- 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
- The motivation is clear. Existing methods correct the current output only, while multistep solvers use previous outputs as well.
- 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
-
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 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.
-
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
-
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?
-
ImageReward drops below the uncorrected model in several cases (e.g., from 0.9131 to 0.8934 on PixArt-, Table 1). How do the authors reconcile this with the claim that generation quality is preserved (Ln405-406)?
-
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.
-
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.
-
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
[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:
| Method | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Full-Precision | - | - | 0.3362 | 0.8865 | 1.0249 |
| Quantized | 10.94 | 1.491 | 0.3351 | 0.8821 | 0.9729 |
| Bias Correction | 19.15 | 5.570 | 0.3329 | 0.8712 | 0.9131 |
| TAC-Diffusion | 18.81 | 5.472 | 0.3327 | 0.8717 | 0.9132 |
| Q-Drift | 10.34 | 1.143 | 0.3347 | 0.8765 | 0.9549 |
| QuAKE (Ours) | 8.73 | 0.463 | 0.3352 | 0.8803 | 0.9664 |
5 steps:
| Method | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Full-Precision | - | - | 0.3268 | 0.8254 | 0.5704 |
| Quantized | 12.64 | 1.908 | 0.3258 | 0.8111 | 0.5257 |
| Bias Correction | 16.51 | 3.651 | 0.3257 | 0.8089 | 0.5234 |
| TAC-Diffusion | 16.43 | 3.642 | 0.3258 | 0.8101 | 0.5249 |
| Q-Drift | 12.11 | 1.638 | 0.3253 | 0.8018 | 0.5064 |
| QuAKE (Ours) | 11.47 | 1.369 | 0.3246 | 0.7964 | 0.5023 |
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 Trajectory | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Full-Precision Trajectory | 7.90 | 0.289 | 0.3363 | 0.8951 | 1.0496 |
| Corrected Trajectory | 7.88 | 0.294 | 0.3362 | 0.8960 | 1.0518 |
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.
[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 Strategy | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Current-Only Correction | 8.34 | 0.456 | 0.3362 | 0.8950 | 1.0455 |
| Full-Window Correction | 7.90 | 0.289 | 0.3363 | 0.8951 | 1.0496 |
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:
| Method | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Full-Precision | - | - | 0.3193 0.0001 | 0.8992 0.0005 | 0.9010 0.0064 |
| Quantized | 7.26 0.05 | 0.268 0.018 | 0.3180 0.0001 | 0.9058 0.0002 | 0.8901 0.0073 |
| Bias Correction | 7.23 0.04 | 0.275 0.020 | 0.3180 0.0002 | 0.9045 0.0004 | 0.8859 0.0065 |
| TAC-Diffusion | 7.27 0.03 | 0.292 0.015 | 0.3180 0.0002 | 0.9053 0.0003 | 0.8915 0.0101 |
| Q-Drift | 7.11 0.07 | 0.201 0.007 | 0.3181 0.0002 | 0.9051 0.0002 | 0.8875 0.0061 |
| QuAKE (Ours) | 6.96 0.05 | 0.158 0.010 | 0.3182 0.0002 | 0.9046 0.0003 | 0.8806 0.0077 |
FID for each of the five runs:
| Run | Quantized | Bias Correction | TAC-Diffusion | Q-Drift | QuAKE (Ours) |
|---|---|---|---|---|---|
| 1 | 7.30 | 7.24 | 7.31 | 7.16 | 7.02 |
| 2 | 7.28 | 7.21 | 7.29 | 7.15 | 6.95 |
| 3 | 7.27 | 7.29 | 7.26 | 7.15 | 7.00 |
| 4 | 7.16 | 7.18 | 7.26 | 7.01 | 6.90 |
| 5 | 7.26 | 7.22 | 7.24 | 7.09 | 6.94 |
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.
| Method | FID | KID () | CLIPScore | CLIP-IQA | ImageReward |
|---|---|---|---|---|---|
| Full-Precision | - | - | 0.3375 | 0.9003 | 1.0990 |
| Quantized | 10.41 | 0.417 | 0.3344 | 0.8808 | 0.9539 |
| Bias Correction | 17.20 | 3.907 | 0.3320 | 0.8802 | 0.9170 |
| TAC-Diffusion | 16.97 | 3.764 | 0.3320 | 0.8789 | 0.9208 |
| Q-Drift | 10.25 | 0.344 | 0.3344 | 0.8882 | 0.9570 |
| QuAKE (Ours) | 10.01 | 0.267 | 0.3342 | 0.8890 | 0.9569 |
References
[1] OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers, arXiv 2026.
[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 Qi 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 Qi by a factor τ as follows:
si=Aisi−1+bi+Gηi,ηi∼N(0,τQi).
We report the experimental results below, where the tested range of τ covers most of the empirically observed prompt-wise variation, as shown in the revised appendix.
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.