obtain the denoising network from scratch and 150k iterations to train the residual predictor on top of the frozen network.

Response for Q5:

We have evaluated the proposed method on two real-world datasets, RealSR and ScreenSR. The quantitative results and the corresponding visual comparisons are provided in the updated PDF, in Appendices A.7 and A.8.

Response for weakness:

  1. Our main contribution is two effective training schemes for intermediate noise prediction, whereas the training scheme of InvSR-Int brings no clear gain. LPNSR-B's mean correction is a natural consequence of our derivation: changing the noise sampling distribution amounts, by construction, to changing the mean and variance terms of the reverse process.
  2. Please see our response to Q2, where we separate the effects of the learned variance, a fixed small variance, and the removal of the stochastic term.
  3. The rationale for the unit noise variance is given in our response to Q3. LPNSR-B targets the per-step alignment between the forward posterior and the reverse process; we do not claim global optimality of the sampling process. We only report that this per-step alignment empirically improves the quality of the generated images.

Thank you for the thorough responses and revision. Q1, Q4 and Q5 are well addressed; please include the rebuttal experiments in the final version. A few brief clarifications:

  1. Q2 shows that σω\sigma_\omega has no effect at inference, so LPNSR-A is effectively deterministic and the gain in Table 8 must come from training. What role does the stochastic term play during training, and why is the deterministic-map baseline even worse than Origin on LPIPS and DISTS?

  2. σω=1\sigma_\omega = 1 being optimal for any input holds for KL(p∥q)\mathrm{KL}(p \Vert q), but Eqs. (10)/(15) minimize KL(q∥p)\mathrm{KL}(q \Vert p), whose optimum is σω2=1+E[Δ2∣xt]/Σθ\sigma_\omega^2 = 1 + \mathbb{E}[\Delta^2 \mid x_t] / \Sigma_\theta. Which direction is intended?

  3. At t=1t = 1, Σθ=0\Sigma_\theta = 0, so the final correction x0′+fωx_0' + f_\omega is not injected noise. How much of LPNSR-B's gain comes from this step?

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.