acceptodds
Under review as a conference paper at ICLR 2027

Accelerating Diffusion Models with Physical Velocity Evolution

Abstract

Diffusion and flow-based generative models deliver strong visual synthesis quality but incur substantial inference cost during iterative sampling. Existing caching methods typically rely on high-dimensional feature differences that can entangle content and noise, overlooking the physical-time velocity dynamics underlying the generative ODE. These representation-level differences do not directly characterize the physical evolution of the latent state, which is governed by a velocity field over continuous sampling time. Instead, grounding cache reuse in physical-time velocity dynamics connects approximation errors to the latent updates they perturb, with the actual physical time intervals determining how velocity errors translate into trajectory deviations. Based on this observation, we introduce Trajectory-Aligned Flow Caching (TAFC), a training-free framework that operationalizes physical-time velocity dynamics through three complementary components. Physical-Time velocity Extrapolation predicts upcoming latent updates by extrapolating velocity dynamics over physical sampling intervals. Curvature-Aware Reuse Gating uses velocity curvature to assess whether the observed dynamics support continued local extrapolation. Self-Calibrating Error Feedback measures a posteriori approximation errors at refresh steps and adapts reuse tolerances, calibrating physical trajectory predictions against model-evaluated updates. Together, these components connect physical-time velocity dynamics, prediction reliability, and online error correction in a closed-loop caching framework. Extensive experiments on large-scale image and video generative models demonstrate that TAFC achieves favorable efficiency–fidelity trade-offs over existing training-free caching methods.

Primary area:
generative models
Keywords:
Diffusion Models, Inference Acceleration, Model Caching, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2LCgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem that existing caching methods decide reuse based on feature differences, which do not match how the sampling trajectory actually evolves. It replaces the zero-order residual reuse of TeaCache-style caching with first-order linear extrapolation in actual time, and decides when to refresh using an extrapolation-magnitude threshold, residual changes, a maximum skip count and PID-adapted thresholds. On FLUX.1-dev, HunyuanVideo and Wan2.1, it shows higher reference fidelity than TeaCache, TaylorSeer and SeaCache at similar latency.

Critical strengths

(1) The evaluations is solid, and the fidelity gains are clear and consistent across models.

(2) The method is lightweight and easy to use: it requires no training, only stores two residuals and a few scalars, and can be plugged directly into existing residual-caching framewokrs.

Critical weaknesses

1、Limited novelty. The method linearly extrapolates the residual using the trend of the last two residuals. In essence, this applies TaylorSeer’s first-order extrapolation to TeaCache’s residual cache, together with a set of heuristic refresh rules (threshold-based refresh and standard PID tuning).

(1) Related work on first-order residual extrapolation: AB-Cache [1], PrediT [2], etc.

(2) Related work on residual reuse with adaptive refresh decisions: EasyCache [3], DiCache [4], etc.

2、 Issues with related-work citations and comparisons.

(1) For example, the citations of event-camera pretraining and optical-flow papers seem unrelated to this work. Please check the related-work citations more carefully.

(2) More related work should be cited and compared, including AB-Cache, PrediT, EasyCache, MeanCache [5] and DiCache. Among them, AB-Cache is the closest method and was also evaluated on HunyuanVideo and FLUX.

3、Key parameters (e.g., the thresholds and the curvature threshold) are missing, and the ablation does not separate the individual contributions of the rules. It is therefore unclear whether the performance gain mainly comes from the extrapolation itself, from a particular refresh rule, or from the PID controller.

[1] AB-Cache: Training-Free Acceleration of Diffusion Models via Adams-Bashforth Cached Feature Reuse

[2] PrediT: Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers

[3] EasyCache: Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching

[4] DiCache: Let Diffusion Model Determine Its Own Cache

[5] MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference

Questions for authors

pls see weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 2LCg [1/2]

We sincerely thank you for your constructive feedback and for recognizing the solid empirical evaluations and lightweight, plug-and-play nature of TAFC. Below, we address your concerns by adding direct conceptual and experimental comparisons with AB-Cache [1], PrediT [2], EasyCache [3], DiCache [4], and MeanCache [5] and providing the full hyperparameter values and rule-level ablation breakdown.

1. Novelty beyond Combining TalyorSeer with Teacache, and More Comparisons (Weakness 1 & Weakness 2(2))

While TaylorSeer forecasts feature evolution and TeaCache uses feature-derived signals to guide reuse, TAFC starts from the physical-time dynamics of the generative ODE: an error in the model velocity perturbs the latent update in proportion to the sampling-time interval. This motivates approximating the reusable Transformer residual over actual sampling time and controlling how far that approximation is extended. Specifically, TAFC forecasts from two model-evaluated residuals using their physical-time separation, assesses reuse through the extrapolated correction and variation, and adapts subsequent budgets using forecast errors observed at refreshes. The residual is the operational quantity, while preserving the velocity-driven sampling trajectory is the underlying objective. Our contribution is this integrated prediction-and-control design motivated by physical-time error propagation, rather than first-order extrapolation alone.

Our contribution lies in integrating physical-time residual forecasting with reliability gating and error-feedback control, rather than introducing numerical extrapolation or adaptive caching individually. For more clarification with related work and comparisons, we have cited all five methods the reviewer mentions and clarified these distinctions in the revised manuscript. The quantitative comparisons are presented below. We did not evaluate AB-Cache because its official implementation was unavailable at the time of our experiments. For MeanCache, the publicly released code supported only Z-Image and Qwen-Image at the time, rather than the backbones evaluated in our study; we therefore provide a conceptual comparison.

New Comparison Table 1: FLUX.1-dev Evaluation with DiCache [4] and PrediT [2] (TAFC rows from Table 1)

Acceleration RegimeMethodLatency (s) ↓\downarrowTFLOPs ↓\downarrowPSNR (dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
∼2.00×\sim 2.00\timesDiCache (δ=50\delta = 50) [4]10.3782219.820.7730.187
∼2.00×\sim 2.00\timesPrediT (∼2.0×\sim 2.0\times) [2]9.8976423.420.8410.151
∼2.00×\sim 2.00\timesTAFC (δ=0.04\delta = 0.04)9.3770329.870.9410.051
∼3.00×\sim 3.00\timesDiCache (δ=70\delta = 70) [4]6.7252818.140.7430.252
∼3.00×\sim 3.00\timesPrediT (∼3.0×\sim 3.0\times) [2]6.4849620.350.7720.243
∼3.00×\sim 3.00\timesTAFC (δ=0.15\delta = 0.15)5.9445121.350.8200.196

New Comparison Table 2: Wan2.1-1.3B Video Evaluation with EasyCache [3] (TAFC row from Table 3)

Acceleration RegimeMethodLatency (s) ↓\downarrowTFLOPs ↓\downarrowPSNR (dB) ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow
∼2.00×\sim 2.00\timesEasyCache (τ=10%\tau = 10\%) [3]155.81571221.140.7380.246
∼2.00×\sim 2.00\timesTAFC (δ=0.03\delta = 0.03)154.21555625.560.8620.092

Across all settings, TAFC achieves lower latency and TFLOPs while outperforming DiCache by +10.05 dB (∼2×\sim 2\times) and +3.21 dB (∼3×\sim 3\times), PrediT by +6.45 dB (∼2×\sim 2\times) and +1.00 dB (∼3×\sim 3\times) on FLUX.1-dev, and EasyCache by +4.42 dB PSNR on Wan2.1-1.3B.


2. Removing Unrelated Citations (Weakness 2(1))

Thank you for pointing this out. Our intention is to provide broader context on differential-equation-based visual modeling and efficient trajectory computation, rather than to present these works as prior methods for diffusion caching. However, we recognize that grouping event-camera pretraining and optical flow within this paragraph made their relevance unclear and distracted from the closest caching literature. We have therefore removed these citations from this paragraph and refocus the discussion on AB-Cache [1], PrediT [2], EasyCache [3], DiCache [4], and MeanCache [5], clarifying their methodological relationships to TAFC and including the additional empirical comparisons where available.

AuthorsOpenReview ↗

Response to Reviewer 2LCg [2/2]

3. Complete Hyperparameters and Rule-by-Rule Ablation Breakdown (Weakness 3)

  • Complete Hyperparameter Specification: All models share a unified controller configuration, while the table threshold δ\delta (tafc_thresh) and, where enabled, the maximum consecutive cache run MM determine the acceleration operating point:

  • Base Thresholds (ebase,κbasee_{\text{base}}, \kappa_{\text{base}}): In Eq. (9), δ\delta directly sets ebase=δe_{\text{base}} = \delta (0.04 / 0.15 / 0.30 for FLUX; 0.02 / 0.15 for HunyuanVideo; 0.03 / 0.15 for Wan2.1), and κbase=2δ\kappa_{\text{base}}=2\delta.

  • Time Modulation & Warmup/Cutoff: In Eq. (9), βk=0.7+1.3(k/N)\beta_k = 0.7 + 1.3(k/N) modulates the budget across normalized step progress k/N∈[0,1]k/N \in [0, 1], and τ(tk)=1\tau(t_k)=1. In Eq. (10), all experiments fix warmup_steps = 3 (full evaluations at k=0,1,2k=0,1,2, with the first valid two-anchor forecast error collected at k=2k=2 for calibration) and cutoff_steps = 49 (Ccut={49}\mathcal{C}_{\text{cut}} = \{49\}). No maximum consecutive-cache-run limit is imposed in the approximately 2×2\times settings. For FLUX, M=4M=4 and M=6M=6 are used in the approximately 3×3\times and 4×4\times settings, respectively; for HunyuanVideo and Wan2.1, M=3M=3 is used in the approximately 3×3\times setting.

  • PID Controller: Fixed at KP=0.4,KI=0.05,KD=0.2K_P = 0.4, K_I = 0.05, K_D = 0.2 with G∈[0.3,3.0]G \in [0.3, 3.0] in Eqs. (14)–(15).

  • Isolating Individual Contributions in the Ablation Study: Our ablations distinguish the effects of extrapolation, refresh rules, and PID feedback: “w/o First-Order Extrapolation” replaces the residual forecast with zero-order reuse; “w/o Adaptive Threshold” removes the prescribed time-dependent budget modulation; and “w/o Closed-Loop Control” disables PID updates by fixing GG constantly 1. To further isolate the refresh rules, we added “w/o κk\kappa_k Alone,” which removes only the anchor-variation gate while retaining Eq. 7 and PID feedback, and “Fixed Uniform + First-Order,” which retains the predictor but uses periodic refreshes. Full TAFC achieves higher PSNR at comparable measured refresh ratios, indicating that the gains do not come from extrapolation alone. We have clarified these implementations and included the additional controls in the revised manuscript. The threshold settings and their mapping to the base budgets are provided in the hyperparameter response above.

[1] AB-Cache: Training-Free Acceleration of Diffusion Models via Adams-Bashforth Cached Feature Reuse

[2] PrediT: Predict to Skip: Linear Multistep Feature Forecasting for Efficient Diffusion Transformers

[3] EasyCache: Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching

[4] DiCache: Let Diffusion Model Determine Its Own Cache

[5] MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer 2LCg

As the deadline of discussion period is approaching, we would kindly like to inquire if you would get a chance to review our response and if there are any remaining questions we can address.

Your insights, both the constructive suggestions and areas of contention, have been crucial for us. We have done our best to address each point and clarify any misunderstandings. We are truly keen to have a constructive dialogue with you to refine our work further.

Best regards,

Authors of Submission 23200

Reviewer_exLbOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper makes diffusion Transformer faster by training-free caching method, TAFC. It uses two recent computed residual to make next residual by linear extrapolation. So it doesn't need to reuse the last residual as it is. And it uses real sampling time gap, not step number. The gate decides when to compute again. In refresh step, it puts forecast error to PID controller to change the budget. The authors compare with TeaCache, TaylorSeer and SeaCache on FLUX.1-dev, HunyuanVideo and Wan2.1. It achieved positive score on PSNR and VBench. Also latency is much more lower than 50-step baseline.

Critical strengths
  1. I think the biggest advantage is that PSNR is more higher when the time is almost same. For example, in Wan2.1 PSNR is 21.17 but SeaCache is only 17.52. I think this is very practical contribution.
  2. The authors marked evaluation very detail. The authors seem to be very honest for showing the result that TAFC is not the best.
  3. The theory part is very simple, and the claim is not too much. Also the performance got worse when any part is removed.
Critical weaknesses
  1. Eq. 7 and Eq. 8 uses same two anchor. However, I couldn't find the result that Eq. 8 helps more than Eq. 7 for predicting forecast error. Figure 7 only shows the combined controller. So it's unclear that why Eq. 8 is needed.
  2. Tables only show δ for TAFC. However, the controller also needs more setting, like base budget and PID gain. I couldn't find how δ is matched with these. So it is hard to reproduce the result only with the paper.
  3. Section 3.1 says it extrapolates Transformer residual, not latent trajectory. However, title and introduction say it extrapolates physical velocity. The connection to velocity only comes in appendix. So reader can think it predicts sampler velocity directly.
Questions for authors
  1. Could you report the full setting for one image setting and one video setting please? I couldn't find how δ is matched with base budget, time modulation, calibration, cutoff, maximum cache run and PID gain.
  2. Is Eq. 8 still helpful for predicting refresh error after Eq. 7 is used? Epost is measured at every refresh, so is there result that shows why both signals are needed with this?
  3. How is "w/o Physical Time" in Figure 7 made? If it uses step index instead of time gap, two forecast is same when step gap is uniform. The paper says schedule is generally nonuniform, but I couldn't find how much the schedule of FLUX and video models is nonuniform.
4 comments
AuthorsOpenReview ↗

Response to Reviewer exLb [1/2]

We sincerely thank you for your positive assessment and for highlighting our practical PSNR and VBench improvements, honest reporting, and clean theoretical formulation. Below we address your three questions in detail.

1. Full Hyperparameter Settings for Image and Video Models & Mapping of δ\delta (Weakness 2 & Question 1)

For the approximately 2×2\times settings, FLUX.1-dev uses δ=0.04\delta=0.04 and HunyuanVideo uses δ=0.02\delta=0.02, with ebase=δe_{\mathrm{base}}=\delta and κbase=2δ\kappa_{\mathrm{base}}=2\delta. Both use βk=0.7+1.3k/N\beta_k=0.7+1.3k/N and τ(tk)=1\tau(t_k)=1, three initial full-evaluation steps, and a forced refresh at step 49 of the 50-step schedule. The first valid two-anchor forecast error is collected at k=2k=2 and used to calibrate the target error. Neither approximately 2×2\times setting imposes a maximum consecutive-cache-run limit; the ck≥Mc_k\ge M refresh condition is disabled in these settings. The PID gains are (KP,KI,KD)=(0.4,0.05,0.2)(K_P,K_I,K_D)=(0.4,0.05,0.2), with the feedback gain clamped to G∈[0.3,3.0]G\in[0.3,3.0]. FLUX generates 1024×10241024\times1024 images with guidance scale 3.5, while HunyuanVideo generates 65-frame videos at 832×480832\times480.

Below is the complete side-by-side configuration table for FLUX.1-dev (Image) and HunyuanVideo / Wan2.1-1.3B (Video):

Hyperparameter / SymbolFLUX.1-dev (2×2\times / 3×3\times / 4×4\times)HunyuanVideo (2×2\times / 3×3\times)Wan2.1-1.3B (2×2\times / 3×3\times)
Inference Steps NN / Resolution50 steps / 1024×10241024 \times 102450 steps / 832×480832 \times 480, 65f50 steps / 832×480832 \times 480, 65f
Threshold δ\delta (tafc_thresh =ebase= e_{\text{base}})0.04 / 0.15 / 0.300.02 / 0.150.03 / 0.15
Maximum Consecutive Cache Run MMNot enforced / 4 / 6Not enforced / 3Not enforced / 3
Calibration Region Ccalib\mathcal{C}_{\text{calib}} (warmup)3 steps (k=0,1,2k = 0, 1, 2)3 steps (k=0,1,2k = 0, 1, 2)3 steps (k=0,1,2k = 0, 1, 2)
Cutoff Region Ccut\mathcal{C}_{\text{cut}} (cutoff)Step 49Step 49Step 49
Time Modulation βk\beta_k0.7+1.3(k/N)0.7 + 1.3(k/N)0.7+1.3(k/N)0.7 + 1.3(k/N)0.7+1.3(k/N)0.7 + 1.3(k/N)
PID Gains (KP,KI,KD)(K_P, K_I, K_D)(0.4,0.05,0.2)(0.4, 0.05, 0.2)(0.4,0.05,0.2)(0.4, 0.05, 0.2)(0.4,0.05,0.2)(0.4, 0.05, 0.2)
Gain Clamping [Gmin⁡,Gmax⁡][G_{\min}, G_{\max}][0.3,3.0][0.3, 3.0][0.3,3.0][0.3, 3.0][0.3,3.0][0.3, 3.0]

We have added this complete table to Appendix A.2 to ensure effortless reproducibility.


2. Regarding Eq. 7 and Eq. 8 (Weakness1 & Question2)

We use Eqs. 7 and 8 to control two aspects of residual extrapolation. One is the extent of the current forecast, and the other is the variation of the evaluated anchors from which it is constructed. Specifically, expanding Eq. 7 we have $\widehat E_k = \frac{|t_k-t_b|}{|t_b-t_a|} \frac{|r_b-r_a|}{|r_b|+\varepsilon}$. It explicitly scales the anchor difference by the current extrapolation horizon. When the anchors are fixed, its value grows as time continues. Therefore, it restricts how far the prediction extends from the latest evaluated residual. Eq. 8 instead checks the relative magnitude and directional variation of the anchor pair without scaling it by the current horizon.

This distinction matters in two cases. First, a short extrapolation horizon can keep Eq. 7 below budget even when the anchors exhibit pronounced variation. Eq. 8 provides an additional condition in this case and introduces an anchor update. Second, even small anchor variation does not prevent a prediction from moving far from the latest evaluation during long time reuse. Thus Eq. 7 supplies this control depending on horizon. We therefore check both the variation underlying the forecast and the extent to which that forecast is used. Refresh-time error feedback calibrates the tolerances of both checks.

As you mentioned, the two share the same anchors and can be correlated, with overlapping decisions in some regimes. Their distinction motivates us including both. We have quantified the incremental benefit of Eq. 8 through an ablation that removes only this condition at matched measured refresh ratios.We have also updated the ablation figure in the revised manuscript to include this comparison. The “w/o κk\kappa_k alone” ablation directly tests this question by retaining Eq. 7 and the EpostE_{\mathrm{post}} feedback while removing only Eq. 8. Full TAFC achieves higher PSNR at comparable measured refresh ratios, supporting the incremental value of Eq. 8 within the controller.

AuthorsOpenReview ↗

Response to Reviewer exLb [2/2]

3. regarding the velocity and residual(Weakness 3)

Our residual formulation and velocity-based motivation are connected by the analysis provided in the paper. Proposition 3.1 characterizes the accuracy of residual forecasting, while Appendix A.1 explains how residual approximation errors affect the sampling trajectory. We have made this connection explicit at the end of Section 3.1.

Section 3.1 defines the cached Transformer residual as r(t)=Fθ(h(t),t)−h(t)r(t)=F_\theta(h(t),t)-h(t). Proposition 3.1 shows that, under the stated local smoothness conditions, two-anchor forecasting admits an O(Δ2)O(\Delta^2) residual approximation bound, compared with the O(Δ)O(\Delta) bound for constant reuse.

We introduce Vθ(x,r,t)V_\theta(x,r,t) in Appendix A.1. This mapping includes the computation downstream of residual insertion and produces the model output used by the sampler. Assumption A.1, Eq. (17), gives

\le L_r\|\widehat r-r\|.$$ A single-step illustration makes this relationship concrete. Starting from the same latent state, the fully evaluated and cached first-order updates differ by at most $$\|x_{k+1}^{\mathrm{cache}}-x_{k+1}^{\mathrm{full}}\| \le \Delta t_k L_r\|\widehat r-r\|, \qquad \Delta t_k=|t_{k+1}-t_k|.$$ Thus, residual approximation changes the sampler output through downstream sensitivity, and the sampling interval determines its contribution to the latent update. This is the mechanism represented by the $L_r\Delta t_k\eta_k$ term in Appendix B.2, Eq. (39). Proposition A.2 then accumulates these contributions into a time-weighted trajectory-deviation bound. In summary, the existing analysis follows a chain, from residual forecasting accuracy, to its effect on the sampler output, and finally to the resulting trajectory perturbation. Residual forecasting is our computational mechanism for approximating the velocity-driven sampling process. --- ### 4. Implementation of `w/o Physical Time` and Schedule Nonuniformity (Question 3) * **How nonuniform are the schedules of FLUX, HunyuanVideo, and Wan2.1?** While the step index $k \in \{0, \dots, 50\}$ is uniform, modern rectified-flow models (FLUX.1-dev, HunyuanVideo, Wan2.1) employ **time-shifted flow-matching schedules** $\sigma_t = \frac{s \cdot t}{1 + (s - 1)t}$ (e.g., `shift` $s = 5.0$ or resolution-dependent shifting). Under `shift = 5.0`, the physical time increments $\Delta t_k = |t_{k+1} - t_k|$ are highly nonuniform across sampling: steps near $t = 1$ (high noise) have compressed physical intervals, whereas steps near $t = 0$ have much wider intervals. Furthermore, when one or more steps are cached between refreshes, the anchor span $|t_b - t_a|$ and extrapolation span $|t_k - t_b|$ cover variable numbers of nonuniform intervals. * **How `w/o Physical Time` is implemented**: In Figure 7 (`tafc_ablation4_integer_steps.py`), `w/o Physical Time` replaces the actual physical scheduler timestamps in Eqs. (3)–(4) with uniform integer step counts ($\Delta t = 1.0$). Because the true shifted schedule is strongly nonlinear, ignoring physical time distorts the finite-difference slope $d_b$ and degrades PSNR in Figure 7. We have explicitly documented the scheduler shift factor and this implementation in Section 4.4.

Thank you for the detailed response. My questions are answered by the hyperparameter table in Appendix A.2, the ablation without κ_k, the new paragraph in Section 3.1, and the explanation of "w/o Physical Time". However, after reading other reviews, I'm still thinking about the novelty. Also new baseline results like MagCache, SpeCa and PrediT are only in the response, not in the paper tables. I will decide my final rating after the discussion with other reviewers.

AuthorsOpenReview ↗

Response to Reviewer exLb

Dear Reviewer exLb,

Thank you very much for your thoughtful feedback, constructive engagement, and for acknowledging the practical contributions, honest evaluation, and clean formulation of our work.

Regarding your concern that the newly introduced baseline results (MagCache, SpeCa, PrediT, DiCache, and EasyCache) were only present in the author response: We have fully integrated these comprehensive baseline comparisons and their corresponding data tables into the revised manuscript (specifically in Appendix D) to ensure complete transparency.

We truly appreciate your positive rating and your valuable questions during the review process, which have significantly strengthened the completeness and presentation of our paper.

Reviewer_61YaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TAFC, a training-free step-level cache for diffusion and flow transformers, arguing that feature-difference criteria ignore the physical-time dynamics of the generative ODE. TAFC linearly extrapolates the Transformer residual r(t)=Fθ(h(t),t)−h(t)r(t)=F_\theta(h(t),t)-h(t) from the last two fully evaluated steps using sampler-time spacing, refreshes when the extrapolated correction or the change between the two anchors exceeds adaptive budgets, and tunes these budgets with a PID controller driven by the forecast error observed at refreshes. A local O(Δ2)O(\Delta^2) error bound supports the forecast. On FLUX.1-dev, HunyuanVideo, and Wan2.1-1.3B at roughly 2–4× speedup, TAFC reports higher PSNR/SSIM/LPIPS against the 50-step output than TeaCache, TaylorSeer, and SeaCache, along with component ablations.

Critical strengths
  1. Consistent fidelity gains on three large backbones at no higher measured latency. As reported in Tables 1–3, TAFC attains the best PSNR, SSIM, and LPIPS in every speed regime on FLUX.1-dev, HunyuanVideo, and Wan2.1, e.g., 29.87 vs. 26.29 dB PSNR over SeaCache on FLUX at about 2.2×. Since training-free caches are judged mainly by their fidelity–speed frontier, consistent gains across image and video models are the main argument for acceptance.
  2. Closed-loop calibration from a free error signal is simple and cheap. Each refresh reveals the true residual at no extra cost, and TAFC uses it to adapt reuse tolerances online while storing only two residual anchors. In the FLUX ablation, removing this feedback costs about 2.5–5 dB PSNR at comparable refresh ratios, which makes the design a practical drop-in for TeaCache-style residual caching.
Critical weaknesses
  1. The velocity framing does not match the method, and novelty is unclear. All signals act on the Transformer residual r(t)r(t), the same representation-level quantity TeaCache caches, and physical time only replaces step counts in the two-anchor extrapolation. The "curvature" κk\kappa_k in Eq. 8 is just a magnitude-and-cosine change between the two anchors. FoCa already extrapolates features from the last two steps as an ODE, MagCache gates reuse by residual magnitude, and SpeCa forecasts then verifies, yet none is discussed or compared. Comparing against them at matched refresh ratios, and adding a variant that extrapolates vθv_\theta itself, would test the velocity premise.
  2. Inconsistent operating points and undefined hyperparameters make the core comparison unverifiable. At the δ values of Table 1's 2× rows, the Fig. 8 heatmaps average about 24.6 full evaluations per sample for TAFC versus 19.9 for SeaCache, yet Table 1's 703 TFLOPs imply only about 21 for TAFC. If Fig. 8 reflects the evaluated runs, the headline 2× comparison gives TAFC more compute than SeaCache. The threshold δ itself is never defined, and βk\beta_k, τ(tk)\tau(t_k), MM, and the PID gains have no reported values or tuning protocol. A full hyperparameter table, tuning on prompts disjoint from the test sets, and measured refresh ratios per entry are needed.
Questions for authors

Clarification (answers could change my assessment)

  1. Please state how the table threshold δ relates to ebasee_{\text{base}} and κbase\kappa_{\text{base}}, give the per-model values of βk\beta_k, τ(tk)\tau(t_k), CcalibC_{\text{calib}}, CcutC_{\text{cut}}, MM, and the PID gains together with the prompts used to tune them, and say whether the baselines received a comparable tuning budget.
  2. Please state whether Fig. 8 and Table 1 come from the same runs, and report for every table entry the measured refresh ratio and the GPU used for latency, since both H200 and A100 are mentioned.
  3. Every TAFC point in the HunyuanVideo ablation (Fig. 7(b), 32–40 dB) exceeds the best HunyuanVideo PSNR in Table 2 (29.24 dB) at comparable refresh ratios; does the 20-prompt ablation use a different resolution, frame count, or reference?
  4. Please describe how each Fig. 7 variant is implemented and add two controls at matched refresh ratios: removing κk\kappa_k alone, and the same first-order forecast with a fixed uniform refresh interval.
  5. As caching becomes more aggressive, TAFC's HPSv3 rises from 10.66 to 10.91, above the 50-step reference, while PSNR drops from 29.87 to 19.41 dB; does this reflect a systematic style shift favored by the reward model, and would a blind human preference study confirm that quality is preserved?
4 comments
AuthorsOpenReview ↗

# Response to Reviewer 61Ya [1/3]

We sincerely thank you for your rigorous review and for recognizing TAFC’s consistent fidelity gains across all three backbones and our lightweight closed-loop error feedback mechanism. Below, we address your concerns and five clarification questions with new baseline comparisons (FoCa, MagCache, SpeCa), full hyperparameter tables, execution log audits, and a 15-participant blind human preference study.

1. Velocity Framing, Novelty, and Comparisons with FoCa[1], MagCache[2], and SpeCa[3] (Weakness 1)

  • Why extrapolate the Transformer residual r(t)r(t) rather than the sampler velocity vθv_\theta directly? Our residual formulation and velocity-based motivation are connected by the analysis provided in the paper. Proposition 3.1 characterizes the accuracy of residual forecasting, while Appendix A.1 explains how residual approximation errors affect the sampling trajectory. Section 3.1 defines the cached Transformer residual as r(t)=Fθ(h(t),t)−h(t)r(t)=F_\theta(h(t),t)-h(t). Proposition 3.1 shows that, under the stated local smoothness conditions, two-anchor forecasting admits an O(Δ2)O(\Delta^2) residual approximation bound, compared with the O(Δ)O(\Delta) bound for constant reuse.

We introduce Vθ(x,r,t)V_\theta(x,r,t) in Appendix A.1. This mapping includes the computation downstream of residual insertion and produces the model output used by the sampler. Assumption A.1, Eq. (17), gives

\le L_r\|\widehat r-r\|.$$ A single-step illustration makes this relationship concrete. Starting from the same latent state, the fully evaluated and cached first-order updates differ by at most $$\|x_{k+1}^{\mathrm{cache}}-x_{k+1}^{\mathrm{full}}\| \le \Delta t_k L_r\|\widehat r-r\|, \qquad \Delta t_k=|t_{k+1}-t_k|.$$ Thus, residual approximation changes the sampler output through downstream sensitivity, and the sampling interval determines its contribution to the latent update. This is the mechanism represented by the $L_r\Delta t_k\eta_k$ term in Appendix B.2, Eq. (39). Proposition A.2 then accumulates these contributions into a time-weighted trajectory-deviation bound. In summary, the existing analysis follows a chain, from residual forecasting accuracy, to its effect on the sampler output, and finally to the resulting trajectory perturbation. Residual forecasting is our computational mechanism for approximating the velocity-driven sampling process. * **Distinctions and Quantitative Comparisons with FoCa, MagCache, and SpeCa:** 1. FoCa [1] combines feature forecasting with numerical calibration, whereas TAFC uses refresh-time forecast errors to adapt subsequent reuse budgets through PID feedback. 2. MagCache [2] bases caching decisions on calibrated residual magnitude ratios, whereas TAFC combines physical-time residual extrapolation with magnitude-and-direction gating and online error feedback. 3. SpeCa [3] uses speculative forecasting and verification, whereas TAFC evaluates reuse gates from cached residual anchors without additional Transformer evaluations and updates its budgets using errors observed at refresh steps. We have cited these papers and clarify these methodological distinctions in the revised manuscript. We conducted new experiments comparing TAFC against open-source **MagCache** and **SpeCa** on FLUX.1-dev (DrawBench) at $\sim 2.00\times$ and $\sim 3.00\times$ acceleration (TAFC rows from Table 1). For **FoCa**, we provide a conceptual comparison because its official implementation is not publicly available at the time of our experiments. | Acceleration Regime | Method | Latency (s) ↓ | TFLOPs ↓ | PSNR (dB) ↑ | SSIM ↑ | LPIPS ↓ | |:---:|:---|---:|---:|---:|---:|---:| | $\sim 2.00\times$ | MagCache ($\epsilon=50$) | 10.12 | 788 | 21.68 | 0.817 | 0.176 | | $\sim 2.00\times$ | SpeCa ($\tau_0=7$) | 9.85 | 771 | 24.63 | 0.862 | 0.128 | | $\sim 2.00\times$ | **TAFC ($\delta=0.04$)** | **9.37** | **703** | **29.87** | **0.941** | **0.051** | | $\sim 3.00\times$ | MagCache ($\epsilon=70$) | 6.64 | 512 | 19.59 | 0.752 | 0.244 | | $\sim 3.00\times$ | SpeCa ($\tau_0=10$) | 6.37 | 508 | 20.74 | 0.783 | 0.237 | | $\sim 3.00\times$ | **TAFC ($\delta=0.15$)** | **5.94** | **451** | **21.35** | **0.820** | **0.196** | ---
AuthorsOpenReview ↗

Response to Reviewer 61Ya [2/3]

2. Definition of δ\delta, Full Hyperparameter Table, and Tuning Protocol (Weakness 2 & Question 1)

The table threshold δ\delta sets ebase=δe_{\mathrm{base}}=\delta, with κbase=2δ\kappa_{\mathrm{base}}=2\delta. Across all three models, we use βk=0.7+1.3k/N\beta_k=0.7+1.3k/N and τ(tk)=1\tau(t_k)=1. The first three steps (k=0,1,2k=0,1,2) are fully evaluated, with the first valid two-anchor forecast error collected at k=2k=2 for calibration; step 49 is also fully evaluated. No maximum consecutive-cache-run limit is imposed in the approximately 2×2\times settings. For FLUX, M=4M=4 and M=6M=6 are used in the approximately 3×3\times and 4×4\times settings, respectively; for HunyuanVideo and Wan2.1, M=3M=3 is used in the approximately 3×3\times setting. All three models use PID gains (KP,KI,KD)=(0.4,0.05,0.2)(K_P,K_I,K_D)=(0.4,0.05,0.2), with G∈[0.3,3.0]G\in[0.3,3.0]. We selected the shared controller settings on 30 manually written validation prompts, disjoint from the evaluation prompts, and then held them fixed. For operating-point selection, TAFC and each baseline are allocated 20 candidate configurations per model on the same validation set, selecting configurations closest to the target latency budgets.

Model BackboneRegimeδ (tafc_thresh)Maximum Cache Run MMWarmupCutoffβₖ ModulationPID (Kₚ, Kᵢ, Kᴅ)G BoundsRefresh Ratio
FLUX.1-dev (50 steps)~2×0.04Not enforced3 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]42%–44%
FLUX.1-dev (50 steps)~3×0.1543 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]26%–28%
FLUX.1-dev (50 steps)~4×0.3063 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]20%–22%
HunyuanVideo (50 steps)~2×0.02Not enforced3 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]46%-48%
HunyuanVideo (50 steps)~3×0.1533 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]30%-32%
Wan2.1-1.3B (50 steps)~2×0.03Not enforced3 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]48%-50%
Wan2.1-1.3B (50 steps)~3×0.1533 stepsStep 490.7+1.3(k/N)0.7+1.3(k/N)(0.4, 0.05, 0.2)[0.3, 3.0]30%-32%

3. Reconciling Figure 8 and Table 1 & Hardware Consistency (Weakness 2 & Question 2)

Thank you for catching this discrepancy, we identify the exact cause:

  • Why Figure 8 and Table 1 differed: All quantitative results in Table 1 are evaluated with warmup_steps = 3, under which TAFC (δ=0.04\delta=0.04) averages approximately 21 full evaluations (703 TFLOPs, 9.37s latency), requiring strictly less computation than SeaCache (δ=0.3\delta=0.3, 782 TFLOPs, 9.82s). In contrast, Figure 8 is plotted from both warmup_steps = 6 for fair comparison, as indicated by the six solid white initial columns at steps 0–5 in Figs. 8(a) and 8(c). Increasing the warm-up length from three to six steps increases TAFC’s average total number of full evaluations from approximately 21 to 24.6, while SeaCache’s remains approximately 19.9 in these runs because it already performs dense refreshes during the initial six steps. Changing the warm-up length affected TAFC’s subsequent refresh schedule, while SeaCache’s schedule remained largely unchanged in these runs.

For completeness, we have clarified the experimental configurations in A.2. All image-generation experiments are conducted on NVIDIA A100 GPUs, while all video-generation experiments are conducted on NVIDIA H200 GPUs. Within each experiment, all compared methods are benchmarked on the same GPU hardware.

AuthorsOpenReview ↗

Response to Reviewer 61Ya [3/3]

4. HunyuanVideo PSNR in Figure 7(b) vs. Table 2 (Question 3)

Following SeaCache’s ablation configuration, we conducted the component ablations using the same resolution (832×480832\times480), frame count (65 frames), and 50-step reference settings as Table 2. During evaluation on the full 944-prompt VBench set, we observed extremely low reference-fidelity scores for certain samples across different caching methods, including PSNR values below 10 dB. We therefore screened the evaluation samples and randomly selected 20 prompts from the retained set, using the same subset for all ablation variants. The resulting curves illustrate the relative contributions of the components on this subset. However, because the subset was selected after screening for reference fidelity, its higher absolute PSNR values should not be directly compared with the full-set averages in Table 2.


5. Implementation of Figure 7 Variants and Two New Controls (Question 4)

We implement the four original ablations as follows:

  1. w/o First-Order Extrapolation: Replace the first-order residual forecast with zero-order reuse, r^(tk)=r(tb)\hat r(t_k)=r(t_b).
  2. w/o Adaptive Threshold: Remove the time-dependent modulation factor βk\beta_k from the reuse budgets.
  3. w/o Closed-Loop Control: Disable PID updates and fix the feedback gain at G=1G=1.
  4. w/o Physical Time: Use sampling-step indices instead of scheduler timestamps to calculate the extrapolation intervals.

Following your suggestion, we added two controls on FLUX.1-dev:

  1. w/o κk\kappa_k Alone: Remove only the κk\kappa_k refresh condition, retaining the extrapolation-risk gate E^k\hat E_k and PID feedback.
  2. Fixed Uniform + First-Order: Retain physical-time first-order forecasting but replace adaptive refresh decisions with a fixed periodic refresh schedule.

As shown in the updated figure, full TAFC achieves higher PSNR than these variants at comparable refresh ratios. In particular, removing κk\kappa_k lowers fidelity, while using a fixed uniform refresh schedule produces a larger drop despite retaining first-order forecasting.


6. HPSv3, Style Shift, and Blind Human Preference Study (Question 5)

Yes, we interpret this trend as a style shift favored by the reward model. Our human ratings are consistent with this interpretation, showing higher color/lighting scores but slightly lower detail-fidelity scores for TAFC. However, these preliminary results do not establish that the shift is systematic across all prompts.

We conducted a blind randomized evaluation with 15 participants, comparing the 50-step baseline with TAFC at δ=0.3\delta=0.3 (∼4×\sim4\times acceleration) on FLUX.1-dev. Participants reported their overall preference and rated four dimensions on a seven-point scale (1 = worst, 7 = best).

Overall preference: 8 participants (53.3%) preferred TAFC, 6 (40.0%) preferred the baseline, and 1 (6.7%) reported a tie.

Evaluation DimensionBaseline (50 steps)TAFC (δ=0.3\delta=0.3)
Detail Fidelity6.075.93
Color & Lighting5.936.27
Text Alignment6.076.13
Visual Naturalness6.006.13

Participants rated the baseline slightly higher for fine details and TAFC higher for color and lighting, with similar ratings for text alignment and visual naturalness. Overall, 8 of 15 participants preferred TAFC, suggesting that lower PSNR does not necessarily imply lower perceived quality. We have included these results in Appendix C.

[1] FoCa: Forecast Then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers.

[2] MagCache: MagCache: Fast Video Generation with Magnitude-Aware Cache.

[3] SpeCa: SpeCa: Accelerating Diffusion Transformers with Speculative Feature Caching.

AuthorsOpenReview ↗

Kind reminder

Dear Reviewer 61Ya

As the deadline of discussion period is approaching, we would kindly like to inquire if you would get a chance to review our response and if there are any remaining questions we can address.

Your insights, both the constructive suggestions and areas of contention, have been crucial for us. We have done our best to address each point and clarify any misunderstandings. We are truly keen to have a constructive dialogue with you to refine our work further.

Best regards,

Authors of Submission 23200

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their responses to your reviews. Please take some time to carefully examine their rebuttals and engage in the discussion by sharing your thoughts, remaining concerns, or follow-up questions.

As the discussion deadline is approaching, I would appreciate your timely participation to allow sufficient time for further exchanges with the authors.

Thank you for your time and valuable contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.