acceptodds
Under review as a conference paper at ICLR 2027

Temporal Quantization for Diffusion Models with Certified Error Bounds

Abstract

Diffusion models generate high-quality samples, but repeated evaluations of large denoisers make deployment computationally expensive. Post-training quantization (PTQ) can reduce this cost, but how to verify and control quantization error through temporal quantization schemes remains challenging. To address this, we study temporal quantization of diffusion models, allowing deterministic, stochastic, and mixed-precision PTQ schemes across timesteps. By leveraging the moment generating function (-MGF), we develop the first closed-form, high-probability error bound for diffusion process under arbitrary quantization schedules. Based on this theoretical guarantee, we build a predictive quantization policy which selects timestep-wise quantization schemes to minimize the terminal error bound under a computational budget. Extensive experiments evaluate both the error bounds and the resulting policies. More than 98.8% of the generated samples are within the 95%-certified quantization error bound. Across selected benchmarks, our proposed method reduces terminal \(L_2\) error by 23.1%, and LPIPS by 23.4% on average. On an NVIDIA RTX 5070Ti GPU, our method achieves at least 16.6% less GPU memory footprint and 12.9% less denoising time than the FP16 baseline.

Primary area:
generative models
Keywords:
Diffusion Models, Quantization, Stochastic Process, Certified Guarantees

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7MArOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies temporal quantization of diffusion models, where the precision, the rounding rule (round-to-nearest or stochastic rounding), and an optional compensator can change across timesteps. The authors split the one-step quantization residual into a conditional bias and a zero-mean fluctuation, and derive a closed-form, high-probability bound on the terminal error under local Lipschitz and sub-Gaussian assumptions (Theorem 4.2). They then select the schedule that minimizes this bound under a computation budget, using dynamic programming with pruning. The parameters of the bound are estimated from samples and verified by split-conformal calibration. On CIFAR-10, SDXL-Turbo, SDXL, and PixArt-Σ\Sigma, the authors report that the bound covers the empirical errors and that the selected schedules reduce LPIPS and the terminal L2 error to the full-precision outputs.

Critical strengths
  1. The method is supported by a theoretical analysis.

  2. The method outperforms existing methods in the most settings.

Critical weaknesses
  1. The certification does not seem to follow directly from Theorem 4.2. Theorem 4.2 holds only if its parameters are upper bounds of their true values. In the paper, however, they are set to empirical statistics, which does not appear to guarantee this. For example:
  • βk\beta_k: Assumption 4.1 requires the bias to be bounded by βk\beta_k almost surely, while βk\beta_k is set to the 0.8 quantile over probe states (Appendix E).

  • ρk\rho_k: The proof requires that each step amplifies the error by at most ρk\rho_k in every generation at deployment. ρk\rho_k, however, is set to 1.1 times the largest amplification observed during calibration under random schedules (Eq. (28)), which may be exceeded on new prompts or under the selected schedule. The factor of 1.1 also seems heuristic, and the paper does not explain how it is chosen or whether it is sufficient.

This affects the main claim of the paper, which is the certified guarantee. It would be helpful if the authors could clarify how the estimated parameters satisfy the assumptions of Theorem 4.2, and adjust the use of "certified" accordingly.

  1. The comparison may not fully account for cost.
  • The cost of each method is not reported. Tables 1 and 4 report quality metrics only, without the latency or memory of each method, and without the schedule and budget that TQDiff selects. TQDiff can use FP16 at some steps (Appendix F.2) and tends to select higher precision at early steps (Sec. 5.2), while most baselines use a single low precision. Fig. 4 compares the runtime with FP16 only, and the memory of TQDiff is higher than that of the single-precision baselines (Ln453-455). Part of the gain may therefore come from additional computation rather than from better scheduling. A similar question applies to TMPDiff [A], whose budget is not reported.

  • Some closely related methods are not compared. PTQD [B] also allocates bit widths across timesteps, and TAC-Diffusion [C] and AccuQuant [D] address the accumulation of quantization errors across steps. These methods are discussed in Secs. 1 and 2, but none of them seems to be evaluated.

Without this information, it is difficult to assess the trade-off between error and cost, which determines the value of a scheduling policy. It would be helpful to report the cost of each method in Tables 1 and 4, compare the methods under matched budgets, and include these methods if feasible.

References

[A] TMPDiff: Temporal Mixed-Precision for Diffusion Models, arXiv 2026

[B] PTQD: Accurate Post-Training Quantization for Diffusion Models, NeurIPS 2023

[C] Timestep-Aware Correction for Quantized Diffusion Models, ECCV 2024

[D] AccuQuant: Simulating Multiple Denoising Steps for Quantizing Diffusion Models, NeurIPS 2025

Questions for authors

Among DiTs, the experiments seem to be conducted on PixArt-Σ\Sigma only. Could the authors evaluate TQDiff on recent flow-matching DiTs such as FLUX.1 or SD3.5?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7MAr (Part 1)

We thank you for your constructive review and for highlighting the need to clarify how our theoretical bound relates to empirically estimated parameters and to evaluate TQDiff against related methods under matched computational budgets. All corresponding revisions are marked in blue in the manuscript.

W1: How Estimated Parameters Connect to Theorem 4.2 and Deployment Certification

Manuscript location: Revised in Abstract (p. 1), Section 4.1 (Assumption 4.1 & Theorem 4.2, pp. 5–6), New Section 4.2 (Proposition 4.3, p. 6), Section 5.2 & Revised Figure 2 (p. 8), Appendix D.2 (p. 19), Appendix E (p. 20), and New Appendix G (Table 5, p. 27).

You are completely right that Theorem 4.2 requires valid per-step upper bounds (ρk,βk,Σk)(\rho_k, \beta_k, \Sigma_k), and that empirical 0.80.8-quantile biases and 1.1×1.1\times secant heuristics alone do not establish almost-sure worst-case conditions. To address your request ("clarify how the estimated parameters satisfy the assumptions of Theorem 4.2, and adjust the use of 'certified' accordingly"), we formalize a two-stage certification framework and clarify the exact scope of "certified":

  1. Three-tier terminology separation: Inside the paper, we reserve "certified" strictly for (a) Theorem 4.2 when valid per-step envelopes hold (referred to as the analytical high-probability bound), and (b) the post-calibration deployment tube q^uk\hat{q} u_k or qηukq_\eta u_k (Proposition 4.3 and Appendix E). The uncalibrated radius r^N\hat{r}_N obtained by substituting finite-sample parameter estimates into Theorem 4.2 is explicitly termed the planning radius (also plug-in radius alongside Figures 2 and 4, and never claimed as a standalone certified bound), while the Title ("...with Certified Error Bounds") reflects the complete two-stage certified deployment guarantee.
  2. Why Theorem 4.2 remains essential and empirically tight (Section 4.2, p. 6; Table 5, p. 27):
    • Planning & tube geometry: Theorem 4.2 composes offline estimates into a planning radius r^N\hat{r}_N that ranks candidate schedules with Spearman 0.99 (Figure 4) and defines the step-wise profile uk=max⁡{r^k,10−3}u_k = \max\{\hat{r}_k, 10^{-3}\}.
    • Empirical accuracy of the propagation law: On the terminal error ∥EN∥\|E_N\| targeted by Theorem 4.2, the uncalibrated terminal radius r^N\hat{r}_N achieves 100% coverage (0/2,0000/2{,}000 violations; q^term=1.000\hat{q}_{\mathrm{term}} = 1.000) across all held-out trajectories on SDXL and PixArt-Σ\Sigma. On the stricter whole-trajectory tube, replacing 0.80.8-quantile β^k\hat{\beta}_k with the probe-state maximum β^kmax⁡\hat{\beta}_k^{\max} (Figure 2) reduces uncalibrated SDXL violations to 1.8% (MJHQ) and 2.0% (sDCI) (q^=1.000\hat{q} = 1.000 at δ=5%\delta = 5\%).
  3. Finite-sample certification at deployment (Proposition 4.3, Section 4.2, p. 6; Appendix E, Eqs. (32)–(33), p. 20):
    • Because finite-sample estimates (such as the 0.80.8-quantile β^k\hat{\beta}_k used for non-conservative planning, or multistep solver dynamics on PixArt-Σ\Sigma) can deviate from almost-sure worst-case bounds at intermediate steps, we freeze the selected schedule and calibrate Si=max⁡1≤k≤N∥Ek(i)∥/ukS_i = \max_{1 \le k \le N} \|E_k^{(i)}\| / u_k on mm independent trajectories.
    • Proposition 4.3 yields simultaneous finite-sample coverage P(∥Ek∥≤q^uk, ∀1≤k≤N)≥1−δ\mathbb{P}(\|E_k\| \le \hat{q} u_k,\ \forall 1 \le k \le N) \ge 1 - \delta without requiring (ρ^k,β^k,Σ^k)(\hat{\rho}_k, \hat{\beta}_k, \hat{\Sigma}_k) to satisfy Assumptions 3.1–4.1 (Appendix E gives the 1−η1-\eta PAC tolerance counterpart). On SDXL/MJHQ (m=1000m = 1000, n=1000n = 1000), q^=1.405/2.194/3.192\hat{q} = 1.405 / 2.194 / 3.192 at δ=5%/1%/0.1%\delta = 5\% / 1\% / 0.1\%, yielding held-out tube violations of 5.9%, 1.4%, and 0.2% (Table 5); on SDXL/sDCI and PixArt-Σ\Sigma/MJHQ (m=n=500m = n = 500), held-out violations are 3.8% / 1.8% and 4.2% / 1.8% at δ=5%/1%\delta = 5\% / 1\%.

W2a: Reporting Costs, Schedules, and Matched-Budget Comparisons

Manuscript location: Revised in Section 5.2 ("Quality at Matched Cost" & "Runtime and GPU Memory", p. 10), New Figure 4 (p. 10), New Appendix H (Tables 6–7, pp. 27–28), and New Appendix I (Table 8, p. 29).

To ensure that quality gains come from scheduling rather than extra compute or extra FP16 steps, Appendix H (Tables 6–7, pp. 27–28) reports the exact action libraries, budgets, selected schedules, FP16 step counts, average bit widths, measured denoising latencies, and peak GPU memory for all methods on a shared RTX 5070 Ti execution engine:

(the rest of the response is in the next block due to the characters limit.)

AuthorsOpenReview ↗

Response to Reviewer 7MAr (Part 2)

  • SDXL (N=50N=50, budget Cˉ=2.120 s\bar{C} = 2.120\text{ s} matched to TMPDiff): TQDiff selects 4 FP16 steps (0–3), 18 W8A8-RN steps (4–21), and 28 W4A4-RN steps (22–49), using fewer FP16 steps (4 vs. 12) and fewer average bits (6.4 vs. 6.9) than TMPDiff (12 FP16 + 38 W4A4-RN). Both schedules run in 1.90 s/img1.90\text{ s/img} (vs. 2.40 s2.40\text{ s} for FP16, 2.02 s2.02\text{ s} for uniform W8A8, and 1.74 s1.74\text{ s} for uniform W4A4), while TQDiff lowers LPIPS from 0.158 to 0.105 on 200 fresh prompts (paired ΔLPIPS=−0.053 [−0.059,−0.046]\Delta\text{LPIPS} = -0.053\ [-0.059, -0.046]).
  • PixArt-Σ\Sigma (N=20N=20, budget Cˉ=2.956 s\bar{C} = 2.956\text{ s} matched to TMPDiff): TQDiff uses 0 FP16 steps, selecting 10 W8A8-RN steps (3–12) and 10 W4A4-RN steps (0–2 and 13–19) (6.0 avg. bits vs. TMPDiff's 7.0 avg. bits with 5 FP16 steps at 2–6), reducing LPIPS from 0.227 to 0.183 (ΔLPIPS=−0.044 [−0.048,−0.039]\Delta\text{LPIPS} = -0.044\ [-0.048, -0.039]) on 500 held-out prompts at 2.89 s2.89\text{ s} vs. 2.93 s2.93\text{ s} denoising latency.
  • Six matched budgets on SDXL (Figure 4, right, p. 10; Table 8, Appendix I, p. 29): Across six budgets matching TMPDiff-family schedules with K∈{2,6,12,17,21,28}K \in \{2, 6, 12, 17, 21, 28\} FP16 steps (planned cost differences <12 ms< 12\text{ ms}) plus uniform W8A8's budget (2254 ms2254\text{ ms}), TQDiff achieves paired LPIPS reductions of 0.028 to 0.069 at every budget, with all 95% bootstrap CIs excluding zero.
  • Memory disclosure (Table 7, p. 28): All-resident SDXL TQDiff uses 7.97 GiB7.97\text{ GiB} peak memory vs. 6.02 GiB6.02\text{ GiB} for TMPDiff and 4.90 GiB4.90\text{ GiB} for FP16 (on PixArt-Σ\Sigma, TQDiff uses 1.12 GiB1.12\text{ GiB} vs. 1.52 GiB1.52\text{ GiB} for TMPDiff and 1.18 GiB1.18\text{ GiB} for FP16). Sequential residency lowers SDXL TQDiff to 4.83 GiB4.83\text{ GiB} (0.77 GiB0.77\text{ GiB} on PixArt-Σ\Sigma) at 2.05 s/img2.05\text{ s/img} (+0.16 s+0.16\text{ s} transfer overhead at batch size 1, or +0.02 s+0.02\text{ s} at batch size 8).

W2b: Comparisons with PTQD, TAC-Diffusion, and AccuQuant

Manuscript location: Revised in Section 2.2 & 2.3 (pp. 3–4), Section 5.2 (p. 10), and New Appendix K (Table 9, p. 30).

  • Controlled comparison with PTQD [B] (Appendix K, Table 9, p. 30): We implemented PTQD's SNR-based mixed-precision allocation and affine step correction ϵq′=(ϵq−μk)/(1+kk)\epsilon'_q = (\epsilon_q - \mu_k)/(1 + k_k) on our shared SDXL engine (using W8A8 in place of W4A8 for kernel compatibility, fitting (μk,kk)(\mu_k, k_k) on 64 prompts, and omitting noise-variance correction since DDIM has η=0\eta = 0). Under the exact same two-precision library {W8A8-RN,W4A4-RN}\{\text{W8A8-RN}, \text{W4A4-RN}\} and planned budget (2118 ms2118\text{ ms}), TQDiff achieves LPIPS 0.130 (L2=36.18L_2 = 36.18) vs. 0.369 (L2=75.20L_2 = 75.20) for PTQD's SNR ordering without correction and 0.371 (L2=75.53L_2 = 75.53) with correction (paired allocation improvement −0.239 [−0.259,−0.219]-0.239\ [-0.259, -0.219] on 200 prompts; full-library TQDiff achieves 0.106). The schedule structure explains why: PTQD's local SNR criterion assigns W8A8 to the final 27 steps (r23827r^{23}8^{27}) where clean-signal SNR is highest, whereas TQDiff accounts for multi-step error amplification Gk=∏j>kρjG_k = \prod_{j>k}\rho_j and assigns W8A8 to the initial 27 steps (827r238^{27}r^{23}), preventing early quantization errors from compounding.
  • Relationship to TAC-Diffusion [C] and AccuQuant [D] (Section 2.2 & 2.3, pp. 3–4; Appendix K, p. 30): TAC-Diffusion (timestep-aware dynamic correction of quantization residuals and exposure bias) and AccuQuant (multi-step quantizer calibration) improve individual quantized operators rather than selecting temporal precision schedules. They are therefore orthogonal to TQDiff and can enter Ak\mathcal{A}_k as candidate actions, just as TCEC does. We now explicitly introduce TAC-Diffusion in Section 2.2 (p. 3) and clarify its complementary role alongside AccuQuant in Section 2.3 (p. 4) without claiming empirical superiority over unevaluated operator-level methods.

Question: Applicability to Recent Flow-Matching DiTs (FLUX.1 or SD3.5)

Manuscript location: Revised in Section 2.1 (p. 3) and Section 6 Conclusion (p. 10).

Mathematically, deterministic flow-matching Euler updates take the form xk+1=Fk(xk;c)=xk+Δtkvθ(xk,tk,c)x_{k+1} = F_k(x_k; c) = x_k + \Delta t_k v_\theta(x_k, t_k, c), which is a direct instance of the discrete-time update in Eq. (1). Thus, Theorem 4.2, Proposition 4.3, and our DP planner apply directly to flow-matching DiTs once per-step action profiles (ρ^k,β^k,Σ^k)(\hat{\rho}_k, \hat{\beta}_k, \hat{\Sigma}_k) are collected. While our PixArt-Σ\Sigma experiments validate TQDiff on a Diffusion Transformer with a multistep solver, we have not yet conducted empirical evaluations on 12B-scale flow-matching models such as FLUX.1 or SD3.5, and we state this scope explicitly in Section 6 (p. 10).

I thank the authors for the detailed responses and additional experiments. However, in the revision, the certified guarantee comes from the split-conformal calibration newly introduced in Sec. 4.2, not from Theorem 4.2. Theorem 4.2 now serves mainly as a planning objective. Since the certified bound derived from the analysis is the main contribution of the paper, this change weakens the contribution. I therefore keep my rating.

Reviewer_dQ5COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates temporal PTQ for diffusion models, which allows varying precision, rounding rules (round-to-nearest vs. stochastic rounding), and error compensators across denoising timesteps. Instead of linearizing the denoiser, the authors model quantization error dynamics by decomposing the one-step residual into a conditional bias and a zero-mean sub-gaussian fluctuation. The authors first adopt a projection-based decomposition and the L_2 moment generating function (called L_2-MGF), and combined with Ville’s supermartingale inequality, the authors derive a closed-form, high-probability bound on the terminal L2L_2 deviation from the full-precision trajectory. This theoretical bound is then used as an objective in an offline dynamic programming planner (with dominance pruning and grid rounding) to select a timestep-wise quantization schedule under a computational budget and a precision-switch constraint. Empirical evaluations on CIFAR-10 DDPM, SDXL, SDXL-Turbo, and PixArt-Σ\Sigma demonstrate that the empirical error falls within the certified bounds (with split-conformal calibration) and that the resulting schedules reduce terminal L2L_2 and LPIPS errors relative to baselines.

Critical strengths
  • The application of L2L_2-MGF and supermartingale techniques (adapting recent stochastic control reachability tools) to diffusion quantization error propagation is mathematically clean and avoids the uncontrolled remainder terms found in prior first-order Taylor linearization approaches.
  • The framework unifies heterogeneous quantization actions (e.g., mixing bit-widths, deterministic rounding, stochastic rounding, and error compensators) into a single state-space optimization problem.
  • The dynamic programming planner is well-designed, utilizing monotone dominance pruning and multiplicative grid rounding to make exponential schedule spaces tractable within seconds.
  • The experimental setup explicitly accounts for hardware realities, such as the latency overhead of reloading models during precision switches and the memory-latency trade-offs measured via CUDA graphs.
Critical weaknesses
  • The paper claims "certified error bounds," which is misleading given how parameters are obtained. The theoretical guarantee in Theorem 4.2 relies on true local Lipschitz constants ρk\rho_k and almost-sure upper bounds βk,Σk\beta_k, \Sigma_k. In Appendices D.2 and E, ρk\rho_k is estimated via 1.1$\times$ the empirical maximum of secants along random trajectories, and βk\beta_k is taken as the 0.8 empirical quantile over probe states. Taking the 0.8 quantile violates the almost-sure bound requirement of Assumption 4.1 on 20% of the calibration data itself.
  • The optimization problem in Eq. (11) assumes the certificates βk(a)\beta_k(a) and Σk(a)\Sigma_k(a) are schedule-independent functions of step kk and action aa. In reality, the conditioning filtration FkF_k and the state XkX_k depend on the sequence of prior actions a0:k−1a_{0:k-1}. Estimating these quantities offline on fixed probe states ignores how early quantization choices shift the input distribution at step kk, invalidating the exactness of the dynamic programming state-space formulation.
  • Optimizing the terminal L2L_2 bound rNr_N does not reliably correlate with perceptual or distributional quality. On CIFAR-10, TQDiff reduces terminal L2L_2 from 6.646 to 2.395, yet the FID degrades catastrophically from 5.895 (W4A8+RN) to 14.360. On PixArt-Σ\Sigma, the complex planner simply degenerates to a static single-precision W8A8+RN schedule, offering no benefit over standard PTQ.
  • The efficiency gains over full-precision models are modest compared to standard static low-bit quantization, and the paper omits direct runtime and memory comparisons against competing temporal methods like TMPDiff or SVDQuant in Figure 4, comparing only against FP16 and static RN baselines.
  • The manuscript contains several presentation and consistency errors, such as swapping the memory and runtime saving percentages between the Abstract (16.6% memory, 12.9% time) and the Introduction (12.9% memory, 16.6% time), and mislabeling the leftmost bar of Figure 2(a–f) as "RN" on the x-axis when the caption states it represents the bound RNR_N.
Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer dQ5C (part 1)

Manuscript location: Revised in Abstract (p. 1), Section 4.1 (Assumption 4.1 & Theorem 4.2, pp. 5–6), New Section 4.2 (Proposition 4.3, p. 6), Section 5.2 & Revised Figure 2 (p. 8), Appendix D.2 (p. 19), Appendix E (p. 20), and New Appendix G (Table 5, p. 27).

We fully agree that taking the 0.80.8 empirical quantile for β^k\hat{\beta}_k (exceeded on 20% of probe states by construction) and empirical secants for ρ^k\hat{\rho}_k does not verify the almost-sure assumptions of Theorem 4.2. Rather than discarding certification, our revision cleanly separates the two stages of our two-stage certification pipeline:

  1. Separating planning from finite-sample certification (Section 4.2, p. 6): Theorem 4.2 supplies the analytical recursion used to compose offline per-step estimates into an estimated (plug-in) planning radius r^N\hat{r}_N for searching ∏k∣Ak∣\prod_k |\mathcal{A}_k| schedules and shaping the step-wise profile uk=max⁡{r^k,10−3}u_k = \max\{\hat{r}_k, 10^{-3}\}. Proposition 4.3 then certifies the frozen selected schedule on mm fresh paired trajectories using the whole-trajectory score Si=max⁡1≤k≤N∥Ek(i)∥/ukS_i = \max_{1 \le k \le N} \|E_k^{(i)}\| / u_k, guaranteeing simultaneous finite-sample coverage P(∥Ek∥≤q^uk, ∀1≤k≤N)≥1−δ\mathbb{P}(\|E_k\| \le \hat{q} u_k,\ \forall 1 \le k \le N) \ge 1 - \delta without relying on exact almost-sure per-step constants. (Appendix E, Eqs. (32)–(33), p. 20, further gives the i.i.d. PAC tolerance bound controlling conditional deployment risk by δ\delta with confidence 1−η1 - \eta.)
  2. Empirical tightness of Theorem 4.2 and effect of β^kmax⁡\hat{\beta}_k^{\max} (Figure 2, p. 8; Table 5, p. 27):
    • Across all 2,0002{,}000 held-out trajectories on SDXL (MJHQ & sDCI) and PixArt-Σ\Sigma (MJHQ), the uncalibrated terminal radius r^N\hat{r}_N has 0% terminal violation (q^term=1.000\hat{q}_{\mathrm{term}} = 1.000).
    • Uncalibrated whole-tube violations under the 0.80.8-quantile profile (17.4%–25.0%17.4\%–25.0\%) occur primarily right after the first quantized step, where r^k\hat{r}_k equals a single 0.80.8-quantile β^k\hat{\beta}_k. Replacing that 0.80.8 quantile with the probe-state maximum β^kmax⁡\hat{\beta}_k^{\max} (Figure 2, second bar) reduces uncalibrated SDXL whole-tube violations to 1.8% (MJHQ) and 2.0% (sDCI) (q^=1.000\hat{q} = 1.000 at δ=5%\delta = 5\%). Split-conformal calibration certifies any profile across all models and δ\delta levels (5.9%/1.4%/0.2%5.9\% / 1.4\% / 0.2\% on SDXL/MJHQ at δ=5%/1%/0.1%\delta = 5\% / 1\% / 0.1\%; Table 5). Inside the text, we reserve "certified" strictly for the analytical theorem under valid envelopes and the calibrated deployment tube, referring to the uncalibrated objective r^N\hat{r}_N as the "planning radius" (or "plug-in radius" alongside Figures 2 and 4).

W2: Schedule-Dependence of Intermediate States XkX_k and Exactness of the DP Formulation

Manuscript location: Revised in Section 4.3 (p. 7), Appendix F.3 ("State-space form", p. 22), and New Appendix I ("On-policy profile stability check", p. 29).

Your observation regarding how prior actions a0:k−1a_{0:k-1} shift the state distribution XkX_k at step kk is spot-on. We now clarify this relationship at both the analytical and empirical levels:

  1. Analytical level vs. surrogate DP objective (Section 4.3, p. 7; Appendix F.3, p. 21): If (βk(a),τk2(a),ℓk(a))(\beta_k(a), \tau_k^2(a), \ell_k(a)) are true uniform envelopes over the reachable tube X\mathcal{X} (Assumption 4.1), they upper-bound the conditional residual across all reachable states Xk∈XX_k \in \mathcal{X} regardless of the prior action sequence a0:k−1a_{0:k-1}; for such uniform envelopes, the DP recursion over (Dk,Tk,Vk)(D_k, T_k, V_k) is exact. In practice, profiling separate cards for all ∣A∣k|\mathcal{A}|^k action prefixes is intractable and prone to small-sample overfitting. Instead, we pool probe states from multiple carrier schedules spanning diverse precision histories and optimize the resulting fixed empirical objective r^N\hat{r}_N in Eq. (11). Crucially, deployment certification does not rely on schedule-independence of the pooled profiles, because Proposition 4.3 certifies the selected schedule end-to-end on its own trajectories.
  2. Empirical on-policy stability check (Appendix I, p. 29): To test whether state-distribution shift under the selected schedule affects planning optimality, we re-profiled all candidate actions along the selected SDXL schedule's own trajectories across 64 fitting prompts:
    • Pooling these on-policy profiles with the carrier profiles leaves the DP's selected schedule completely unchanged.
    • Re-planning solely from on-policy profiles changes only 2 of 50 steps (under our random-direction secant fallback at early steps) or 7 of 50 steps (when retaining pooled ρ^k\hat{\rho}_k at early steps), yielding negligible paired LPIPS differences of +0.002+0.002 ([−0.001,+0.006][-0.001, +0.006]) and +0.003+0.003 ([−0.002,+0.008][-0.002, +0.008]) on 200 held-out prompts.

(the rest of the response is in the next block due to the characters limit.)

AuthorsOpenReview ↗

Response to Reviewer dQ5C (part 2)

W3: Correlation of rNr_N with Perceptual/Distributional Quality, CIFAR-10 FID, and PixArt-Σ\Sigma Scheduling

Manuscript location: Revised in Section 5.2 (pp. 9–10), New Figure 4 (p. 10), New Appendix H (Table 7, p. 28), New Appendix I (Table 8, p. 29), and New Appendix J (p. 29).

  1. Strong correlation with perceptual/trajectory fidelity (r^N\hat{r}_N vs. LPIPS and L2L_2): Across 41 candidate SDXL schedules on 70 prompts (Figure 4, left, p. 10; Appendix I, p. 28), r^N\hat{r}_N achieves a Spearman correlation of 0.99 ([0.98,0.99][0.98, 0.99]) with both mean terminal L2L_2 and mean LPIPS, far exceeding simple cost or bit-width proxies (≤0.55\le 0.55).
  2. Why CIFAR-10 FID degrades despite lower L2L_2 and LPIPS (Section 5.2, p. 9; Appendix J, p. 29): We explicitly acknowledge that controlling pairwise trajectory L2L_2 deviation does not guarantee lower distributional FID. In a 50,000-sample single-backend diagnostic fixing the 100-step CIFAR schedule against a shared FP32 reference (Appendix J), TQDiff reduces terminal L2L_2 from 6.6466.646 to 2.4252.425 and LPIPS from 0.11010.1101 to 0.06430.0643 relative to W4A8-RN, while clean-FID is 15.7815.78 vs. 5.905.90. Spectral analysis reveals that late low-precision/compensated steps leave a small, high-frequency residual (36.9%36.9\% of error energy at frequencies ≥8\ge 8 cycles/image vs. 20.1%20.1\% for W4A8-RN) that Inception features penalize heavily; replacing only the last 20 steps with full precision restores clean-FID to 5.91 (L2=2.130L_2 = 2.130).
  3. Non-trivial mixed schedule on PixArt-Σ\Sigma under a tighter budget (Section 5.2, pp. 9–10; Appendix H, Table 7, p. 28): In Table 1, the default budget was loose enough to afford W8A8+RN at every step, which is already near-lossless (LPIPS 0.09560.0956). When evaluated at TMPDiff's tighter latency budget (Cˉ=2.956 s\bar{C} = 2.956\text{ s}, where uniform W8A8 at 3.170 s3.170\text{ s} planned cost is infeasible), TQDiff selects a genuine mixed schedule—10 W8A8-RN steps (3–12) and 10 W4A4-RN steps (0–2 and 13–19) with zero FP16 steps (6.06.0 vs. TMPDiff's 7.07.0 avg. bits)—reducing LPIPS from 0.227 to 0.183 (paired ΔLPIPS=−0.044 [−0.048,−0.039]\Delta\text{LPIPS} = -0.044\ [-0.048, -0.039]) on 500 held-out prompts at 2.89 s2.89\text{ s} vs. 2.93 s2.93\text{ s} denoising latency.

W4: Direct Runtime and Memory Comparisons Against Static PTQ, SVDQuant, and TMPDiff

Manuscript location: Revised in Section 4.3 (p. 7), Section 5.2 ("Quality at Matched Cost" & "Runtime and GPU Memory", p. 10), New Appendix H (Tables 6–7, pp. 27–28), and New Appendix I (Table 8, p. 29).

Table 7 (Appendix H, p. 28) benchmarks FP16, static W4A4-RN (SVDQuant-style rank-32 group-64 NVFP4 backend), static W8A8-RN, TMPDiff, and TQDiff on the same RTX 5070 Ti engine:

  • Denoising latency & quality: On SDXL, static W4A4-RN takes 1.741 s1.741\text{ s} (LPIPS 0.3710.371), static W8A8-RN 2.020 s2.020\text{ s} (LPIPS 0.1040.104), TMPDiff 1.902 s1.902\text{ s} (LPIPS 0.1580.158), and TQDiff 1.897 s1.897\text{ s} (LPIPS 0.1050.105) vs. 2.403 s2.403\text{ s} for FP16 (−33%-33\% LPIPS vs. TMPDiff at matched latency). Across all six matched budgets (Table 8, p. 29; including uniform W8A8's budget, where TQDiff lowers LPIPS from 0.1040.104 to 0.0480.048), TQDiff consistently outperforms every baseline.
  • Memory residency trade-off: Holding all active precisions resident uses 7.97 GiB7.97\text{ GiB} on SDXL vs. 6.02 GiB6.02\text{ GiB} (TMPDiff) and 4.90 GiB4.90\text{ GiB} (FP16); on PixArt-Σ\Sigma, TQDiff uses 1.12 GiB1.12\text{ GiB} vs. 1.52 GiB1.52\text{ GiB} (TMPDiff) and 1.18 GiB1.18\text{ GiB} (FP16). Sequential residency lowers peak memory to 4.83 GiB4.83\text{ GiB} (SDXL) and 0.77 GiB0.77\text{ GiB} (PixArt-Σ\Sigma), adding 0.16 s/img0.16\text{ s/img} at batch size 1 (2.054 s2.054\text{ s}) or 0.02 s/img0.02\text{ s/img} at batch size 8.

W5: Presentation and Consistency Fixes

Manuscript location: Revised in Abstract & Section 1 (pp. 1–2), Section 3.1 (p. 4), Section 4.3 (p. 7), Section 5.2 & Revised Figure 2 (p. 8), and Appendices B–E (pp. 15–20).

Thank you for catching these issues: (1) we removed the swapped memory/runtime savings in the Abstract and Section 1, replacing them with exact residency-specified latency and memory results (Section 5.2, p. 10; Table 7, p. 28); (2) we replaced Figure 2 (p. 8) to fix the "RN" tick and show each schedule's planning radius r^N\hat{r}_N, β^kmax⁡\hat{\beta}_k^{\max} radius, and calibrated tube q^r^N\hat{q}\hat{r}_N with consistent rNr_N notation; and (3) we standardized step indexing (k=0,…,N−1k = 0, \ldots, N-1 in Section 3.1, p. 4 and Section 4.3, p. 7), positive semidefinite Σk\Sigma_k (Assumption 4.1, p. 5), and ρ0=1\rho_0 = 1 (Appendix D.2, p. 19).

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

Temporal quantization: each step picks an action from a library (bit width, RN or SR, optional TCEC compensator) and steps may differ, so one trajectory mixes deterministic and stochastic schemes. The goal is a high-probability bound on the quantized trajectory's deviation from full precision, not better FID. Splitting the residual into a conditional bias bkb_k and a zero-mean ξk\xi_k and decoupling them through an L2L^2-MGF gives a supermartingale that Ville's inequality closes into rN=DN+TN+2VNlog⁡(1/δ)r_N = D_N + \sqrt{T_N} + \sqrt{2V_N\log(1/\delta)}. rNr_N is minimized by DP under budget Cˉ\bar C (lossless dominance pruning, ≤5.2% grid relaxation, ≤3 precision switches); certificates cost O(N∣A∣(R+n))O(N|\mathcal{A}|(R+n)) once and the DP takes 3.3 s at N=50N=50. Four models, NN from 4 to 100; over single-precision PTQ, SDXL + MJHQ goes LPIPS 0.358 → 0.098 and terminal L2 76.0 → 29.4.

Critical strengths
  1. The bound is real and stronger than what surrounds it. No linearization, no independence across steps, clean reduction to the deterministic case; TCEC leaves the remainder uncontrolled and TMPDiff's error accumulation is an assumption. Verified line by line — the central claim holds, which is why I do not score lower.

  2. The theory runs. Certificates estimated once and reused for every schedule; DP in 3.3 s at N=50N=50; gains over single-precision PTQ large enough to matter (LPIPS 0.358 → 0.098, terminal L2 76.0 → 29.4).

Critical weaknesses
  1. The certificates are estimated, not certified. ρk\rho_k is 1.1× the empirical maximum of paired secants; βk\beta_k is the 0.8 quantile over probe states; Σ^k\hat\Sigma_k is a sample covariance with n≪dn \ll d, so its rank is at most n−1n-1 and tr Σ^\mathrm{tr}\,\hat\Sigma systematically underestimates the true dd-dimensional variance. The assumption asks for an almost-sure bound and receives an empirical quantile; split-conformal covers only the final scalar radius. "The theorem holds given the true (ρ,β,Σ)(\rho,\beta,\Sigma)" and "the deployed bound holds" are different claims, and the line is never drawn. The certificate is the entire contribution, so this is decisive.

  2. Nothing is cost-matched. A\mathcal{A}, Cˉ\bar C and the realized cost of the selected schedule are all unreported, so I cannot separate smarter scheduling from spending more compute. Efficiency is only compared against FP16, while memory is conceded to be worse than under single-precision quantization.

Questions for authors
  1. (Clarification) A\mathcal{A} and Cˉ\bar C per experiment, and what the selected schedule costs — average bit width, number and position of FP16 steps — placed beside the baselines.
  2. (Clarification) What are RR and nn, and is the tr Σ^\mathrm{tr}\,\hat\Sigma underestimate corrected? Report the conformal score distribution and empirical violation rate at δ=5%\delta = 5\%, and at 1%1\% and 0.1%0.1\%.
  3. (Clarification) CIFAR FID is 14.360 against 4.990 (FP16) and 5.895 (W4A8+RN), while LPIPS, PSNR and L2 on that row all improve. Which schedule was selected, and does pinning it while forcing a single backend reproduce the number?
  4. (Clarification) Plot rNr_N against realized terminal error across candidate schedules. If they are uncorrelated the bound does no better than a cheap proxy, and TMPDiff already ties or beats TQDiff on SDXL-Turbo + sDCI (L2 36.132 vs 36.556).
4 comments
AuthorsOpenReview ↗

Response to Reviewer Moda (part 1)

We deeply appreciate your thorough, line-by-line verification of our theoretical analysis and your sharp, constructive questions. Your distinction between "the theorem holds given the true (ρ,β,Σ)(\rho, \beta, \Sigma)" and "the deployed bound holds" directly shaped our two-stage certification architecture in the revised manuscript.

W1 & Q2: Two-Stage Certification, tr⁡Σ^\operatorname{tr}\hat{\Sigma} Analysis, and Calibration Statistics

Manuscript location: Revised in Abstract (p. 1), Section 4.1 (Assumption 4.1 & Theorem 4.2, pp. 5–6), New Section 4.2 (Proposition 4.3, p. 6), Section 5.2 & Revised Figure 2 (p. 8), Appendix D.2 (Eqs. (28)–(30), p. 19), Appendix E (Eqs. (31)–(33), p. 20), and New Appendix G (Table 5, p. 27).

  1. Drawing the exact line via two-stage certification (Section 4.2, p. 6): We completely agree that empirical quantiles (0.80.8-quantile β^k\hat{\beta}_k) and empirical secant estimates do not verify the almost-sure envelopes of Theorem 4.2. We now explicitly separate the two complementary stages:
    • Stage 1 (Theorem 4.2 — Analytical Recursion & Planning Radius): Theorem 4.2 supplies the closed-form propagation law used to compose one-step estimates across ∏k∣Ak∣\prod_k |\mathcal{A}_k| schedules into the plug-in planning radius r^k\hat{r}_k (ranking 41 SDXL schedules with Spearman 0.99; Figure 4) and define the time-varying profile uk=max⁡{r^k,10−3}u_k = \max\{\hat{r}_k, 10^{-3}\}.
    • Stage 2 (Proposition 4.3 & Appendix E — Finite-Sample Deployment Certification): Freezing the selected schedule and calibrating on mm fresh paired trajectories certifies the deployment tube q^uk\hat{q} u_k without assuming that (ρ^k,β^k,Σ^k)(\hat{\rho}_k, \hat{\beta}_k, \hat{\Sigma}_k) satisfy almost-sure conditions.
  2. Whole-trajectory certification rather than terminal-only scalar calibration (Proposition 4.3, p. 6; Appendix D.2, Eqs. (29)–(30), p. 19): Our calibration scores the whole-trajectory maximum Si=max⁡1≤k≤N∥Ek(i)∥/ukS_i = \max_{1 \le k \le N} \|E_k^{(i)}\| / u_k. Taking q^=max⁡{1,S(j)}\hat{q} = \max\{1, S_{(j)}\} at j=⌈(m+1)(1−δ)⌉j = \lceil(m+1)(1-\delta)\rceil guarantees simultaneous finite-sample coverage P(∥Ek∥≤q^uk, ∀1≤k≤N)≥1−δ\mathbb{P}(\|E_k\| \le \hat{q} u_k,\ \forall 1 \le k \le N) \ge 1 - \delta across all NN steps under exchangeability, without step-wise independence. (An i.i.d. PAC tolerance bound controlling conditional deployment risk by δ\delta with confidence 1−η1-\eta is also given in Appendix E, Eqs. (32)–(33), p. 20.)
  3. Values of P,R,nP, R, n and behavior of tr⁡Σ^\operatorname{tr}\hat{\Sigma} when n≪dn \ll d (Appendix E, p. 20; Appendix G, p. 27):
    • Sample counts: On SDXL (d=16,384d = 16{,}384), each (k,a)(k, a) cell is profiled on P=128P = 128 probe states pooled across four carrier schedules; on PixArt-Σ\Sigma (d=65,536d = 65{,}536), we use P=64P = 64 probe states across three carriers. At each probe state, stochastic actions use R=n=8R = n = 8 independent replays to debias β^k\hat{\beta}_k (Eq. (31)) and form Σ^k\hat{\Sigma}_k, while deterministic actions use one evaluation (ξk≡0\xi_k \equiv 0). Probe prompts are disjoint from conformal calibration (mm) and evaluation (nevaln_{\text{eval}}).
    • Does rank deficiency (rank⁡(Σ^)≤n−1\operatorname{rank}(\hat{\Sigma}) \le n - 1) cause tr⁡Σ^\operatorname{tr}\hat{\Sigma} to underestimate variance? Mathematically, no: with the n−1n - 1 denominator, linearity of expectation gives E[tr⁡Σ^∣Fk]=1n−1∑i=1nE[∥ξ^(i)∥2∣Fk]=tr⁡Cov⁡(ξk∣Fk)\mathbb{E}[\operatorname{tr}\hat{\Sigma} \mid \mathcal{F}_k] = \frac{1}{n-1}\sum_{i=1}^n \mathbb{E}[\|\hat{\xi}^{(i)}\|^2 \mid \mathcal{F}_k] = \operatorname{tr}\operatorname{Cov}(\xi_k \mid \mathcal{F}_k) for every n≥2n \ge 2 and dd (the n−1n-1 non-zero eigenvalues capture the full expected trace across all dd coordinates, while λmax⁡(Σ^)\lambda_{\max}(\hat{\Sigma}) is biased upward by convexity). Nevertheless, your broader point stands: a finite-sample covariance is not a verified almost-sure sub-Gaussian proxy satisfying Eq. (9). We clarify both points in Appendix E ("Proxy estimates τk2,ℓk\tau_k^2, \ell_k", p. 20) and rely on Proposition 4.3 for deployment certification.

(the rest of the response is in the next response block due to the characters limit.)

AuthorsOpenReview ↗

Response to Reviewer Moda (part 2)

  1. Conformal score distributions, β^kmax⁡\hat{\beta}_k^{\max} tightness, and violation rates at δ=5%,1%,0.1%\delta = 5\%, 1\%, 0.1\% (Table 5, Appendix G, p. 27; Figure 2, p. 8):
Setting & Profilem/nevalm / n_{\text{eval}}r^N\hat{r}_NScore SiS_i (q50 / q95 / q99 / max)Uncalib. Viol. (Term. / Tube)Calib. q^\hat{q} (δ=5/1/0.1%\delta=5/1/0.1\%)Held-out Viol. (δ=5/1/0.1%\delta=5/1/0.1\%)
SDXL / MJHQ (0.80.8-q β^k\hat{\beta}_k)1000 / 1000157.10.690 / 1.393 / 2.137 / 3.1920.0% / 17.4%1.405 / 2.194 / 3.1925.9% / 1.4% / 0.2%
SDXL / sDCI (0.80.8-q β^k\hat{\beta}_k)500 / 500157.10.721 / 1.446 / 2.159 / 3.1110.0% / 18.8%1.454 / 2.180 / ∞\infty3.8% / 1.8% / —
PixArt-Σ\Sigma / MJHQ (0.80.8-q β^k\hat{\beta}_k)500 / 500325.70.976 / 1.660 / 2.390 / 4.7130.0% / 25.0%1.662 / 2.489 / ∞\infty4.2% / 1.8% / —
SDXL / MJHQ (β^kmax⁡\hat{\beta}_k^{\max})1000 / 1000257.8—0.0% / 1.8%1.000 / 1.114 / 1.8681.8% / 1.1% / 0.1%
SDXL / sDCI (β^kmax⁡\hat{\beta}_k^{\max})500 / 500257.8—0.0% / 2.0%1.000 / 1.127 / ∞\infty2.0% / 1.6% / —
PixArt-Σ\Sigma / MJHQ (β^kmax⁡\hat{\beta}_k^{\max})500 / 500437.8—0.0% / 11.0%1.555 / 2.315 / ∞\infty4.2% / 1.8% / —
  • Empirical validation of Theorem 4.2's propagation law: On the terminal error ∥EN∥\|E_N\|, the uncalibrated planning radius r^N\hat{r}_N achieves 100% coverage (0/2,0000/2{,}000 violations; q^term=1.000\hat{q}_{\mathrm{term}} = 1.000) across all held-out trajectories. For the whole 50-step tube on SDXL, simply replacing 0.80.8-quantile β^k\hat{\beta}_k with β^kmax⁡\hat{\beta}_k^{\max} (Figure 2, second bar) achieves 98.2% (MJHQ) and 98.0% (sDCI) uncalibrated whole-tube coverage (q^=1.000\hat{q} = 1.000 at δ=5%\delta = 5\%). On PixArt-Σ\Sigma (11.0%11.0\% under β^kmax⁡\hat{\beta}_k^{\max}) and at δ=1%,0.1%\delta = 1\%, 0.1\%, Proposition 4.3 closes the remaining finite-sample gap, bringing held-out violations to 5.9%/1.4%/0.2%5.9\%/1.4\%/0.2\% on SDXL/MJHQ (within expected binomial variation around a marginal guarantee, SE≈0.7%\text{SE} \approx 0.7\% at 5%5\%; finite q^\hat{q} at δ=0.1%\delta = 0.1\% requires m≥⌈1/δ⌉−1=999m \ge \lceil 1/\delta \rceil - 1 = 999, explaining q^=∞\hat{q}=\infty for m=500m=500).

W2 & Q1: Action Libraries A\mathcal{A}, Budgets Cˉ\bar{C}, Selected Schedules, Bit Widths, Latency, and Memory

Manuscript location: Revised in Section 4.3 (p. 7), Section 5.2 ("Quality at Matched Cost" & "Runtime and GPU Memory", p. 10), New Appendix H (Tables 6–7, pp. 27–28), and New Appendix I (Table 8, p. 29).

In Appendix H, Table 6 (p. 28) specifies the shared RTX 5070 Ti action libraries and budgets Cˉ\bar{C} matched to TMPDiff's planned cost (SDXL, N=50N=50: A={FP16:52.6, W8A8-RN:45.1, W4A4-RN:39.2, W4A4-SR:39.9} ms/step\mathcal{A} = \{\text{FP16}: 52.6,\ \text{W8A8-RN}: 45.1,\ \text{W4A4-RN}: 39.2,\ \text{W4A4-SR}: 39.9\}\text{ ms/step}, Cˉ=2.120 s\bar{C} = 2.120\text{ s}; PixArt-Σ\Sigma, N=20N=20: A={FP16:184.8, W8A8-RN:158.5, W4A4-RN:135.5} ms/step\mathcal{A} = \{\text{FP16}: 184.8,\ \text{W8A8-RN}: 158.5,\ \text{W4A4-RN}: 135.5\}\text{ ms/step}, Cˉ=2.956 s\bar{C} = 2.956\text{ s}). Table 7 (p. 28) reports all schedules (0-based indices), bit widths (W=A), latencies, memory footprints, and LPIPS side by side:

ModelMethodSelected Schedule (0-based steps)FP16Avg. BitsPlan (s)Time: All-Res. / Seq. (s)Mem.: All-Res. / Seq. (GiB)LPIPS ↓\downarrow
SDXLFP16FP16 (0–49)5016.02.6302.403 / —4.90 / —0.000
W4A4-RNW4A4-RN (0–49)04.01.9591.741 / —2.48 / —0.371
W8A8-RNW8A8-RN (0–49)08.02.2542.020 / —3.32 / —0.104
TMPDiffFP16 (0–11), W4A4-RN (12–49)126.92.1201.902 / 2.0076.02 / 4.820.158
Our TQDiffFP16 (0–3), W8A8-RN (4–21), W4A4-RN (22–49)46.42.1191.897 / 2.0547.97 / 4.830.105
PixArt-Σ\SigmaFP16FP16 (0–19)2016.03.6963.677 / —1.18 / —0.000
W4A4-RNW4A4-RN (0–19)04.02.7102.688 / —0.54 / ——
W8A8-RNW8A8-RN (0–19)08.03.1703.114 / —0.80 / ——
TMPDiffW4A4-RN (0–1, 7–19), FP16 (2–6)57.02.9562.927 / 2.9921.52 / 1.190.227
Our TQDiffW4A4-RN (0–2, 13–19), W8A8-RN (3–12)06.02.9402.885 / 3.0061.12 / 0.770.183

At matched latency with fewer average bits and fewer FP16 steps, TQDiff lowers LPIPS over TMPDiff by 33% on SDXL (0.1050.105 vs. 0.1580.158; L2L_2: 30.830.8 vs. 40.540.5) and 19% on PixArt-Σ\Sigma (0.1830.183 vs. 0.2270.227); at Uniform W8A8's budget (2.254 s2.254\text{ s}), TQDiff reduces SDXL LPIPS from 0.1040.104 to 0.0480.048 (Table 8, p. 29).

(the rest is in the next block due to the characters limit.)

AuthorsOpenReview ↗

Response to Reviewer Moda (part 3)

  • Memory–Latency Trade-off (Table 7, p. 28): Keeping all active precisions GPU-resident eliminates weight-reload overhead and preserves 1.90 s1.90\text{ s} SDXL denoising latency, at the cost of higher peak memory on SDXL (7.97 GiB7.97\text{ GiB} vs. 6.02 GiB6.02\text{ GiB} for TMPDiff and 4.90 GiB4.90\text{ GiB} for FP16) because three precisions reside simultaneously; on PixArt-Σ\Sigma, TQDiff skips FP16 altogether and already uses less memory than both baselines (1.12 GiB1.12\text{ GiB} vs. 1.521.52 and 1.18 GiB1.18\text{ GiB}). Conversely, sequential residency keeps only one precision in GPU memory at a time, cutting peak memory to 4.83 GiB4.83\text{ GiB} on SDXL and 0.77 GiB0.77\text{ GiB} on PixArt-Σ\Sigma, while amortized weight-transfer overhead is +0.16 s/img+0.16\text{ s/img} at batch size 11 (2.054 s2.054\text{ s} total) and drops to +0.02 s/img+0.02\text{ s/img} at batch size 88.

Q3: CIFAR-10 FID vs. L2L_2/LPIPS and Single-Backend Pinned-Schedule Diagnostic

Manuscript location: Revised in Section 5.2 (p. 9) and New Appendix J (p. 29).

  • Selected 100-step schedule: Listed step-by-step in Appendix J (p. 29), the CIFAR schedule uses 25 full-precision steps (0–21 and 25–27), 27 SR steps, 20 RN+TCEC steps, and 28 RN steps (50 W4A8 and 25 W4A6 quantized steps).
  • Single-backend 50,000-sample reproduction check: Evaluating the pinned schedule and W4A8-RN on a single quantized backend against a shared FP32 reference across 50,000 samples confirms the exact phenomenon: W4A8-RN reproduces clean-FID 5.90 ([5.82,5.98][5.82, 5.98]), LPIPS 0.11010.1101, and terminal L2=6.646L_2 = 6.646, while the pinned TQDiff schedule yields clean-FID 15.78 ([15.62,15.94][15.62, 15.94]) alongside substantially lower terminal L2L_2 (2.425) and LPIPS (0.0643). (Table 1 retains the original 14.36014.360 value, as this diagnostic uses FP32 full-precision steps and a refitted compensator on the unified backend.)
  • Why FID diverges from L2L_2/LPIPS: Spectral analysis shows that late low-precision and compensated steps leave a small, high-frequency residual (36.9%36.9\% of error energy at spatial frequencies ≥8\ge 8 cycles/image vs. 20.1%20.1\% for W4A8-RN) that barely affects pixel L2L_2 but shifts Inception feature covariance. Confirming this, a diagnostic replacing only the final 20 steps with full precision restores clean-FID to 5.91 ([5.82,6.01][5.82, 6.01], L2=2.130L_2 = 2.130). We report this higher-cost intervention strictly as a diagnostic showing that our bound controls trajectory L2L_2 fidelity rather than distributional FID.

Q4: Correlation of r^N\hat{r}_N with Realized Error vs. Cheap Proxies, and SDXL-Turbo + sDCI

Manuscript location: Revised in Section 5.2 ("Does the Radius Rank Schedules?", p. 10; Table 1 caption, p. 9), New Figure 4 (p. 10), Appendix F.4 (Table 4 caption, p. 23), and New Appendix I (p. 28).

  • Schedule ranking across 41 SDXL schedules (Figure 4, left, p. 10; Appendix I, p. 28): Across 41 diverse schedules evaluated on 70 fresh prompts, r^N\hat{r}_N achieves a Spearman rank correlation of 0.99 ([0.98,0.99][0.98, 0.99], Kendall τ=0.91\tau = 0.91) with realized mean terminal L2L_2, 0.95 ([0.87,0.96][0.87, 0.96], τ=0.82\tau = 0.82) with q95 terminal L2L_2, and 0.99 with mean LPIPS. By contrast, cheap proxies that ignore multi-step amplification ∏j>kρj\prod_{j>k}\rho_j perform poorly: planned cost reaches 0.55, average bit width 0.54, FP16 step count 0.32, and unpropagated bias sum ∑kβ^k\sum_k \hat{\beta}_k only 0.21.
  • SDXL-Turbo + sDCI (N=4N=4): We fully agree that at N=4N=4 steps, the horizon is too short for multi-stage precision scheduling to separate from TMPDiff (L2L_2 36.13236.132 vs. 36.55636.556, LPIPS 0.16550.1655 vs. 0.16410.1641). We now explicitly note this parity in Section 5.2 (p. 9) and the captions of Tables 1 (p. 9) & 4 (p. 23).

All three concerns I raised are resolved by the new runs. I decided to raise my rating from 2 to 3 as well.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (part 1)

We sincerely thank all three reviewers for their rigorous, constructive evaluations and for recognizing both the theoretical and practical contributions of our work—including verifying our nonlinear L2L_2-MGF error propagation bound "line by line" (Reviewer Moda), highlighting the clean supermartingale formulation and efficient dynamic programming (DP) planner (Reviewer dQ5C), and noting the theoretical grounding and empirical gains across settings (Reviewer 7MAr). All three reviews converge on two central questions: (1) how finite-sample parameter estimates bridge to formal certification at deployment, and (2) whether TQDiff's quality gains hold under strictly matched computational budgets with full cost disclosure.

We have revised the manuscript with targeted additions and clarifications marked in blue, while preserving the original narrative, proofs, and benchmark tables. Below we summarize our two core improvements.

1. Two-Stage Certification Architecture: Analytical Propagation + Finite-Sample Deployment Certification

All three reviewers rightly emphasized the essential distinction between Theorem 4.2 (which holds under valid per-step envelopes (ρk,βk,Σk)(\rho_k, \beta_k, \Sigma_k)) and guarantees computed from finite-sample empirical estimates (such as 0.80.8-quantile biases β^k\hat{\beta}_k and empirical secants ρ^k\hat{\rho}_k). Rather than weakening the paper's certification claim, our revision formalizes a complete two-stage certification pipeline (Section 4.2, p. 6)—mirroring standard analytical-bound plus order-statistic certification in certified ML (e.g., Randomized Smoothing):

  • Stage 1 — Irreplaceable Role & Empirical Tightness of Theorem 4.2 (Section 4, pp. 5–6): Finite-sample calibration alone cannot search an exponential space (∏k=0N−1∣Ak∣=450≈1030\prod_{k=0}^{N-1}|\mathcal{A}_k| = 4^{50} \approx 10^{30} schedules on SDXL) nor determine the per-step tube profile uku_k. Theorem 4.2 supplies the closed-form nonlinear propagation law that composes one-step empirical profiles into an estimated (plug-in) planning radius r^N\hat{r}_N with Spearman rank correlation 0.99 against realized terminal L2L_2 and LPIPS across 41 schedules (Figure 4, p. 10). Moreover, Theorem 4.2's propagation law is remarkably accurate in practice:
    • For the terminal error ∥EN∥\|E_N\| targeted by Theorem 4.2, the uncalibrated planning radius r^N\hat{r}_N attains 100% coverage (0/2,0000/2{,}000 violations; q^term=1.000\hat{q}_{\mathrm{term}} = 1.000) across all held-out SDXL and PixArt-Σ\Sigma trajectories.
    • Even for the stricter whole-trajectory tube (all NN steps simultaneously), replacing the 0.80.8-quantile β^k\hat{\beta}_k with the probe-state maximum β^kmax⁡\hat{\beta}_k^{\max} immediately achieves 98.2% (MJHQ) and 98.0% (sDCI) uncalibrated whole-tube coverage on SDXL (q^=1.000\hat{q} = 1.000 at δ=5%\delta = 5\%; Table 5, p. 27).
  • Stage 2 — Finite-Sample Deployment Certification (Proposition 4.3, Section 4.2, p. 6; Appendix D.2, p. 19; Appendix E, Eqs. (32)–(33), p. 20): Once the planner selects a schedule, we freeze it together with uk=max⁡{r^k,10−3}u_k = \max\{\hat{r}_k, 10^{-3}\} and certify the whole-trajectory score Si=max⁡1≤k≤N∥Ek(i)∥/ukS_i = \max_{1 \le k \le N} \|E_k^{(i)}\|/u_k on mm independent calibration trajectories. Proposition 4.3 guarantees simultaneous finite-sample coverage P(∥Ek∥≤q^uk, ∀k≤N)≥1−δ\mathbb{P}(\|E_k\| \le \hat{q} u_k,\ \forall k \le N) \ge 1 - \delta under exchangeability—closing any finite-sample estimation gap without requiring (ρ^k,β^k,Σ^k)(\hat{\rho}_k, \hat{\beta}_k, \hat{\Sigma}_k) to satisfy almost-sure worst-case bounds—and Appendix E (Eqs. (32)–(33)) provides the complementary i.i.d. PAC tolerance bound with confidence 1−η1 - \eta.
  • Layered Terminology: We preserve "Certified Error Bounds" in the Title because the two-stage pipeline delivers genuine finite-sample certified deployment bounds, while inside the text we strictly separate: (i) analytical high-probability bound (Theorem 4.2 under valid envelopes), (ii) planning / plug-in radius (r^N\hat{r}_N during DP search), and (iii) calibrated / finite-sample certified deployment tube (q^uk\hat{q} u_k or qηukq_{\eta} u_k).

(the rest of the general response is in the next response block: General Response (part 2) due to the characters limit.)

AuthorsOpenReview ↗

General Response (part 2) + Summary Roadmap of Revisions

2. Controlled Cost-Matched Evaluations and Full Schedule Disclosure

To separate scheduling gains from compute budget, kernel choice, and hardware differences, we report controlled comparisons on one NVIDIA RTX 5070 Ti GPU using a shared execution engine (FP16, SVDQuant-style rank-32 group-64 NVFP4 W4A4, and INT8 W8A8 with cached embeddings and CUDA graphs) in Section 5.2 (pp. 9–10), Appendix H (Tables 6–7, pp. 27–28), and Appendix I (Table 8, p. 29):

Setting & SplitMethodSelected Schedule (0-based steps)FP16 StepsAvg. Bits (W/A)Denoising Time (s)Memory: All-Res. / Seq. (GiB)LPIPS ↓\downarrowPaired Δ\DeltaLPIPS [95% CI]
SDXL (N=50N=50, Cˉ=2.120\bar{C}=2.120 s, 200200 prompts)TMPDiffFP16 (0–11), W4A4-RN (12–49)126.9 / 6.91.9026.02 / 4.820.158—
Our TQDiffFP16 (0–3), W8A8-RN (4–21), W4A4-RN (22–49)46.4 / 6.41.8977.97 / 4.830.105−0.053-0.053 [−0.059,−0.046-0.059, -0.046]
PixArt-Σ\Sigma (N=20N=20, Cˉ=2.956\bar{C}=2.956 s, 500500 prompts)TMPDiffW4A4-RN (0–1, 7–19), FP16 (2–6)57.0 / 7.02.9271.52 / 1.190.227—
Our TQDiffW4A4-RN (0–2, 13–19), W8A8-RN (3–12)06.0 / 6.02.8851.12 / 0.770.183−0.044-0.044 [−0.048,−0.039-0.048, -0.039]

At matched denoising latency and with fewer average bits and fewer FP16 steps, TQDiff reduces LPIPS by 33% on SDXL and 19% on PixArt-Σ\Sigma. Furthermore, across six matched budgets on SDXL (cost gaps <12<12 ms; Table 8, p. 29), TQDiff outperforms the TMPDiff family at every budget (ΔLPIPS\Delta\text{LPIPS} from −0.028-0.028 to −0.069-0.069, all 95% CIs excluding zero), and lowers LPIPS from 0.3690.369 to 0.1300.130 against an adapted PTQD SNR-allocation baseline on the identical two-precision library (Appendix K, Table 9, p. 30).

The summary roadmap of each response/revision is listed in the following table:

Summary of Revisions in the Manuscript (highlighted in blue)

Reviewer ConcernReviewersRevised Manuscript Location (in Blue)
Two-stage certification: analytical bound vs. planning radius vs. deployed certificationModa W1/Q2, dQ5C W1, 7MAr W1Abstract (p. 1); §1 (pp. 1–2); §4.1 (Assumption 4.1 & Thm. 4.2, pp. 5–6); New §4.2 / Prop. 4.3 (p. 6); App. D.1–D.2 (pp. 18–19); App. E (p. 20)
Calibration parameters (P,R,nP, R, n), tr⁡Σ^\operatorname{tr}\hat{\Sigma} unbiasedness, β^kmax⁡\hat{\beta}_k^{\max}, & violation ratesModa W1/Q2, dQ5C W1, 7MAr W1§5.2 & Revised Fig. 2 (p. 8); App. E (p. 20); New App. G / Table 5 (p. 27)
Action libraries A\mathcal{A}, budgets Cˉ\bar{C}, selected schedules, bit widths, latency & memory residencyModa W2/Q1, dQ5C W4, 7MAr W2a§4.3 (p. 7); §5.2 (pp. 9–10); New App. H / Tables 6–7 (pp. 27–28); New App. I / Table 8 (p. 29)
History dependence of βk(a),Σk(a)\beta_k(a), \Sigma_k(a), DP state formulation & on-policy profile stabilitydQ5C W2§4.3 (p. 7); App. F.3 (pp. 21–22); New App. I ("On-policy profile stability check", p. 29)
Correlation of r^N\hat{r}_N with realized error vs. cheap proxies & SDXL-Turbo/sDCI parityModa Q4, dQ5C W3§5.2 & Table 1 (pp. 9–10); New Fig. 4 (p. 10); App. F.4 / Table 4 (p. 23); New App. I (p. 28)
CIFAR-10 FID vs. L2L_2/LPIPS & pinned-schedule 50k diagnostic; PixArt-Σ\Sigma mixed scheduleModa Q3, dQ5C W3§5.2 (pp. 9–10); New App. H / Table 7 (p. 28); New App. J (p. 29)
Comparison with PTQD, TAC-Diffusion, AccuQuant & scope on flow-matching DiTs7MAr W2b, Q§2.1–2.3 (pp. 3–4); §5.2 (p. 10); §6 Conclusion (p. 10); New App. K / Table 9 (p. 30)
Presentation fixes (swapped Abstract/Intro percentages, Fig. 2 RN vs. rNr_N, indexing & notation)dQ5C W5Abstract & §1 (pp. 1–2); §3.1 (p. 4); §4.3 (p. 7); §5.2 & Revised Fig. 2 (p. 8); App. B–E (pp. 15–20)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.