acceptodds
Under review as a conference paper at ICLR 2027

Quantization-Aware Kalman Estimation for Diffusion Sampling

Abstract

Quantization offers a practical path to deploying diffusion models with reduced memory and computation, but aggressive compression can cause quantized outputs to deviate substantially from their full-precision counterparts. Sampling-stage correction methods seek to compensate for such deviations during sampling, but existing approaches rely primarily on local information and underexploit trajectory history, limiting their ability to correct errors that propagate across timesteps. In this work, we formulate sampling with a quantized denoiser as an online estimation problem, using the history of quantized denoiser outputs to recover the underlying full-precision outputs required by the sampler. We propose QuAKE, a Quantization-Aware Kalman Estimator that combines a smooth trajectory prior with a conditional Gaussian observation model. At each sampling step, QuAKE recursively updates the posterior over the output window in closed form and feeds its posterior mean to the sampler. QuAKE is a lightweight plug-and-play corrector that requires no modification to the quantized network and naturally supports arbitrary high-order multistep ODE samplers. Experiments across W4A4-quantized text-to-image diffusion models show that QuAKE consistently outperforms existing methods in reducing the distributional discrepancy from full-precision sampling.

Primary area:
generative models
Keywords:
Diffusion Models, Model Quantization, Online Estimation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FrkmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes QuAKE, a sampling-stage correction method for quantized diffusion models. QuAKE regards the full-precision (FP) output window of a multistep solver as a latent state and the quantized outputs as its noisy observations. At each step, it combines a prediction extrapolated from previous estimates with the current quantized output through a Kalman filter, and supplies the posterior mean to the solver. All statistics are calibrated per timestep and channel on FP trajectories. The authors report that QuAKE achieves lower FID and KID than existing correction methods on four W4A4 text-to-image models with both DPM-Solver++ and UniPC.

Critical strengths
  1. The motivation is clear. Existing methods correct the current output only, while multistep solvers use previous outputs as well.
  2. The method is lightweight and solver-agnostic. It requires no modification to the quantized network and adds about 2 ms per step on FLUX.1-dev (Fig. 4).
Critical weaknesses
  1. QuAKE relies on two assumptions that are not validated. First, all statistics are shared across samples and spatial positions (Appendix C). The Kalman gains are thus fixed before sampling, and every prompt receives the same weighting between the prediction and the quantized output. The smoothness of trajectories, however, is likely to vary across prompts, as shown in [A]. For a prompt with a less smooth trajectory, the calibrated QQ underestimates the prediction error, and the filter trusts the prediction more than it should. Second, the quantization noise is assumed to be independent across steps (Sec. 3.4). Since the same quantized weights process similar inputs at consecutive steps, the noise is likely to be correlated over time. The filter would then treat a repeated error as new evidence, becoming overconfident in a biased estimate. Discussing these issues would be beneficial.

  2. The robustness to fewer sampling steps is not examined. All experiments use 20 sampling steps (Sec. 4.1). QuAKE predicts the current output by extrapolating past outputs, and this prediction becomes less accurate as the gap between steps grows. With fewer steps, QuAKE would therefore gain less from past outputs. In addition, the first few steps have too few past outputs for full-order extrapolation (Appendix D), and these steps make up a larger part of a short trajectory. It would be helpful to report results with 10 and 5 steps, and to show whether the gain over Bias Correction persists.

[A] DiCache: Let Diffusion Model Determine Its Own Cache, ICLR 2026.

Questions for authors
  1. The statistics are estimated on FP trajectories (Appendix C), but QuAKE applies them along corrected quantized trajectories. How large is this mismatch, and how does it affect the filter?

  2. ImageReward drops below the uncorrected model in several cases (e.g., from 0.9131 to 0.8934 on PixArt-α\alpha, Table 1). How do the authors reconcile this with the claim that generation quality is preserved (Ln405-406)?

  3. QuAKE uses past observations and also corrects the past outputs in the window. How much does each component contribute? An ablation that uses past observations but corrects the current output only would answer this, and would also test the claim in Ln410-419 that the window inconsistency degrades the performance of Bias Correction.

  4. Some margins are small (e.g., FID on FLUX.1-dev, Table 1), and reporting the variance over multiple seeds would make the comparison more convincing.

  5. All models are quantized with SVDQuant only. Results with another quantizer would show whether QuAKE depends on the error characteristics of a specific method.

6 comments
AuthorsOpenReview ↗

[Part 1/3] Response to Reviewer Frkm

We thank the reviewer for the thoughtful feedback.

1. Validation of the modeling assumptions (W1). We provide additional evidence supporting our modeling assumptions.

(1) Robustness to process noise variation. We examine the effect of using a shared process noise covariance QiQ_i across different prompts. Because different prompts can induce trajectories with different degrees of smoothness and thus different process noise variances, the variance along an individual sampling trajectory may differ from the calibrated average. To test the robustness of QuAKE to this mismatch, we deliberately introduce a mismatch by scaling the calibrated covariance QiQ_i by a factor τ\tau as follows:

si=Aisi−1+bi+Gηi,ηi∼N(0,τQi).s_i = A_i s_{i-1} + b_i + G\eta_i,\qquad \eta_i \sim \mathcal{N}(0,\tau Q_i).

We report the experimental results below, where the tested range of τ\tau covers most of the empirically observed prompt-wise variation, as shown in the revised appendix.

τ\tauFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
0.57.030.1660.31830.90460.8836
17.020.1580.31840.90420.8822
26.980.1470.31810.90340.8811

According to the results, QuAKE remains stable across this range, showing that moderate discrepancies between the calibrated and actual process noise variances have little effect on performance. Given the simplicity and practicality of calibration, using a shared covariance estimated across prompts is a reasonable choice.

(2) Temporal independence of quantization noise. In Appendix F.3 of the revised manuscript, we added experiments examining the temporal correlation of quantization noise. The measured correlations across sampling steps are generally weak, supporting our independence assumption. Although the inputs at consecutive steps may be similar, different timestep embeddings and complex nonlinear transformations within the denoiser can yield weakly correlated quantization errors at the output.

AuthorsOpenReview ↗

[Part 2/3] Response to Reviewer Frkm

2. Robustness to fewer sampling steps (W2). We added an ablation on the number of sampling steps. QuAKE maintains its advantage at both 10 and 5 steps. These results show that explicitly using historical denoiser outputs remains beneficial for estimating the underlying full-precision outputs even with fewer sampling steps. We report the results below and in the revised appendix.

10 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.33620.88651.0249
Quantized10.941.4910.33510.88210.9729
Bias Correction19.155.5700.33290.87120.9131
TAC-Diffusion18.815.4720.33270.87170.9132
Q-Drift10.341.1430.33470.87650.9549
QuAKE (Ours)8.730.4630.33520.88030.9664

5 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.32680.82540.5704
Quantized12.641.9080.32580.81110.5257
Bias Correction16.513.6510.32570.80890.5234
TAC-Diffusion16.433.6420.32580.81010.5249
Q-Drift12.111.6380.32530.80180.5064
QuAKE (Ours)11.471.3690.32460.79640.5023

3. Calibration trajectory mismatch (Q1). We added an experiment in which we recalibrated the statistics along QuAKE-corrected quantized trajectories. The results show that the trajectory mismatch has little effect on correction performance. We report the results below and in the revised appendix.

Calibration TrajectoryFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision Trajectory7.900.2890.33630.89511.0496
Corrected Trajectory7.880.2940.33620.89601.0518

4. Slight decrease in ImageReward (Q2). Our claim in Ln405-406 of the original manuscript is that QuAKE is better at "preserving the behavior of the full-precision model," as reflected in FID and KID. Preserving the full-precision output distribution and maintaining generation quality are distinct objectives, since quantization-induced shifts can change the generated content without uniformly degrading its quality, as illustrated by the qualitative results in the paper. Our claim of "without compromising overall generation quality" is supported by the quality metrics taken together. Given the relatively wide dynamic range of ImageReward, the decrease of about 0.02 is small, with little perceptible difference in visual quality. We therefore consider our claim that QuAKE maintains overall generation quality to be justified.

AuthorsOpenReview ↗

[Part 3/3] Response to Reviewer Frkm

5. Ablation of historical output correction (Q3). We added an ablation that applies correction only to the current output. This variant performs worse than full-window correction, demonstrating the contribution of correcting historical outputs. We report the results below and in the revised appendix.

Correction StrategyFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Current-Only Correction8.340.4560.33620.89501.0455
Full-Window Correction7.900.2890.33630.89511.0496

6. Results over multiple seeds (Q4). Given the available time and compute, we repeated the FLUX.1-dev experiment on MJHQ-30K using the DPM-Solver++ sampler with five random seeds. We report the mean and standard deviation for each metric, followed by the FID for each run. With the limited multi-seed evaluation, we believe that the existing results across models, datasets, and samplers already provide sufficient evidence that QuAKE's gains are attributable to the method itself rather than random variation across runs.

Results across five random seeds:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.3193 ±\pm 0.00010.8992 ±\pm 0.00050.9010 ±\pm 0.0064
Quantized7.26 ±\pm 0.050.268 ±\pm 0.0180.3180 ±\pm 0.00010.9058 ±\pm 0.00020.8901 ±\pm 0.0073
Bias Correction7.23 ±\pm 0.040.275 ±\pm 0.0200.3180 ±\pm 0.00020.9045 ±\pm 0.00040.8859 ±\pm 0.0065
TAC-Diffusion7.27 ±\pm 0.030.292 ±\pm 0.0150.3180 ±\pm 0.00020.9053 ±\pm 0.00030.8915 ±\pm 0.0101
Q-Drift7.11 ±\pm 0.070.201 ±\pm 0.0070.3181 ±\pm 0.00020.9051 ±\pm 0.00020.8875 ±\pm 0.0061
QuAKE (Ours)6.96 ±\pm 0.050.158 ±\pm 0.0100.3182 ±\pm 0.00020.9046 ±\pm 0.00030.8806 ±\pm 0.0077

FID for each of the five runs:

RunQuantizedBias CorrectionTAC-DiffusionQ-DriftQuAKE (Ours)
17.307.247.317.167.02
27.287.217.297.156.95
37.277.297.267.157.00
47.167.187.267.016.90
57.267.227.247.096.94

7. Results with another quantizer (Q5). We added experiments with OrbitQuant [1], and QuAKE maintained its advantage. We report the results below and in the revised appendix.

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.33750.90031.0990
Quantized10.410.4170.33440.88080.9539
Bias Correction17.203.9070.33200.88020.9170
TAC-Diffusion16.973.7640.33200.87890.9208
Q-Drift10.250.3440.33440.88820.9570
QuAKE (Ours)10.010.2670.33420.88900.9569

References

[1] OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers, arXiv 2026.

Reviewer_r2adOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies inference-stage correction of quantization error in diffusion models and proposes QuAKE, which uses a Kalman filter to estimate the full-precision denoiser outputs. Such improved estimation enhances generated image quality. In addition, the proposed method can smoothly work with existing high order solvers. Experimental results show that the method effectively improves the generated image quality.

Critical strengths

With the assumption of local smoothness, this paper uses a Kalman filter to leverage the history of quantized output to estimate the full-precision outputs. The proposed QuAKE shows a way to leverage information in historical observations to reduce the prediction variance. This idea is novel and supported by experiments. Additionally, the method is easy to use and therefore able to serve as a plug-and-play correction for existing solvers.

Critical weaknesses

The proposed estimation (extrapolation) depends on the historical outputs. Therefore, the amount of history should affect the performance. However, the current experiments are done only with a 20-step setup with no history length ablation. Also, it is unknown whether the method will work under a few-step setup such as 4-step sampling using SDXL-Turbo. These experiments are necessary to show where this method applies. It’s fine even if the method does not achieve the best performance under those settings, but such analysis is necessary for users as guidance.

Questions for authors
  • While the derivations are correct, the readability can be further improved. The dimensions of most notation are never provided, and some of the notation is used without introduction (m_{i|j}, P_{i|j}).
  • The idea of this work is simply to use a Kalman filter to get a better estimation. However, the current writing kind of makes it hard to follow. Like making assumptions (smooth trajectory prior, Gaussian observation) without clear motivation. I would suggest providing an introduction to the Kalman filter in the preliminary section to guide the reader. I believe this would greatly improve the readability.
  • Appendix A.2 states TCEC [a] as a DDIM-specific method, while their paper actually provides experiments for higher-order solvers. A comparison should be added.

[a] Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion Models, ICML 2026.

  • One of the advantages of the proposed method is the flexibility to work with arbitrary high order solvers. However, current experiments are limited to second-order solvers. Since first-order samplers are still widely used, such as DDIM or the Euler sampler, experiments of first order should also be provided. This setup should also enable comparisons with other training-free methods such as [b, c, d].

[b] Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps, ICLR 2024.

[c] QNCD: Quantization Noise Correction for Diffusion Models, ACM MM 2024.

[d] StepbaQ: Stepping backward as Correction for Quantized Diffusion Models, NeurIPS 2024.

4 comments
AuthorsOpenReview ↗

[Part 1/2] Response to Reviewer r2ad

We thank the reviewer for the valuable feedback.

1. Ablation of history length (W1). We added experiments with 30, 20, 10, and 5 sampling steps, and QuAKE maintained its advantage. We report the results below and in the revised appendix.

30 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.33740.90241.1045
Quantized9.721.2200.33620.89571.0589
Bias Correction22.377.7540.33210.87610.9585
TAC-Diffusion22.157.5880.33190.87670.9632
Q-Drift9.280.9550.33610.89121.0473
QuAKE (Ours)7.740.2600.33640.89671.0602

20 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.33750.90031.0990
Quantized10.231.3940.33600.89361.0481
Bias Correction22.507.7550.33210.87640.9522
TAC-Diffusion22.367.6390.33200.87660.9576
Q-Drift9.731.0720.33590.88901.0356
QuAKE (Ours)7.900.2890.33630.89511.0496

10 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.33620.88651.0249
Quantized10.941.4910.33510.88210.9729
Bias Correction19.155.5700.33290.87120.9131
TAC-Diffusion18.815.4720.33270.87170.9132
Q-Drift10.341.1430.33470.87650.9549
QuAKE (Ours)8.730.4630.33520.88030.9664

5 steps:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.32680.82540.5704
Quantized12.641.9080.32580.81110.5257
Bias Correction16.513.6510.32570.80890.5234
TAC-Diffusion16.433.6420.32580.81010.5249
Q-Drift12.111.6380.32530.80180.5064
QuAKE (Ours)11.471.3690.32460.79640.5023

2. Few-step distilled models (W1). We added experiments with 4-step SDXL-Turbo [1]. The results suggest that explicitly using historical denoiser outputs helps estimate the underlying full-precision outputs even in few-step models. We report the results below and in the revised appendix.

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.32930.84910.6605
Quantized10.241.8930.32880.80530.5540
Bias Correction8.680.9480.32980.82580.6210
TAC-Diffusion9.371.3020.33060.81870.5976
Q-Drift9.641.4970.32900.81200.5624
QuAKE (Ours)8.350.7880.33010.84480.6482

3. Dimensions and notation (Q1). As mentioned in Section 3.1, following common practice in prior work, we perform the estimation element-wise, so fif_i is treated as a scalar in the derivation. The subscript i∣ji\mid j is standard notation in filtering theory and denotes estimation of a variable at step ii using information available through step jj. We use this notation for consistency, and each quantity of the form mi∣jm_{i\mid j} or Pi∣jP_{i\mid j} can be read as a separate variable. The corresponding definitions are given when these quantities are introduced, as in equations (20) and (24) of the revised manuscript (which correspond to equations (19) and (23) of the original submission).

AuthorsOpenReview ↗

[Part 2/2] Response to Reviewer r2ad

4. Kalman filtering preliminaries (Q2). Following the reviewer’s suggestion, we added a brief introduction to Kalman filtering in Section 2.3 to motivate the smooth trajectory prior and conditional Gaussian observation model used in our theoretical formulation.

5. Reasons for excluding TCEC (Q3). Although TCEC [2] briefly discusses an extension to higher-order samplers in its appendix, we excluded TCEC from our experimental comparisons mainly for two reasons. First, compared with the more commonly used multistep second-order sampler, the single-step second-order sampler they use requires twice as many network evaluations, reaching 40 evaluations for the 20 sampling steps used in our experiments. A direct comparison with the other methods would be computationally unfair. Second, to the best of our knowledge, no public implementation is available, and the paper omits key implementation details, including the values of important hyperparameters such as rr and λ\lambda and how the Frobenius norm of the Jacobian is estimated. This makes it difficult to reproduce the method reliably.

6. First-order samplers (Q4). We added experiments with the DDIM sampler, including the suggested baselines, and QuAKE maintained its advantage. We report the results below and in the revised appendix.

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.33770.90191.0716
Quantized9.381.0700.33640.89551.0242
Bias Correction19.135.9540.33330.88180.9508
TAC-Diffusion18.965.7560.33310.88110.9525
Q-Drift8.740.7380.33670.89491.0236
TS-DDIM [3]9.730.8600.33600.88650.9860
QNCD [4]9.351.0490.33660.89541.0267
StepbaQ [5]8.050.4430.33680.89361.0185
QuAKE (Ours)7.600.2430.33670.89731.0308

References

[1] Adversarial diffusion distillation, ECCV 2024.

[2] Error Propagation Mechanisms and Compensation Strategies for Quantized Diffusion Models, ICML 2026.

[3] Alleviating Exposure Bias in Diffusion Models through Sampling with Shifted Time Steps, ICLR 2024.

[4] QNCD: Quantization Noise Correction for Diffusion Models, ACM MM 2024.

[5] StepbaQ: Stepping backward as Correction for Quantized Diffusion Models, NeurIPS 2024.

Thank you for the detailed response and the additional experiments, which address my concerns. I will raise my score.

Out of curiosity, it would be interesting to see how StepbaQ performs in the SDXL-Turbo experiment, since it is one of the strongest baselines in the DDIM table and also reports results on SDXL-Turbo in its original paper. This is only an optional suggestion.

Thank you for your positive feedback on our work. We evaluated StepbaQ on SDXL-Turbo under our experimental settings and found that it helps mitigate quantization errors, while our method still performs better overall. We report the updated results below and have included them in the latest revision of the manuscript.

Updated SDXL-Turbo results:

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision––0.32930.84910.6605
Quantized10.241.8930.32880.80530.5540
Bias Correction8.680.9480.32980.82580.6210
TAC-Diffusion9.371.3020.33060.81870.5976
Q-Drift9.641.4970.32900.81200.5624
StepbaQ8.781.0260.32930.81940.5964
QuAKE (Ours)8.350.7880.33010.84480.6482
Reviewer_h6LNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes QuAKE, a Kalman filtering-based method for correcting quantization errors during diffusion sampling. By combining a smooth trajectory prior with a conditional Gaussian observation model, QuAKE estimates full-precision denoiser outputs using historical quantized observations. Experiments demonstrate consistent improvements in distributional similarity with negligible computational overhead.

Critical strengths
  • Practical and efficient: QuAKE requires no additional training or network modifications and introduces negligible computational overhead.

  • Consistent experimental improvements: The method achieves superior FID and KID results across multiple diffusion models and datasets.

Critical weaknesses
  • Limited methodological novelty: The proposed method primarily applies conventional Kalman filtering to diffusion sampling, and its novelty beyond the state-space formulation is somewhat limited.

  • Insufficient validation of modeling assumptions: The linear transition, Gaussian observation, and temporal independence assumptions are not thoroughly validated across different models and sampling steps.

  • Limited evaluation beyond fast samplers: The experiments are restricted to fast samplers such as DPM-Solver++ and UniPC. It would be interesting to investigate whether QuAKE can also provide performance improvements when applied to conventional, non-fast samplers such as DDIM etc.

Questions for authors
  • I believe that validating the modeling assumptions is one of the most important aspects of this work. Could the authors provide a more detailed justification and empirical validation of these assumptions? (Linear transition, gaussian observation, temporal independence, local smoothness...)

  • Could the authors discuss whether the proposed method can be extended to stochastic samplers? Exploring this possibility would further strengthen the paper's contribution and broaden its applicability.

3 comments
AuthorsOpenReview ↗

Response to Reviewer h6LN

We thank the reviewer for the constructive feedback.

1. Methodological novelty (W1). Our novelty lies in being the first to formulate sampling-stage correction for quantized diffusion models as a history-conditioned online state-estimation problem and in proposing a correction method applicable to any high-order multistep ODE sampler, a common default for many modern diffusion models. We use standard Kalman filtering as a tool to solve this estimation problem, and we do not claim novelty in filtering theory itself.

2. Validation of the modeling assumptions (W2/Q1). The local smoothness assumption is consistent with the Taylor expansion assumption widely used in high-order samplers (as discussed in Appendix A.1), while the conditional Gaussian observation model is based on phenomena commonly observed in prior works in this domain [1, 2]. We provide additional empirical evidence for our modeling assumptions in the revised appendix. Experiments across models and sampling steps support the assumptions used in our theoretical formulation.

3. Evaluation beyond fast samplers (W3). We added experiments with the DDIM sampler, and QuAKE maintained its advantage. We report the results below and in the revised appendix.

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.33770.90191.0716
Quantized9.381.0700.33640.89551.0242
Bias Correction19.135.9540.33330.88180.9508
TAC-Diffusion18.965.7560.33310.88110.9525
Q-Drift8.740.7380.33670.89491.0236
TS-DDIM9.730.8600.33600.88650.9860
QNCD9.351.0490.33660.89541.0267
StepbaQ8.050.4430.33680.89361.0185
QuAKE (Ours)7.600.2430.33670.89731.0308

4. Extension to stochastic samplers (Q2). Our formulation applies to sampler updates of the form xi+1=Ψi(xi;si)x_{i+1}=\Psi_i\left(x_i;s_i\right), where xix_i is the current sample state and sis_i contains the current denoiser output and any past outputs used by the sampler. QuAKE can be used with any stochastic sampler whose update takes this form. We added experiments with the stochastic SDE-DPM-Solver++ sampler [3], and QuAKE maintained its advantage. We report the results below and in the revised appendix.

MethodFID ↓\downarrowKID (×103\times 10^3) ↓\downarrowCLIPScore ↑\uparrowCLIP-IQA ↑\uparrowImageReward ↑\uparrow
Full-Precision--0.33560.89911.0691
Quantized15.483.4220.33280.88610.9749
Bias Correction37.2916.7290.32510.85190.7535
TAC-Diffusion36.9816.3920.32520.84880.7538
Q-Drift14.773.0830.33270.87630.9526
QuAKE (Ours)10.410.8210.33430.88961.0048

References

[1] D2-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models, AAAI 2025.

[2] Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling, arXiv 2026.

[3] DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models, MIR 2025.

As I understand it, the linear transition and Gaussian assumptions are approximate modeling choices, and the additional analyses provide empirical support for their validity. Although these assumptions are not guaranteed to hold exactly, the experimental results demonstrate the practical effectiveness of the resulting correction method. Similarly, while temporal independence has not been established, the observed weak temporal correlations provide empirical justification for using the standard Kalman filtering approximation.

I would also like to bring the concurrent work [1] to the authors’ attention. I suggest citing it and briefly discussing it in the related work section of the revised manuscript.

Overall, my concerns have been sufficiently addressed, and I will raise my score to 3.

[1] DiFA: Inference-Time Forward-Process Alignment for Diffusion Models

We thank the reviewer for the positive assessment of our work. We have added a brief discussion of the relationship between the concurrent work DiFA and our approach in the Related Work section of the latest revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their thoughtful and constructive feedback. We have revised the manuscript in response to their comments. The main changes are as follows.

1. Kalman filtering preliminaries (Section 2.3). We added a brief introduction to Kalman filtering to motivate the smooth trajectory prior and conditional Gaussian observation model used in our theoretical formulation.

2. Empirical support for the modeling assumptions (Section 3 and Appendix F). We provide additional empirical evidence supporting the local smoothness of denoiser trajectories, the conditional Gaussian observation model, and the assumptions that process and observation noise are each independent across sampling steps and are mutually independent.

3. Related work (Appendix A). We expanded the related work discussion to include recent studies on leveraging historical denoiser predictions and clarify their relationship to our approach.

4. Additional experiments and analyses (Appendix G). In response to the reviewers’ requests, we added experiments with DDIM and SDE samplers, different numbers of sampling steps, an additional quantization method, and a few-step distilled model. We also added an ablation of historical denoiser output correction, an analysis of matched-trajectory calibration, and a study of our method's robustness to the process noise variance.

All changes are marked in blue in the revised manuscript. We hope these revisions address the reviewers’ comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.