acceptodds
Under review as a conference paper at ICLR 2027

CalibCache: Cache-to-Full Feature Dynamics Calibration for Diffusion Transformers

Abstract

Diffusion Transformers generate high-quality images, but iterative sampling remains computationally expensive. Feature caching reduces the number of full network evaluations, yet its approximation error can become more severe under short sampling schedules, where fewer full-compute steps and a coarser sampling grid can make history-based extrapolation less reliable. In this paper, we propose CalibCache, a lightweight offline calibration method that corrects systematic discrepancies between cached and full-compute feature dynamics. Specifically, CalibCache calibrates feature acceleration, defined as the second-order finite difference across sampling steps. Using only a single pair of full-compute and naive-cache trajectories at each setting, CalibCache fits scalar acceleration maps offline by centered least squares. These maps are then applied during sampling without additional full-network evaluations or online optimization. Across FLUX.1-dev and Qwen-Image on DrawBench and DiT-XL/2 on ImageNet-256, CalibCache consistently improves fidelity or generation quality over TaylorSeer under 20-step sampling at cache intervals . At , it reduces FLUX.1-dev LPIPS from to while achieving a denoising speedup over full-compute sampling, and lowers DiT-XL/2 FID from to relative to TaylorSeer.

Primary area:
generative models
Keywords:
Diffusion Transformers, Feature Caching, Inference Acceleration, Offline Calibration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_RNbBOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CalibCache, an offline calibration method for feature caching in diffusion Transformers. For each deployment configuration it builds one paired full-compute and naive-cache trajectory and uses the discrepancy between their second-order residual-feature differences ("feature acceleration") as supervision. A centered least-squares estimator fits step- and branch-specific scalar maps in closed form. At inference, these frozen scalars correct cached residuals at the first cache step of each interval, combined with corrected writeback and an isolated full-anchor predictor history, adding no extra full-network evaluation and no online optimization.

Critical strengths
  • Very low deployment cost with substantive gains.

  • Requires only one full-compute/naive-cache trajectory pair per deployment configuration.

  • Evaluates FLUX.1-dev, Qwen-Image, and DiT-XL/2, with useful ablations and calibration sensitivity analyses.

Critical weaknesses
  • The experiments cover only 20-step sampling, without validation under other sampling schedules. Although the paper is motivated by the potential unreliability of history-based extrapolation under short sampling schedules, it does not show how this issue and the benefits of calibration vary with the sampling grid. Quality–latency comparisons across different numbers of sampling steps would help establish the method’s applicability and validate its motivation for short sampling schedules.

  • The paper compares only against caching methods and does not report full-compute baselines with fewer sampling steps on FLUX.1-dev or Qwen-Image. Consequently, it remains unclear whether CALIBCACHE provides better generation quality than simply reducing the sampling schedule from 20 steps to 7–9 steps at matched latency.

  • The latency benefits vary substantially across models and fall considerably short of the FLOPs reductions. For example, on DiT-XL/2 at N=5N=5, a 3.31× FLOPs reduction yields only a 1.81× denoising speedup. The paper should further analyze prediction, cache read/write, and other execution overheads to explain these cross-model differences and clarify the operating conditions under which substantial practical gains can be achieved.

  • The authors are encouraged to include a focused analysis of GOC[1] and ERTACache[2] in the related work section. Discussing these representative studies would help trace the development of diffusion feature caching and clarify the technical positioning of this submission.

[1]. Qiu, Junxiang, et al. "Accelerating diffusion transformer via gradient-optimized cache." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[2]. Peng, Xurui, et al. "Ertacache: Error rectification and timesteps adjustment for efficient diffusion." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

See the Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer RNbB (1/4): Sampling Grids and Calibration Benefits

We thank the reviewer for recognizing our method's low deployment cost, single-trajectory calibration design, and multi-model evaluation. We address the comments with sampling-grid diagnostics, comparable-latency baselines, runtime measurements, and a clarified comparison with prior work.

1. Sampling grids, extrapolation reliability, and quality–latency trade-offs

CalibCache remains effective at 30–50 steps, while the tested 20-step configurations show greater local extrapolation error and larger PSNR, LPIPS, and ImageReward gains over TaylorSeer than the 50-step configurations.

(a) Local extrapolation across sampling grids

Both models exhibit higher local prediction error at 20 than at 50 steps, even when predictions do not feed back into sampling.

We evaluate the same 32 DrawBench prompt–seed pairs at 20, 30, 40, and 50 steps. TaylorSeer receives features from a full-compute trajectory only at the anchors prescribed by N = 5, warmup-3, and predicts each interval's first cache step with maximum order two. Predictions never advance sampling. Errors are normalized by each sample's teacher-feature energy before averaging.

Model20 steps30 steps40 steps50 steps
FLUX.1-dev0.0364900.0182440.0120990.015685
Qwen-Image0.0072780.0022010.0015650.000851

On both models, the paired 95% bootstrap confidence intervals for the 20-minus-50-step error difference lie entirely above zero. Appendix D.1 details the diagnostic and statistical procedures.

(b) Effectiveness under longer schedules

At 50 steps on FLUX.1-dev, CalibCache achieves the best means on all five quality/fidelity metrics among the compared caching methods and is faster than TaylorSeer and HiCache, with a 3.54× denoising speedup.

We evaluate 30, 40, and 50 steps on both models using DrawBench-200 at 1024×1024 with matched prompt–seed pairs, retaining one calibration trajectory pair per model–schedule configuration. The FLUX 50-step results follow. All caching methods use N = 5, with warmup-3 for TaylorSeer, HiCache, and CalibCache. Fidelity uses the matched Native-50 outputs.

MethodTime (s)↓Speedup↑PSNR↑SSIM↑LPIPS↓CLIP↑IR↑
Full compute16.9131.00×∞1.00000.000031.2150.9799
FORA3.8864.35×14.100.60480.499331.3110.9097
TaylorSeer5.4903.08×16.520.69080.362031.3240.9902
HiCache5.3503.16×16.860.70460.339231.2771.0003
CalibCache4.7813.54×17.460.70470.333431.5351.0029

Complete results and protocols for both models at 30–50 steps appear in Appendix C, Tables 16–21.

(c) More pronounced calibration gains under shorter schedules

CalibCache improves all five metrics over TaylorSeer in every listed setting. PSNR, SSIM, LPIPS, and ImageReward gains are larger at 20 than at 50 steps on both models.

The table reports CalibCache minus TaylorSeer, with the LPIPS sign reversed so that every positive value indicates improvement. Fidelity uses each grid's corresponding full-compute reference.

ModelStepsPSNR gain (dB)SSIM gainLPIPS reductionCLIP gainIR gain
FLUX.1-dev500.936950.013890.028630.211190.01277
FLUX.1-dev400.974560.013300.034300.140390.01386
FLUX.1-dev301.090190.019600.044500.078970.01624
FLUX.1-dev201.832100.082610.133630.060400.12406
Qwen-Image503.527120.092220.130980.389390.04781
Qwen-Image403.841690.101400.156060.627220.12343
Qwen-Image304.625540.130470.181391.204610.30824
Qwen-Image205.475100.123930.198941.846300.54488

Together, the local extrapolation diagnostic and generation results support our motivation for calibration under short schedules. Complete longer-schedule baseline and latency results are provided in Appendix C.

AuthorsOpenReview ↗

Response to Reviewer RNbB (2/4): Reduced-Step Full-Compute Baselines

2. Reduced-step full-compute baselines at comparable latency

At comparable denoising latencies, CalibCache-20 improves the mean results on all three fidelity metrics and ImageReward over the reduced-step full-compute baselines on both FLUX.1-dev and Qwen-Image.

We bracket CalibCache's 20-step, N = 5, warmup-3 H800 denoising latency with integer-step Native baselines, using 7/8 steps for FLUX and 8/9 for Qwen. All methods use matched DrawBench-200 prompts and initial noise at 1024×1024, with identical model weights and guidance. Fidelity is measured against each model's Native-20 outputs.

ModelMethodTime (s)↓PSNR↑SSIM↑LPIPS↓CLIP↑IR↑
FLUX.1-devNative-72.348216.17740.65530.459531.36910.7572
FLUX.1-devNative-82.681916.51080.67180.429831.47750.8327
FLUX.1-devCalibCache-202.464419.10140.67350.429231.94610.8989
Qwen-ImageNative-85.388515.17750.63270.445732.57320.8433
Qwen-ImageNative-96.058615.28320.63860.429932.53480.8841
Qwen-ImageCalibCache-205.629519.24600.71480.304132.47041.0008

On FLUX, all five improvements over Native-7 have paired 95% bootstrap intervals excluding zero. Against Native-8, PSNR improves by 2.5906 dB, also supported by the paired intervals. On Qwen, CalibCache improves all three fidelity metrics and ImageReward over both baselines, reducing LPIPS by 0.14159 versus Native-8 and 0.12578 versus Native-9. The corresponding ImageReward gains are 18.7% and 13.2%, while CLIP decreases by only 0.32% and 0.20%, respectively.

Appendix D.2 provides the complete evaluation and timing protocols, along with paired confidence intervals.

AuthorsOpenReview ↗

Response to Reviewer RNbB (3/4): FLOPs Reduction and Practical Speedup

3. The gap between FLOPs reduction and denoising speedup

Cache steps retain low-FLOP operations with non-negligible runtime. Differences in backbone workloads and overhead outside forward calls explain the variation in practical speedup across models.

(a) Retained cache-step costs

For DiT-XL/2 at 20 steps, N = 5, warmup-3, full evaluations decrease from 20 to six, yielding approximately 3.31× FLOPs reduction. The remaining 14 cache steps retain modulation, gating, residual updates, and cache operations. Matched-schedule measurements over 18 paired inputs on H800 at batch size 1 give the following cumulative costs:

ComponentTime (ms/image)Share
Direct-reuse cost31.8961.25%
CalibCache increment20.1838.75%

Even without correction, retained operations incur substantial latency after the main attention/MLP recomputation is skipped. CalibCache adds execution costs from prediction, history maintenance, and corrected-feature writeback.

In the same DiT setting, FORA uses four full evaluations without extra warmup, versus six for TaylorSeer, HiCache, and CalibCache. This smaller full-compute budget contributes to its speed advantage, while its generation quality is lower in this comparison. Its 4.96× FLOPs reduction yields 2.53× denoising speedup, consistent with the costs retained even by direct reuse.

(b) Cross-model runtime composition

The following shares use each configuration's total CalibCache denoising time as the denominator.

ModelResolutionFull forwardCache forwardOutside forward
FLUX.1-dev1024×102489.39%10.59%0.02%
Qwen-Image1024×102466.52%11.65%21.83%
DiT-XL/2256×25648.46%38.02%13.52%

DiT's 256×256 setting has fewer image tokens and smaller attention/MLP workloads than the 1024×1024 FLUX/Qwen settings, whereas AdaLN modulation-parameter projections do not shrink proportionally with token count. Retained operations therefore occupy a larger runtime share. Qwen also incurs substantial costs outside forward calls, including guidance computation, coordination between conditional/unconditional passes, and sampler updates.

The reported CalibCache speedups are 2.86×, 2.27×, and 1.81× on FLUX, Qwen, and DiT. Caching generally provides greater practical gains when attention/MLP recomputation dominates full-compute runtime and retained operations and outside-forward costs are relatively small. Appendix D.3 provides the complete protocols and breakdowns from diagnostics conducted separately from the original latency measurements.

Reviewer_QrXPOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an offline calibration method for accelerating diffusion Transformers named CalibCache. It fits step- and branch-specific scalar coefficients from paired full-compute and naive-cache trajectories. During inference, these coefficients correct cached residuals without requiring additional full-network evaluations. Reported results show better speed and visuall quality, i.e., at cache interval N=5N=5, the method reports 2.86× denoising speedup on FLUX.1-dev and reduces DiT-XL/2 FID from 16.39 with TaylorSeer to 10.37.

Critical strengths

This paper shows strength in following aspects:

  1. The calibration is simple and lightweight. The method estimates step- and branch-specific scalar coefficients through closed-form least-squares fitting, avoiding the need for additional full-network evaluations during inference.
  2. The calibration parameters have a small memory footprint, making the approach practical for accelerating diffusion transformers.
Critical weaknesses

It shows following weakness:

  1. The method relies primarily on a single paired calibration trajectory per model configuration. Its robustness across prompts, random seeds, noise trajectories, and cache configurations is not clearly evaluated.

  2. The improvements may arise partly from the corrected feature writeback and isolated predictor history, rather than the proposed calibration coefficients alone. More controlled ablations are needed to show the contribution of each component.

  3. The reported inference speedups exclude the offline calibration cost. The practical end-to-end efficiency gains should be reported, which should include quantifying calibration time and analyzing amortization over repeated inference workloads.

  4. The CalibCache achieves FID improvements on DiT-XL/2. However, gains on text-to-image models are more mixed and involve additional inference overhead. This should be discussed to support the claimed advantages of efficiency and visual quality.

Questions for authors

Based on weakness, the authors should clarify following questions:

  1. Please clarify the robustness across prompts, random seeds, noise trajectories, etc.
  2. More ablation studies are required to show the contribution of each component.
  3. Provide more experiments on inference speed.
  4. More comparisions across different different models to support the claimed advantages of efficiency and visual quality.
7 comments
AuthorsOpenReview ↗

Response to Reviewer QrXP (1/4): Robustness to Calibration Samples, Generation Noise, and Cache Configurations

We thank the reviewer for recognizing our lightweight calibration design, closed-form solution, and low parameter-storage overhead, and for raising important questions about robustness, component contributions, and practical inference efficiency. In response, we have added comprehensive robustness evaluations, controlled component ablations, end-to-end efficiency measurements, and cross-model experiments on Sana. We respond point by point below.

1. Robustness Across Calibration Prompts, Calibration Seeds, Generation Noise, and Cache Configurations

The tested calibration prompts and noise seeds yield similar performance. Within each tested cache configuration, frozen coefficients maintain similar Native-relative fidelity across generation noise realizations.

(a) Calibration prompts

Ten independent single-prompt calibrations show limited variation on the same evaluation set.

The original Appendix A.8 (Table 13) evaluates ten FLUX.1-dev coefficient sets, each fitted with a different COCO prompt and fixed calibration noise. On identical DrawBench-200 inputs, LPIPS is 0.43368±0.009560.43368\pm0.00956 and ImageReward is 0.89312±0.010330.89312\pm0.01033 (mean ± sample standard deviation across the ten evaluation-set aggregates).

(b) Calibration random seeds

Changing only the calibration noise yields similar results across all five metrics.

We fit three coefficient sets with the same prompt and different noise seeds (FLUX.1-dev, 20 steps, N=5N=5, warmup-3), then evaluate them on identical DrawBench-200 prompts and generation noise.

Calibration seedPSNR ↑SSIM ↑LPIPS ↓CLIP ↑ImageReward ↑
36415919.10140.67350.429231.94610.8989
36416019.14730.67110.431331.88150.8775
36416119.06550.66870.436931.93720.8947

(c) Generation noise trajectories

One frozen coefficient set maintains similar evaluation metrics across three tested generation seed configurations.

We retain the paper's coefficients and vary only generation noise on DrawBench-200. All three evaluated schedules are reported below; each seed is its listed base plus the prompt index. Fidelity uses Native outputs with the same prompt and initial noise.

Generation seed basePSNR ↑SSIM ↑LPIPS ↓CLIP ↑ImageReward ↑
31415919.10140.67350.429231.94610.8989
161803319.21700.67400.434231.94230.8809
141421319.25590.67360.421731.82590.9216

(d) Cache configurations

Under a second cache configuration, calibration changes again yield similar metrics.

We refit for N=3N=3, warmup-2, and separately test three calibration prompts, three calibration noise seeds, and three generation seed sets. Each fit still uses one trajectory pair; the corresponding coefficients remain frozen during generation-noise tests.

Protocols and complete results are in Appendix F.1, Tables 30–31. These measurements describe sensitivity within the tested settings.

AuthorsOpenReview ↗

Response to Reviewer QrXP (2/4): Contributions of Fitted Coefficients, Corrected Writeback, and Predictor-History Isolation

2. Contributions of Corrected Writeback, Predictor-History Isolation, and Fitted Calibration Coefficients

Controlled ablations support contributions from writeback and history isolation, while coefficient fitting provides further gains on both DiT and Qwen.

(a) Corrected writeback and history isolation

Writeback and history isolation each improve FID with the other components fixed.

The original Table 6 uses DiT-XL/2, ImageNet-256, DDIM-20, CFG 1.5, N=5N=5, and warmup-3. All variants share coefficients fitted from one trajectory pair, full-compute positions, and correction positions, with 50K images evaluated per variant.

Application policyFID ↓sFID ↓IS ↑Precision ↑Recall ↑
No writeback + isolated history18.6328313.06961124.419910.597020.53530
Writeback + ordinary history14.635789.01988133.635960.624580.58280
Writeback + isolated history10.369156.40612151.756440.668280.57380

Rows 1 and 3 isolate the contribution of retaining corrected features for later cache steps; rows 2 and 3 isolate the contribution of preventing direct feedback of corrected values into predictor history.

(b) Additional contribution of fitted coefficients

With writeback and history isolation retained, fitting reduces DiT FID by 7.65% and Qwen LPIPS by approximately 31.9%; all five Qwen paired improvements have 95% bootstrap intervals excluding zero.

We replace only the fitted coefficients with fixed s=0s=0, retaining writeback, isolated history, and all computation positions. Setting s=0s=0 yields unit-coefficient extrapolation along f−pf-p. Both DiT variants use 50K images under the same generation and evaluation protocol; the fitted row retains the original Table 6 values.

Coefficient settingFID ↓sFID ↓IS ↑Precision ↑Recall ↑
Fixed s=0s=011.228286.79978146.341830.654400.58210
Offline-fitted ss10.369156.40612151.756440.668280.57380

Fitting also improves sFID, IS, and Precision, with a small Recall decrease. This comparison uses one global generation seed; cross-seed FID uncertainty was not estimated.

The Qwen comparison uses 20 steps, N=5N=5, warmup-3, and identical DrawBench-200 prompt-seed pairs. Full settings are in Appendix F.2.

Qwen-Image variantPSNR ↑SSIM ↑LPIPS ↓CLIP ↑ImageReward ↑
Fixed s=0s=014.76220.62550.446431.46870.7402
Offline-fitted ss19.24600.71480.304132.47041.0008

On Qwen, fitting increases PSNR by 4.4838 dB and ImageReward from 0.7402 to 1.0008. Paired statistics use 10,000 bootstrap resamples of the 200 matched prompt-seed pairs. The relatively close performance of fixed extrapolation on DiT does not carry over to Qwen, supporting the additional value of fitting across the tested models.

Appendix F.2 gives the protocols and fixed-coefficient comparisons (Tables 32–33); Table 34 reports paired differences and 95% confidence intervals.

AuthorsOpenReview ↗

Response to Reviewer QrXP (3/4): Offline Calibration Cost, End-to-End Efficiency, and Amortization

3. Offline Calibration Cost, End-to-End Efficiency, and Amortization

Based on measured end-to-end latencies, the estimated cumulative speedups are 2.65× for FLUX and 2.12× for Qwen when generating 1,000 images, and 1.62× for DiT when generating 50K images, with the one-time preparation cost included.

Measurements use one H800 GPU with 20 steps, N=5N=5, and warmup-3. Preparation is timed once per model at its fixed configuration, from process startup to exit, including initialization, model loading, trajectory collection, fitting, saving, validation, and cleanup.

Per-image latency uses GPU-resident models and batch size 1, from input initialization to returning the image on the CPU. It includes conditioning encoding where applicable, denoising, and VAE decoding. Native and CalibCache alternate execution order on ten matched inputs; results are post-warmup means.

ModelOne-time preparation (s)Native (s/image)CalibCache (s/image)End-to-end speedupBreak-even (images)
FLUX.1-dev94.496.967672.538682.74×22
Qwen-Image239.0912.794655.787052.21×35
DiT-XL/263.060.244180.149321.64×665

Let CC be preparation cost, LNativeL_{\mathrm{Native}} and LCalibCacheL_{\mathrm{CalibCache}} the per-image latencies, and MM the number of generated images. The estimated break-even count and cumulative speedup including preparation are:

Mbreak-even=⌈CLNative−LCalibCache⌉.M_{\text{break-even}}=\left\lceil\frac{C}{L_{\mathrm{Native}}-L_{\mathrm{CalibCache}}}\right\rceil.

S(M)=MLNativeC+MLCalibCache.S(M)=\frac{M L_{\mathrm{Native}}}{C+M L_{\mathrm{CalibCache}}}.

Detailed measurement conditions and results are in Appendix F.3, Table 35.

Reviewer_pz7kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CalibCache, a lightweight offline calibration method for accelerating Diffusion Transformers. The authors use paired full-computation and naive-cache trajectories to estimate scalar coefficients that calibrate second-order feature differences through centered least squares. During inference, these coefficients are used to correct cached residual features without additional full-network evaluations.

Experiments on FLUX.1-dev, Qwen-Image, and DiT-XL/2 demonstrate improvements in generation quality and reconstruction fidelity over existing caching methods, including TaylorSeer and HiCache.

Critical strengths

Strengths

  1. The paper is well-written, with clear and fluent explanations. The proposed method is simple and computationally lightweight.
  2. The experimental results are promising, particularly on Qwen-Image and DiT-XL/2. The generated images also demonstrate good visual quality.
  3. The ablation studies are relatively comprehensive, including comparisons of supervision targets, coefficient choices, correction policies, and calibration-set sizes.
Critical weaknesses

Weaknesses

  1. According to Equation (7) in Appendix A.4, the proposed correction can be rewritten as r~τ=f+(1−sg)(f−p)\tilde{r}_\tau=f+(1-s_g)(f-p), which is equivalent to a two-anchor linear extrapolation with a learned coefficient. Furthermore, Table 11 shows that a single shared coefficient already achieves performance close to the step-by-branch coefficients, particularly in terms of SSIM and LPIPS. This makes me wonder how much of the improvement actually comes from the proposed acceleration calibration, rather than simply learning appropriate extrapolation coefficients. Although the authors compare against L2P in Appendix A.9, they also acknowledge that their adaptation is not sufficiently validated to serve as a reliable baseline. A more direct comparison with conventional learned linear predictors under the same two-anchor formulation would be helpful.
  2. Table 4 shows that acceleration supervision substantially outperforms residual and velocity supervision. However, I find this large difference somewhat surprising. According to Table 9, if the preceding full-computation and cached features coincide, all three supervision formulations should yield equivalent online updates under centered least-squares fitting. Specifically, let u=f−pu=f-p. Under this condition, aC=−ua^C=-u and aF=(rτF−f)−ua^F=(r^F_\tau-f)-u, so the fitted acceleration and residual coefficients satisfy s=1−αs=1-\alpha, resulting in identical online predictions. Therefore, the substantial differences in Table 4 seem closely related to the divergence between the independently evolving full-computation and naive-cache trajectories. I would appreciate a more detailed explanation of why acceleration supervision is more effective under this setting, and whether its advantage persists when the supervision targets are evaluated at the same latent states.
  3. In Section 3.2, the calibration coefficients are fitted using naive-cache trajectories, whereas during inference they are applied to corrected-cache trajectories. Since the corrected features are written back into the cache and affect subsequent sampling steps, the actual feature dynamics may differ from those observed during offline calibration. Although Appendix A.8 examines sensitivity to calibration prompts, it does not directly address this discrepancy between calibration and deployment. Could the authors provide some analysis showing whether the acceleration relationships learned from naive-cache trajectories remain valid on corrected-cache trajectories?
Questions for authors

Question

My main concerns have been detailed in the Weaknesses section. In particular, I would like to see a controlled comparison between acceleration supervision and direct feature-reconstruction supervision under matched latent states and an identical two-anchor predictor. This would help clarify the actual source of the improvements reported in Table 4.

6 comments
AuthorsOpenReview ↗

Response to Reviewer pz7k (1/4): Learned Two-Anchor Prediction and Matched-Latent Generation Results

We thank the reviewer for recognizing the paper's writing, computational efficiency, and experimental analysis. We address the questions through controlled generation experiments, an exact analysis of supervision equivalence, single-position interventions, and tests of frozen calibration on corrected-cache trajectories.

1. Contribution of acceleration calibration and comparison with learned two-anchor predictors

Acceleration supervision improves generation with the same learned two-anchor predictor. In Table 4, changing the supervision target from direct reconstruction to acceleration reduces FID from 41.80262 to 10.36915, with the remaining settings fixed.

Our online update can indeed be written as a learned two-anchor linear predictor. The contribution is learning its coefficients from the acceleration relationship between full-compute and naive-cache trajectories, together with corrected writeback and isolated predictor history. Table 4's “Residual supervision” is precisely a learned two-anchor predictor trained by direct feature reconstruction. It holds the predictor form, parameter count, fitting procedure, and deployment policy fixed to isolate supervision choice.

Table 11's shared coefficient achieves SSIM and LPIPS close to step-by-branch coefficients, indicating that a lower-dimensional parameterization retains much of the benefit. Following the reviewer's suggestion, we further isolate supervision by matching calibration inputs while fixing parameterization and deployment.

2. Supervision equivalence under matched latents and the source of acceleration supervision's benefit

The matched-input experiment retains acceleration supervision's generation advantage. The derivation identifies a history-projection term that explains coefficient differences, and interventions test how these differences affect sampling error.

(a) Acceleration supervision still improves generation with matched calibration inputs

With matched calibration latents, acceleration supervision achieves FID 10.3911 versus 35.7894 for reconstruction/velocity with 50,000 images per evaluated coefficient pack.

We save every input latent of a naive-cache trajectory and independently replay the full network on each saved input to obtain teacher features. Replay outputs do not advance the next input. Reconstruction, velocity, and acceleration coefficients are fitted on this common sequence and frozen for generation, where each method evolves naturally under the same predictor and deployment policy.

The experiment uses DiT-XL/2 on ImageNet-256, DDIM-20, CFG=1.5, N = 5, warmup=3, and the same single calibration sample. Each distinct coefficient pack generates 50,000 images with global seed 0.

SupervisionFID ↓sFID ↓IS ↑Precision ↑Recall ↑
Direct reconstruction / Velocity35.789438.713175.46630.46770.4402
Acceleration10.39116.5014151.98470.67020.5740

Reconstruction and velocity produce identical eight-coefficient packs and share one generation result. These matched-input refits are distinct from the independently evolving calibration trajectories used in Table 4. Appendix G.1–G.2 provides the centered fitting formulation and equivalence derivation. The “Matched-input generation experiment” paragraph and its table in G.2 document this experiment.

AuthorsOpenReview ↗

Response to Reviewer pz7k (2/4): Supervision Equivalence and the History-Projection Identity

2(b). Matched input latents do not imply matching historical features

The coefficient difference equals the projection coefficient of the centered history discrepancy onto the centered prediction direction. This recovers the reviewer's conditional equivalence while explaining why matched inputs can yield different coefficients at later correction positions.

Let TtT_t be the full-network teacher feature on the saved input sequence, and f,pf,p the two most recent full-compute anchors. All variants use

u=f−p,r^t=f+ku,u=f-p,\qquad \hat r_t=f+ku,

with kR,kV,kAk_R,k_V,k_A fitted by reconstruction, velocity, and acceleration supervision. At the first cache step of an interval, the preceding step is full-compute, so matched inputs give Tt−1=fT_{t-1}=f. The cache reference signal and teacher acceleration satisfy

atC=f−2f+p=−u.a_t^C=f-2f+p=-u.

atF=Tt−2f+Tt−2=(Tt−f)−u+(Tt−2−p).a_t^F=T_t-2f+T_{t-2}=(T_t-f)-u+(T_{t-2}-p).

Reparameterizing kA=1−sAk_A=1-s_A gives the equivalent targets

yR=yV=Tt−f,yA=(Tt−f)+(Tt−2−p).y_R=y_V=T_t-f,\qquad y_A=(T_t-f)+(T_{t-2}-p).

Here yAy_A is the target in the common extrapolation parameterization. The original acceleration fit uses input atCa_t^C and target atFa_t^F.

Within a step–branch group, let HH denote centering. With identical observations, weights, and centering, ∥Hu∥2>0\|Hu\|^2>0, and no numerical safeguard or postprocessing altering the closed-form solution,

kR=kV=⟨Hu,H(Tt−f)⟩∥Hu∥2.k_R=k_V=\frac{\langle Hu,H(T_t-f)\rangle}{\|Hu\|^2}.

kA=⟨Hu,H[(Tt−f)+(Tt−2−p)]⟩∥Hu∥2.k_A=\frac{\langle Hu,H[(T_t-f)+(T_{t-2}-p)]\rangle}{\|Hu\|^2}.

Therefore,

kA−kR=⟨Hu,H(Tt−2−p)⟩∥Hu∥2.\boxed{k_A-k_R=\frac{\langle Hu,H(T_{t-2}-p)\rangle}{\|Hu\|^2}.}

Equality holds when this centered history projection vanishes. At the first correction, t=3t=3, both preceding steps are full-compute, giving Tt−2=pT_{t-2}=p and numerically verified equivalence up to floating-point rounding. At execution index 8, the latest anchors are f=T7f=T_7 and p=T2p=T_2, while teacher acceleration uses T8−2T7+T6T_8-2T_7+T_6. The discrepancy T6−T2T_6-T_2 yields measured attention coefficients kR=0.13794k_R=0.13794 and kA=0.94717k_A=0.94717, consistent with the identity.

Appendix G.1–G.2, Equations (13)–(22), provides the centered objective, target derivations, equality conditions, and numerical example. All execution indices are zero-based.

AuthorsOpenReview ↗

Response to Reviewer pz7k (3/4): Single-Position Interventions and Exact Error Decomposition

2(c). Testing the effect of coefficient differences on sampling error

All three positive single-position interventions reduce mean final Native-reference error, with Bonferroni-adjusted intervals excluding zero. An exact decomposition shows how the induced change can offset existing trajectory deviation even when local approximation error increases.

On 1,000 held-out class–noise pairs, one per ImageNet class, we use the frozen matched-input coefficients from the 50K experiment. Let γcal=kA−kR\gamma_{\mathrm{cal}}=k_A-k_R. Starting from reconstruction, we intervene separately at indices 8, 13, and 18, replacing the attention and MLP coefficients at one position by kR+γcal=kAk_R+\gamma_{\mathrm{cal}}=k_A or kR−γcalk_R-\gamma_{\mathrm{cal}}. Coefficients remain shared across layers, all other coefficients and update rules stay fixed, and subsequent latents, anchors, and caches evolve naturally.

From the reconstruction baseline's common prefix and current input, full-network evaluation, reconstruction, and positive intervention produce post-update latents zF,zR,z+z_F,z_R,z_+ using the same sampler update. Let zNz_N be the independent Native trajectory's post-update latent at the same step, with the same class and initial noise. Define

b=zF−zN,e=zR−zF,δ=z+−zR.b=z_F-z_N,\qquad e=z_R-z_F,\qquad \delta=z_+-z_R.

Here bb is the existing trajectory deviation propagated through the current full update, ee is the local approximation error, and δ\delta is the intervention-induced latent change. For dd latent elements,

=\frac{\|b+e+\delta\|^2-\|b+e\|^2}{d}.$$ $$\Delta E_{\mathrm{local},t} =\frac{\|e+\delta\|^2-\|e\|^2}{d}.$$ Expanding gives the exact identity, without linearizing the network or sampler, $$\boxed{\Delta E_{\mathrm{Native},t}=\Delta E_{\mathrm{local},t} +\frac{2\langle b,\delta\rangle}{d}.}$$ The interaction term measures whether the intervention-induced change offsets or reinforces existing trajectory deviation. A negative $\langle b,\delta\rangle$ indicates an offsetting component. Even if local MSE increases, Native-reference MSE decreases when the magnitude of the negative interaction term exceeds that increase. At index 8, mean local MSE increases by 0.0001765, but the mean interaction term is -0.0007277, yielding a mean current-step Native-reference MSE change of **-0.0005511** (values independently rounded). The mean decompositions at indices 13 and 18 show the same offsetting pattern. After continuing each rollout, mean final latent-MSE changes are reported below as intervention minus reconstruction, using Native with the same class and initial noise as the reference: | Execution index | Positive: mean final MSE change [98.333% CI] | Negative: mean final MSE change | |---|---:|---:| | 8 | -0.002302 [-0.004238, -0.000447] | +0.047421 | | 13 | -0.016952 [-0.018093, -0.015807] | +0.035448 | | 18 | -0.012436 [-0.012968, -0.011921] | +0.020789 | Intervals use 5,000 paired bootstrap resamples of the class–noise pairs, with Bonferroni adjustment for the three positive-intervention comparisons. All negative interventions increase mean final error, providing directional controls. On the same 1,000 samples, the complete acceleration policy reduces mean final latent MSE from **0.088017 to 0.074263**, a **15.63% reduction** (95% bootstrap CI: **[13.64%, 17.67%]**). Single-position effects need not add to the complete-policy effect. **Appendix G.3, Equations (23)–(25)**, provides the error decomposition, offsetting condition, and illustrative example.
Reviewer_SSrjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper improves generation quality over TaylorSeer under a 20-step inference setting by modifying feature caching in diffusion models. It uses the same prompt and initial noise to collect full-compute and naive-cache feature trajectories, and fits correction coefficients using least squares. The coefficients are specific to each sampling step and branch, and shared across layers. Under 20-step inference, the method reports substantial improvements in generation quality over TaylorSeer and other baselines on FLUX.1-dev, Qwen-Image, and DiT-XL/2.

Critical strengths
  1. Only a small amount of data is needed for offline calibration, after which 20-step generation quality is substantially better than TaylorSeer and other baselines.
  2. The final inference formula is relatively simple.
Critical weaknesses
  1. The main experimental comparisons use 20 steps, which is a relatively short sampling setting. For these models, I would usually expect 30 or even 50 steps when aiming for high generation quality. The 20-step baseline may already involve quality loss. Since users are usually sensitive to generation quality in practical applications, the practical value of improvements in this 20-step setting may be limited.
  2. Compared with training-based acceleration methods such as DMD, this class of methods, including the proposed method, has no advantage in speedup or image quality. This further limits its value in practical applications.
Questions for authors
  1. Can the authors provide comparisons with more common sampling step counts, such as 30 or more steps?
  2. Can the authors compare the proposed method with DMD and similar methods in terms of speedup and image quality?
16 comments
AuthorsOpenReview ↗

Response to Reviewer SSrj (1/2): Longer Sampling Schedules

We thank the reviewer for recognizing the low calibration data requirements, simple inference formula, and improvements in generation quality of our method. To address your concerns regarding sampling step counts and practical value, we have added experiments with longer sampling schedules and a direct comparison with a representative training-based acceleration method. We respond to both points below.

1. Effectiveness under longer sampling schedules

Across six configurations (30, 40, and 50 steps on FLUX.1-dev and Qwen-Image), CalibCache achieves higher mean ImageReward, PSNR, and SSIM, lower LPIPS, and faster denoising than TaylorSeer and HiCache.

All three caching methods use DrawBench-200 at 1024×10241024\times1024, N=5N=5, three warmup steps, and identical test prompts and seeds. CalibCache requires only one full-compute/naive-cache trajectory pair per configuration.

The 50-step results are shown below. Fidelity is measured against each model's 50-step full-compute outputs. Time denotes denoising latency.

FLUX.1-dev (50 steps)

MethodTime (s) ↓Speedup ↑PSNR ↑SSIM ↑LPIPS ↓CLIP ↑IR ↑
Full compute16.9131.00×∞1.00000.000031.2150.9799
TaylorSeer5.4903.08×16.520.69080.362031.3240.9902
HiCache5.3503.16×16.860.70460.339231.2771.0003
CalibCache (Ours)4.7813.54×17.460.70470.333431.5351.0029

Qwen-Image (50 steps)

MethodTime (s) ↓Speedup ↑PSNR ↑SSIM ↑LPIPS ↓CLIP ↑IR ↑
Full compute31.7231.00×∞1.00000.000032.9281.2289
TaylorSeer14.6512.17×14.650.64870.397332.5051.1449
HiCache16.3731.94×14.930.64950.377132.5961.1638
CalibCache (Ours)13.9972.27×18.180.74100.266332.8951.1927

Complete results, calibration settings, cache schedules, and timing details are in Appendix C (Tables 16–21).

AuthorsOpenReview ↗

Response to Reviewer SSrj (2/2): Training-Based Acceleration and Practical Value

2. Comparison with training-based acceleration methods and practical value

To address the request for comparisons with DMD and related training-based methods, we evaluate the publicly available 8-step Hyper-FLUX distilled model based on the same FLUX.1-dev backbone [1]. CalibCache (20 steps, N=4N=4) achieves higher reference-output fidelity at comparable denoising time and CLIP scores, with substantially lower preparation cost.

(1) Direct comparison with Hyper-FLUX

We evaluate all 200 DrawBench prompts at 1024×10241024\times1024, using identical test prompts and seeds. PSNR, SSIM, and LPIPS use the same 50-step full-compute FLUX.1-dev outputs (Naive-50) as the reference. Time denotes denoising latency on one H800 GPU at batch size 1.

MethodTime (s) ↓IR ↑CLIP ↑PSNR ↑SSIM ↑LPIPS ↓
Hyper-FLUX (8 steps)2.90671.025031.597913.02360.57680.5199
CalibCache (20 steps, N=4N=4)2.81730.966731.611615.47470.64380.4343

Relative to Hyper-FLUX, CalibCache increases PSNR by 2.4511 dB and SSIM by 0.0670, and reduces LPIPS by 0.0856. Although its mean ImageReward is slightly lower, we observe no noticeable visual quality degradation in the qualitative comparisons. In some cases, it better preserves reference subject appearance and spatial layout while more closely satisfying prompt requirements, illustrating the practical value of higher fidelity. Qualitative comparisons and experimental details are in Appendix E.

(2) Low preparation cost and straightforward adaptation

GLANCE reports approximately 1.1 million training samples and 800 GPU-hours for Hyper-FLUX [1, Table 2]. CalibCache instead fits scalar coefficients from one full-compute/naive-cache trajectory pair via closed-form least squares, without backpropagation, iterative optimization, or pretrained-weight updates. For FLUX.1-dev with 20 steps and N=4N=4, the recorded offline preparation took approximately 92 seconds on one H800 GPU. This lightweight procedure eases recalibration for different checkpoints or sampling configurations, particularly when suitable distilled weights are unavailable or training resources are limited. Its one-time cost is amortized over subsequent inference requests.

(3) Complementarity between distillation and caching

Distillation reduces sampling steps, while caching reuses intermediate features to reduce repeated computation. Prior work has combined these approaches: DisCa combines Restricted MeanFlow step distillation with a trainable cache predictor [2], while FlowCache applies training-free caching to the 16-step distilled MAGI-1 model for further acceleration [3].

Both training-based few-step distillation and feature caching are therefore important approaches to efficient diffusion generation; combining them offers a promising direction for improving the trade-off between generation quality and inference efficiency.

[1] Dong et al., Glance: Accelerating Diffusion Models with 1 Sample, Table 2.

[2] Zou et al., DisCa: Accelerating Video Diffusion Transformers with Distillation-Compatible Learnable Feature Caching.

[3] Ma et al., Flow Caching for Autoregressive Video Generation, Appendix E.1.

I acknowledge that CalibCache has an advantage over methods such as DMD in terms of training compute requirements. However, I remain skeptical that it can outperform these training-based methods in inference time and generation quality.

For the comparisons presented in the response, could the authors provide quantitative human evaluations or evaluations using advanced multimodal models? For industrial-grade image generation models, metrics such as ImageReward, CLIP, PSNR, and SSIM may not adequately reflect perceived image quality.

I also have reservations about the claimed complementarity between distillation and caching. Cache-based methods rely on the assumption that intermediate features evolve predictably during inference, and this assumption may be disrupted by distillation training. Applying cache-based prediction to a distilled sampling process compressed to only eight or even four steps also poses practical difficulties.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.