acceptodds
Under review as a conference paper at ICLR 2027

Q-Drift: Quantization-Aware Drift Correction for Diffusion Model Sampling

Abstract

Post-training quantization (PTQ) is a practical path to deploy large diffusion models, but quantization noise can accumulate over the denoising trajectory and degrade generation quality. We propose Q-Drift, a principled sampler-side correction that treats quantization error as an implicit stochastic perturbation on each denoising step and applies the paired drift adjustment prescribed by generalized marginal-preserving SDEs. Unlike prior sampler-side corrections, which aim to cancel the quantization error of each individual sample, Q-Drift aims to preserve the distribution. Q-Drift estimates a timestep-wise variance statistic from calibration, in practice requiring as few as 10 paired full-precision/quantized calibration runs. The resulting sampler correction is plug-and-play with common samplers, diffusion models, and PTQ methods, while incurring negligible overhead at inference. Across six diverse text-to-image models (spanning DiT and U-Net), three samplers (Euler, flow-matching, DPM-Solver++), and two PTQ methods (SVDQuant, MixDQ), Q-Drift improves FID over the corresponding quantized baseline in most settings, with up to 4.79 FID reduction on PixArt- (SVDQuant W3A4), while preserving CLIP scores. Code is available at https://anonymous.4open.science/r/Q-Drift.

Primary area:
generative models
Keywords:
diffusion model, quantization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fdeFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Q-Drift, a training-free fix for sampling with quantized diffusion models. During a calibration phase, it runs the full-precision and quantized models side by side and estimates the variance of the quantization error at each step. At sampling time, this variance is used to rescale the update. Model weights are not touched. Experiments cover six models and three samplers. FID goes down in all seven, and CLIP score stays about the same. Under milder quantization, the results are mixed.

Critical strengths

1.The method is simple and cheap. After calibration it only needs one scalar per step, so it is easy to drop into an existing quantized pipeline. I also liked the SDXL check with only 10 calibration runs: even in the three stress cases the authors picked, the result stays close to the 1,000-run version (Table 3). This suggests calibration cost is not a real concern.

2.The gains are not just one FID number. The method is tested on both U-Net and transformer backbones, e.g. PixArt-Sigma goes from 48.85 to 44.06 FID and Sana from 17.32 to 16.54 (Table 1). Appendix E adds KID and paired bootstrap intervals. Six of the seven intervals support an improvement, and the authors are upfront that the small FLUX.1-dev gain is not significant.

Critical weaknesses

1.My main concern is the gap between the theory and what is actually implemented. In Section 4.2, the correction strength is derived by treating the quantization error as independent, zero-mean Gaussian noise. But the actual update keeps the real error, does not subtract its mean, and scales it together with the prediction. So the error is neither zero-mean nor independent across steps in practice. Matching the variance of an idealized noise model does not show that this update brings the sampling distribution closer to the full-precision one. Appendix D looks at spatial statistics, but not the correlation across steps, which is what the SDE argument depends on. Right now I see this as a reasonable heuristic motivated by the SDE view, not a correction with a marginal-preserving guarantee. The paper should say this more clearly.

2.The experiments don't show that this specific scaling rule is what matters. Table 3 compares different variance estimates, but there is no simple baseline such as a constant multiplier or a tuned per-step schedule. Without that, it is hard to know whether the gain comes from the variance-based rule or just from damping the update a bit. The comparison with earlier correction methods is only on SDXL/Euler, where some parts of those methods can't be used anyway. Also, in Table 5, four of the six SVDQuant W4A4 settings get slightly worse FID after correction. So it is not clear when Q-Drift should be preferred over a simpler adjustment.

Questions for authors

1.Along the corrected trajectories, how large is the mean of the quantization error, and how correlated is it across steps? If the mean is not removed, why should the residual variance alone decide the correction strength?

2.Could you compare against a constant multiplier and a simple step-dependent multiplier, both tuned on a separate validation set and run with the same number of steps? I think this is the key experiment for judging the contribution.

3.Appendix C.1 calls the flow-matching version an empirical extension. Which of the marginal-preservation arguments are meant to hold for that version, and for the multistep solver?

2 comments

Thank you for the constructive review, and for recognizing the simplicity of Q-Drift, the robustness of calibration with only 10 runs, and the statistical care of our evaluation. We address your questions below.

W1/Q1. Error mean and cross-step correlation. Following the reviewer's suggestion, we measured them along the trajectories (FLUX.1-schnell W3A4, 1,000 paired trajectories):

Step1234
|mean| / std0.0100.0380.0320.019
Var(ε_q) / Var(ε_FP)0.780.980.970.92
corr(Δε, ε_FP)−0.47−0.16−0.18−0.28

The mean is small (1–4% of the std), which is why the residual variance sets the correction strength. The error is, however, correlated across steps (0.17–0.51; PixArt-Σ 0.26–0.39). The quantized prediction also carries less variance than FP and is anti-correlated with its error, so each Euler step is too short. We agree that Q-Drift is a heuristic motivated by the SDE view, not a correction with a marginal-preservation guarantee, and the paper will state this clearly.

W2/Q2. Constant and step-dependent multipliers tuned on a validation set. We ran the key experiment as specified, tuning on 1,000 validation prompts and evaluating on held-out test prompts (5,000 for FLUX, 1,000 for SDXL-Turbo) with the same steps and seeds:

FLUX.1-schnell W3A4Tuned by FIDTest FID ↓Test ImageReward ↑
Quantized–23.310.830
Q-Drift (λ = 1)no22.310.825
Constant multiplier (c = 0.20)yes21.170.782
Step-dependent multiplieryes22.310.811
SDXL-Turbo W3A4 (1k test)
Quantized–59.310.819
Q-Drift (λ = 1)no57.400.813
Constant multiplier (c = 0.066)yes53.990.747
Step-dependent multiplieryes54.010.758

A constant tuned by FID reaches a lower FID, but it over-lengthens the step and loses perceptual quality (ImageReward −0.05 on FLUX, −0.07 on SDXL-Turbo). Tuning the constant by ImageReward instead does not help: on SDXL-Turbo validation, every constant (1–16×) scores below the uncorrected model (ImageReward 0.655–0.812 vs 0.823), and the best one (1×: FID 58.69, ImageReward 0.812) is worse than Q-Drift on both metrics (58.38, 0.824). Q-Drift improves FID with ImageReward essentially unchanged and needs no FID search; the tuned constant also changes about 20× across models (0.20 on FLUX, 0.066 on SDXL-Turbo, 0.02 on SDXL, 0.01 on Sana). On FLUX, the step-dependent multiplier matches Q-Drift's FID with lower ImageReward. Note that the gain comes from lengthening the step, the opposite of damping: shrinking it (c* < 0) worsens FID (54.29 vs 52.90). We will add these baselines to the main table and move Table 5 into the main text.

Q3. Scope of the arguments.

  • VE Euler: the continuous-time drift–diffusion pairing with per-step variance matching.
  • Rectified flow: the score is affine in the velocity, so the same pairing exists, but our VE-form coefficient is an approximation (hence "empirical extension").
  • DPM-Solver++: the coefficient is local and ignores the correlation between increments.

Given the measured correlations, none of these is a guarantee, and we will state this explicitly.

AuthorsOpenReview ↗

Kind request to check our response

Dear Reviewer, we ran the comparison with a tuned constant multiplier, which you called the key experiment. We would be grateful if you could check our response and consider reconsidering your score.

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

This paper proposes Q-Drift, a sampler-side correction method for post-training quantized diffusion models that aims to mitigate the distributional mismatch caused by accumulated quantization errors during sampling. The authors model quantization errors using their conditional residual variance and interpret these errors as additional stochastic perturbations in the sampling process. Based on this interpretation, Q-Drift derives a timestep-dependent drift correction coefficient by matching the residual variance with the diffusion term of a generalized SDE. The resulting method rescales the sampling update without requiring model retraining or weight modification. Experimental results across six text-to-image models, multiple samplers, and different quantization methods demonstrate improvements in FID, with additional ablation studies examining calibration size, variance estimation, and correction granularity.

Critical strengths

(+) Simple and practical method: The proposed method is simple and easy to apply to existing quantized diffusion models. Q-Drift only introduces a timestep-dependent scaling factor to correct the sampling update, without requiring retraining, weight modification, or additional model evaluations. This makes the method practical and compatible with existing quantization pipelines.

(+) Comprehensive experimental evaluation: The paper provides extensive experiments across six diffusion models, three samplers, and two PTQ methods, demonstrating consistent FID improvements under aggressive quantization settings. In addition, the authors provide ablation studies on calibration size, variance estimation, and correction granularity, as well as statistical analyses to support the effectiveness of the proposed method.

Critical weaknesses

(-) Insufficient justification for the Gaussian perturbation assumption: The paper approximates residual quantization errors as Gaussian perturbations (L221–222) and uses their variance to derive the drift correction coefficient. However, it is unclear whether these errors can be treated as stochastic perturbations equivalent to the diffusion term in the generalized SDE. In particular, matching their variances alone does not necessarily imply that the resulting drift correction preserves the intended sampling distribution. Since this assumption is central to the proposed derivation, the paper would be stronger if the authors provided further theoretical justification or empirical validation of this connection.

(-) Limited methodological novelty: The proposed method applies a timestep-dependent scaling factor to the quantized model's predictions during the sampling process, which is similar to the correction mechanisms used in existing methods such as PTQD[A] and TAC-Diffusion[B]. In addition, the conditional residual variance used to compute the scaling factor has already been explored in D$^2$-DPM[C]. Although the SDE-based derivation provides a different perspective, the main methodological contribution appears to be the formulation of the scaling coefficient cic_i, rather than a fundamentally new correction mechanism. The paper would be stronger if it more clearly demonstrated the advantages of this formulation over existing compensation approaches.

(-) Insufficient ablation on the correction coefficient: The paper provides ablations on variance estimation and correction granularity, but does not sufficiently investigate the proposed scaling coefficient cic_i itself. In particular, it would be useful to compare the SDE-derived coefficient with simpler alternatives, such as heuristic variance-based scaling or a coefficient optimized to minimize the reconstruction error between full-precision and quantized predictions.

(-) Numerous incorrect references: The paper contains several incorrect references, including substantial errors in author names and paper titles. For example, the references for DPM-Solver-v3[D], PTQ4DM[E], Q-Diffusion[F], and Consistency Models[G] contain incorrect author lists, and some also have incorrect titles. Similar issues appear in other references, such as TFMQ-DM[H]. These are not merely minor formatting issues, as several entries do not accurately identify the original papers. The authors should carefully verify and correct the entire reference list.

(minor) Writing and presentation issues: The paper is difficult to follow, as the main idea is presented through several theoretical formulations before introducing the relatively simple correction mechanism. A more intuitive explanation of the overall approach and clearer connections between the theoretical formulations would improve readability.


References:

[A] Yefei He et al., “PTQD: Accurate Post-Training Quantization for Diffusion Models”, NeurIPS 2023

[B] Yuzhe Yao et al., “Timestep-Aware Correction for Quantized Diffusion Models”, ECCV 2024

[C] Qian Zeng et al., “D$^2$-DPM: Dual Denoising for Quantized Diffusion Probabilistic Models”, AAAI 2025

[D] Kaiwen Zheng et al., “DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics”, NeurIPS 2023

[E] Yuzhang Shang et al., “Post-training Quantization on Diffusion Models”, CVPR 2023

[F] Xiuyu Li et al., “Q-Diffusion: Quantizing Diffusion Models”, ICCV 2023

[G] Yang Song et al., “Consistency Models”, ICML 2023

[H] Yushi Huang et al., “TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models”, CVPR 2024

Questions for authors

Q1: How does the correction coefficient vary across denoising timesteps? The proposed method computes a timestep-dependent correction coefficient cic_i based on the conditional residual variance of quantization errors. However, the paper does not visualize how cic_i changes throughout the sampling process. Could the authors provide timestep-wise plots of cic_i for different models and quantization settings? Such an analysis would help clarify when and to what extent the proposed method modifies the sampling trajectory.

Q2: Why does Q-Drift degrade reconstruction quality on SDXL-Turbo? In Table 1, Q-Drift improves FID on SDXL-Turbo with MixDQ W4A8, but substantially degrades PSNR, SSIM, and LPIPS compared to the quantized baseline. This raises the concern that the proposed drift correction may improve distribution-level metrics at the cost of reconstruction fidelity. Could the authors provide a more detailed analysis of this behavior, including qualitative comparisons, to clarify why these metrics deteriorate despite the FID improvement?

Q3: What is the total calibration cost? Q-Drift requires paired full-precision and quantized model predictions to estimate the conditional residual variance at each denoising timestep. Although the inference overhead is minimal, the calibration process may introduce additional computational costs. Could the authors report the calibration time and memory requirements for different models? This would help assess the practical efficiency of the proposed method.

Q4: Can additional perceptual quality metrics be included? The experimental evaluation mainly relies on FID, CLIP score, and reconstruction-based metrics. Could the authors additionally report perceptual quality metrics such as ImageReward? These metrics would provide a more comprehensive assessment of whether the proposed correction improves perceptual image quality beyond distribution-level statistics.

4 comments

Thank you for the detailed review, and for recognizing that Q-Drift is simple, practical and broadly evaluated. We address your concerns below.

W1. Justification of the Gaussian perturbation assumption. Following the reviewer's suggestion, we tested it directly on FLUX.1-schnell W3A4 (1,000 paired trajectories):

  • Mean: small, 1–4% of the error std.
  • Cross-step correlation: 0.17–0.51 (PixArt-Σ: 0.26–0.39, decaying within about 5 steps).
  • Effect on the distribution: Q-Drift reduces the MMD between intermediate distributions and FP by 20–32% at every step.

The error is therefore not i.i.d. noise. We will present the SDE derivation as motivation for the correction, not as a guarantee of marginal preservation.

W2/W3. Novelty vs PTQD, TAC-Diffusion and D²-DPM; ablation of the coefficient. We credit D²-DPM for the conditional variance statistic. The key difference is the direction of the correction: PTQD, TAC and D²-DPM cancel or shrink the error, whereas Q-Drift lengthens the step. Following the reviewer's suggestion, we compared the coefficient with the alternatives you proposed:

FLUX.1-schnell W3A4, valc per stepFID-1k ↓ImageReward ↑
Quantized052.900.873
Reconstruction-optimal c*−0.052 … −0.00454.290.868
Q-Drift0.068, 0.018, 0.027, 0.09251.610.865

Minimizing the reconstruction error shrinks the step and makes FID worse, which confirms that quantized samplers under-step. Lengthening the step with Q-Drift's calibrated coefficient improves FID while keeping ImageReward within 0.01. We will clarify this in Section 4.

W4. References. We have corrected and verified the full reference list, including DPM-Solver-v3, PTQ4DM, Q-Diffusion, Consistency Models and TFMQ-DM.

W5. Writing. Section 4 will open with the one-line intuition: quantized models under-step, so each step is lengthened by (1 + c).

Q1. c across timesteps. On FLUX: 0.068, 0.018, 0.027, 0.092 (U-shaped). On SDXL-Turbo: 0.0011, 0.0031, 0.0087, 0.0203 (rising toward the end). The profile is model-specific, which is why it is calibrated rather than fixed. We will add plots for all models.

Q2. PSNR/LPIPS on SDXL-Turbo (MixDQ). This is a bug in our evaluation. The Q-Drift images were generated with different seeds from the FP and quantized images. We will regenerate this row with matching seeds. Across models, Q-Drift does not hurt fidelity to the FP sample:

W3A4PSNR ↑ (quant → Q-Drift)LPIPS ↓ (quant → Q-Drift)
FLUX.1-schnell13.83 → 13.840.425 → 0.413
SDXL-Turbo (1k test)14.55 → 14.600.386 → 0.384
PixArt-Σ, 30 steps13.47 → 13.540.494 → 0.483

We will add qualitative comparisons for MixDQ.

Q3. Calibration cost. FLUX: 5.2 min with K = 10 prompts and 70 min with K = 1,000, on one GPU including model loading (PixArt-Σ, K = 1,000: 112 min). As Table 3 shows, K = 10 is close to K = 1,000.

Q4. ImageReward. FLUX.1-schnell, 5k test: quantized 0.830, Q-Drift 0.825 with FID improving from 23.31 to 22.31. A constant tuned by FID reaches FID 21.17 but drops ImageReward to 0.782.

I thank the authors for the rebuttal. Some of the concerns are addressed, including references, calibration cost, and c across timesteps. However, the main concerns, including justification of gaussian perturbation (W1) and the contribution of the coefficient (W2/W3) are still not addressed, and these are critical since these are the main contributions of the paper. Therefore, I will maintain my score.

AuthorsOpenReview ↗

Answers to W1 and W2/W3

Thank you for your reply. We are glad that some of your concerns are addressed. Below, we clarify how our results address the two remaining ones.

W1. Why does variance matching give the right correction?

Our measurements on FLUX.1-schnell W3A4 show three things:

  1. The error mean is negligible (1–4% of the std), so the variance is the relevant statistic.
  2. Quantized steps are too short. The quantized prediction has less variance than FP (ratio 0.78–0.98), so each step under-moves.
  3. Q-Drift fixes exactly this. Its coefficient c is largest where the variance deficit is largest (steps 1 and 4), and the corrected trajectories are closer to FP at every step (MMD −20% to −32%).

We agree the error is not i.i.d. noise, so we will present the SDE view as motivation. The variance-based coefficient itself is supported by these measurements.

W2/W3. What does the coefficient contribute?

We compared it with the alternatives you suggested (FLUX.1-schnell W3A4):

CoefficientStepFID ↓ImageReward ↑Needs search?
None (quantized)–52.900.873–
Reconstruction-optimal (as in PTQD, TAC, D²-DPM)shorter54.290.868no
Q-Driftlonger51.610.865no
  • Minimizing reconstruction error makes FID worse. Q-Drift corrects in the opposite direction and improves it.
  • A constant tuned by FID reaches lower FID, but loses ImageReward (−0.05 to −0.07), and the best value varies about 20× across models.
  • On two new seeds (SDXL-Turbo W3A4, 1,000 prompts each), Q-Drift improves FID (59.67 → 58.39; 58.91 → 57.22) and ImageReward (0.753 → 0.762; 0.776 → 0.783), while the tuned constant drops ImageReward to 0.688 / 0.712.

In short, the coefficient sets both the direction and the size of the correction from calibration, with no search. None of the simpler alternatives does this.

We hope this clarifies both points, and we would be grateful if you could take it into account and reconsider your score.

AuthorsOpenReview ↗

Kind request to check our response

Dear Reviewer, as the discussion period is closing soon, we wanted to briefly follow up. We ran both experiments you asked for: a test of the Gaussian assumption (W1) and a comparison with a coefficient optimized for reconstruction error (W2/W3). We would be grateful if you could take a look and let us know if it addresses your concern.

Reviewer_3uvBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Q-Drift is a training-free sampler-side correction. Offline, paired FP and quantized trajectories give the conditional residual variance VV of the quantization error; VV is read as injected-noise strength and matched to the Langevin term of the marginal-preserving SDE of Karras et al., and keeping the paired drift turns this into a per-step rescaling c=∣Δσ∣⋅V/(2σ)c = |\Delta\sigma| \cdot V / (2\sigma) of the increment xi+1=xi+Δσ(1+c)ϵ^θx_{i+1} = x_i + \Delta\sigma(1+c)\hat\epsilon_\theta. At inference: one scalar multiply per step, no injected noise, no extra NFE, extended to Euler, flow matching and DPM-Solver++. FID improves in all seven aggressive settings across three samplers (largest 4.79 on PixArt-Σ), KID drops everywhere, six of seven bootstrap intervals for ΔFID lie below zero — but the gains are a small fraction of the quantization gap (SDXL 7%, FLUX.1-dev 2%), four of seven milder settings in Appendix Table 5 get worse, and the correction only ever scales the increment, never its direction.

Critical strengths
  1. Correct problem, correct math, almost free. The gap between what PTQ optimizes and what the dynamics accumulate is real; I verified the derivations, including the second-order multistep coupling. One scalar multiply per step, no extra NFE, stacks onto SVDQuant and MixDQ — that cost profile is why the paper is worth arguing about.

  2. Breadth and statistical care above the norm. Seven settings across three samplers, KID alongside FID, paired bootstrap intervals reported, and the FLUX.1-dev crossing-zero interval stated plainly. This is why I am at weak rather than clear rejection.

Critical weaknesses
  1. Mechanically it is step-size modulation, compared against nothing trivial. Under all three samplers it only scales the increment's magnitude. The paired drift is meaningful only if quantization error really acts as Langevin noise; drop that and what remains is a plain per-step step-size rescaling — yet there is no comparison against a global λ\lambda tuned on the calibration set, a per-step constant cc, or schedule-shift / timestep-rescale tricks. The unconditional-variance ablation makes it worse: a larger VV gives 31.95 against 30.69, so the method is sensitive to VV's magnitude and a c→λcc \to \lambda c sweep is unavoidable. If the optimal λ\lambda is not 1, the derivation loses most of its force.

  2. The one principled justification is never tested, and the gains fail under milder quantization. Marginal preservation is the entire reason for matching the Langevin term, and nothing measures it: the only evidence is terminal FID and KID. This matters because at inference the quantization error is a deterministic function of state and weights, correlated across steps, across space and with the state — not the fresh i.i.d. noise the Langevin term requires. Meanwhile four of seven milder settings in Appendix Table 5 get worse (FLUX.1-dev 20.49 → 20.55, FLUX.1-schnell 18.62 → 18.69, SDXL 17.05 → 17.17, PixArt-Σ W4A4 19.61 → 19.65), unreported in the main text.

Questions for authors
  1. (Clarification) Report the magnitude and across-step profile of cc, and sweep c→λcc \to \lambda c for λ∈{0,0.25,0.5,1,2}\lambda \in \{0, 0.25, 0.5, 1, 2\}. Is λ=1\lambda = 1 optimal?
  2. (Clarification) Compare against a single global λ\lambda tuned on the calibration set, a per-step constant cc, and schedule shift / timestep rescale. If Q-Drift does not clearly beat one tuned scalar, what justifies the per-step machinery?
  3. (Clarification) Test marginal preservation directly, compare the distribution of xx under FP, quantized and Q-Drift at intermediate noise levels, via MMD or per-level statistics. Without this the Langevin interpretation is unfalsifiable.
  4. (Clarification) How much does calibration on a full-precision trajectory matter? Try advancing with the quantized model and iterating to a fixed point, and report what happens when cc calibrated at 30 steps is used at 20 or 50. As written cc is tied to model, PTQ method, sampler, schedule and step count, and none of that is discussed.
3 comments

Thank you for the thorough review, for verifying the derivations, and for recognizing the cost profile and the statistical care of our evaluation. We address your questions below.

Q2/W1. Comparison with a tuned global λ, a per-step constant, and schedule shift. Following the reviewer's suggestion, we tuned each baseline by FID on 1,000 held-out validation prompts (FLUX.1-schnell, SVDQuant W3A4, freshly quantized), then evaluated the best value on the 5,000 MJHQ test prompts with the same seeds:

FLUX.1-schnell W3A4Tuned by FIDVal FID-1k ↓Test FID-5k ↓Test ImageReward ↑
Quantized–52.9023.310.830
Q-Drift (λ = 1)no51.6122.310.825
Q-Drift, λ tuned (λ = 4)yes51.5221.850.779
Global constant c (c = 0.20)yes50.4821.170.782
Per-step constant c = κ|Δσ|/(2σ)yes51.6822.310.811
Timestep shift (best s = 0.5)yes52.5523.120.763

The same protocol on SDXL-Turbo W3A4 (SVDQuant, 4 steps; tuned on 1k validation prompts, tested on 1k held-out prompts):

SDXL-Turbo W3A4Tuned by FIDTest FID-1k ↓Test ImageReward ↑Test LPIPS ↓
Quantized–59.310.8190.386
Q-Drift (λ = 1)no57.400.8130.384
Global constant c (c = 0.066)yes53.990.7470.416
Per-step constantyes54.010.7580.415

On both models, a scalar tuned by FID reaches a lower FID, but it over-lengthens the step: ImageReward drops by 0.05–0.07 and LPIPS worsens. Tuning the constant by ImageReward instead does not help: on SDXL-Turbo validation, every constant (1–16×) scores below the uncorrected model (ImageReward 0.655–0.812 vs 0.823), and the best one (1×: FID 58.69, ImageReward 0.812) is worse than Q-Drift on both metrics (58.38, 0.824). Q-Drift, with no FID tuning, improves FID by 1.0–1.9 while keeping ImageReward within 0.006. Schedule shift does not help (s = 2: 54.41). The best constant also differs across models (0.20 on FLUX, 0.066 on SDXL-Turbo, 0.02 on SDXL, 0.01 on Sana), so a scalar must be searched by FID for every model, while Q-Drift sets the magnitude from a few minutes of calibration. We will add these baselines to the main table.

Q1. Magnitude, profile and λ sweep. On FLUX, c = 0.068, 0.018, 0.027, 0.092 (U-shaped; each step is lengthened by 1.8–9%).

λ00.250.5124
Val FID-1k ↓52.9052.6152.0751.6151.6451.52
Val ImageReward ↑0.8730.8720.8660.8650.8580.820

FID plateaus from λ = 1, while ImageReward starts to fall beyond λ = 1, so λ = 1 is a good operating point without tuning.

Q3/W2. Testing marginal preservation directly. RBF-MMD² between the x_t distributions (FLUX, 256 prompts, shared noise, ×10²):

After step1234
FP vs quantized1.443.374.104.36
FP vs Q-Drift1.152.583.072.95
FP split-half floor1.151.211.231.24

Q-Drift moves every intermediate distribution toward FP (−20 to −32%), but does not reach the floor. As the reviewer anticipated, the error is correlated across steps (0.17–0.51), so we will present the Langevin view as motivation, not as a guarantee.

Q4. Step-count transfer. We calibrated PixArt-Σ W3A4 once at 20 steps and sampled at 10 and 30 steps without re-calibration (1,000 prompts, same seeds):

StepsMethodFID-1k ↓ImageReward ↑PSNR ↑LPIPS ↓
10Quantized83.730.06713.660.485
10Q-Drift77.910.06213.790.473
30Quantized86.240.16113.470.494
30Q-Drift80.210.19913.540.483

(FP: FID 46.15 / 46.14.) The calibrated coefficient transfers across step counts: FID improves by about 6 at both, with PSNR and LPIPS also improving. We will add this table to the paper.

W2. Milder settings (Table 5). We will move Table 5 into the main text and state that the gains concentrate in aggressive quantization, where the quantization gap is large.

The rebuttal addresses the raised points with additional experiments; my assessment of the contribution is unchanged, so I keep my rating at 3.

AuthorsOpenReview ↗

Thank you, and an additional robustness result

Thank you for your reply, and for confirming that the rebuttal addresses your points. As an additional robustness check, we repeated the SDXL-Turbo W3A4 comparison on two more seeds (1,000 prompts each):

SeedMethodFID ↓ImageReward ↑LPIPS ↓
1042Quantized59.670.7530.384
1042Q-Drift (λ = 1)58.390.7620.384
1042Tuned constant54.160.6880.416
2042Quantized58.910.7760.385
2042Q-Drift (λ = 1)57.220.7830.384
2042Tuned constant53.790.7120.414

Q-Drift improves FID and ImageReward on both seeds without any tuning, while the tuned constant again trades perceptual quality for FID. We will add these results to the paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

We thank all reviewers for their careful and constructive feedback. We appreciate the recognition that Q-Drift is simple, practical and almost free at inference (all reviewers), that the derivations are correct (3uvB), and that the evaluation is broad and statistically careful (3uvB, yzFJ, fdeF). Following the reviewers' suggestions, we ran the requested experiments:

  • Tuned scalar baselines (3uvB, fdeF, yzFJ). On FLUX.1-schnell and SDXL-Turbo W3A4, a constant multiplier or λ tuned on validation by FID reaches a lower FID than Q-Drift, but it over-lengthens the step and loses perceptual quality: ImageReward drops by 0.05–0.07 (FLUX 0.830 → 0.78; SDXL-Turbo 0.819 → 0.75). Q-Drift's calibrated coefficient improves FID (FLUX 23.31 → 22.31; SDXL-Turbo 59.31 → 57.40) with ImageReward essentially unchanged, and it needs no FID search. The best constant also varies up to 20× across models (0.20 on FLUX, 0.066 on SDXL-Turbo, 0.02 on SDXL, 0.01 on Sana), so it must be re-tuned per model.
  • Direction of the correction (yzFJ, fdeF). Quantized samplers under-step. Lengthening the step helps on every model, while the reconstruction-optimal coefficient, which shrinks the step as prior corrections do, makes FID worse (54.29 vs 52.90).
  • Assumptions tested directly (3uvB, yzFJ, fdeF). The error mean is small (1–4% of the std), but the error is correlated across steps (0.17–0.51). Q-Drift reduces the MMD between intermediate distributions and FP by 20–32% at every step.
  • Step-count transfer (3uvB). Calibrated once at 20 steps on PixArt-Σ, Q-Drift improves FID by about 6 at both 10 and 30 steps.
  • Cost and metrics (yzFJ). Calibration takes 5.2 min with 10 prompts on FLUX; we add ImageReward, PSNR and LPIPS.

We will present the SDE argument as motivation rather than a guarantee, add the tuned baselines and the assumption tests to the paper, and fix all references.

AuthorsOpenReview ↗

Summary of the discussion

Dear Area Chair(s),

As the discussion period is closing soon, we would like to respectfully summarize the status of the discussion.

Reviewer 3uvB (weak accept):

  • Replied that the rebuttal addresses the raised points with additional experiments; keeps 3.
  • We added a robustness check over two new seeds, where Q-Drift improves over the quantized model on every metric.

Reviewer yzFJ (clear reject):

  • Confirmed that the references, calibration cost and the behaviour of c across timesteps are addressed.
  • Stated that W1 and W2/W3 are "still not addressed," but did not comment on the experiments we ran for both:
    • W1: the requested empirical validation of the Gaussian assumption. The error mean is 1–4% of its std, and Q-Drift moves every intermediate distribution toward full precision (MMD −20% to −32%).
    • W2/W3: the requested comparison with a coefficient optimized for reconstruction error. It makes FID worse (54.29 vs 52.90), while Q-Drift improves it (51.61).

Reviewer fdeF (weak reject): no response yet.

  • Called the comparison with a tuned constant multiplier "the key experiment." We ran it: the tuned constant lowers ImageReward by 0.05–0.07, and its best value varies about 20× across models, so it must be re-tuned per model. Q-Drift needs no tuning.

We would be very grateful if you could take this into account in your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.