- Rating
- 2(−1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes TORE, a training-free diffusion acceleration method that selects token-wise Taylor orders using estimated prediction errors of the full-model residual. It estimates truncation and finite-difference derivative errors from cached residuals, addressing the mismatch between shallow-proxy scoring and residual prediction. Orders are preselected at full-computation steps, eliminating probes at skip steps. Experiments cover text-to-image, text-to-video, and class-to-image generation.
Critical strengths
[S1] Presentation
The paper is generally easy to follow. In particular, Figure 2 illustrates TORE’s improvement over shallow-proxy selection, while Figure 3 demonstrates the limitations of the shallow proxy, making the motivation easy to understand.
[S2] Clear Novelty
The order selection criterion and the elimination of probing at skip steps are clear distinctions from prior work.
Critical weaknesses
[W1] Limited Scope of Novelty
Token-wise selection of different Taylor predictors already exists in TAP. Moreover, the basic perspective of decomposing Taylor prediction error is already present in TaylorSeer. Therefore, several of the paper’s main contribution claims appear to require more restrained wording.
[W2] Limited Experimental Evidence
a) The experimental evaluation is not sufficiently extensive. For FLUX.1-dev, only 200 images are generated and evaluated, and no confidence intervals are reported. This makes it difficult to establish that TORE consistently outperforms the baselines in ImageReward and other metrics. Although I acknowledge that the 200-prompt DrawBench evaluation is a widely used protocol, I believe a more robust evaluation is needed. For example, RALU (CVPR 2026) and JiT (CVPR 2026) evaluate over 4,000 generated images, employing various metrics such as GenEval and T2I-CompBench.
b) The experimental setup is insufficiently described. It is unclear how many initial steps use full computation and how higher-order finite differences are handled before sufficient history becomes available.
Questions for authors
[Q1] Figure 7 and Table 4 demonstrate the effectiveness of adaptive selection, but they do not isolate the advantage of the proposed criterion over TAP’s proxy criterion. Please provide a comparison that keeps all other settings identical and changes only the selection criterion between TAP’s proxy and TORE’s criterion.
[Q2] What are the confidence intervals for the results in Table 1? Among the 200 images, on how many does TORE achieve the highest ImageReward score across the compared methods?
[Q3] If possible, include additional images for FLUX.1-dev and report the evaluation results calculated based on a total of 1,000 or more images.
[Q4] How many initial generation steps use full computation? How are higher-order finite differences handled before sufficient history becomes available?
[Q5] How do performance, latency, FLOPs, and peak VRAM usage vary with M? In particular, since finite differences from order 0 through M+1 are cached, peak VRAM usage may increase with M. Does this occur in practice?
6 comments
Response to reviewer
For question 2: We thank the reviewer for this suggestion. In our original DrawBench evaluation, we used five different random seeds for each of the 200 prompts, generating five images per prompt and a total of 1,000 images per method. However, we did not clearly specify this evaluation protocol or report confidence intervals in Table 1. We now provide 95% confidence intervals for all generation quality metrics, as presented in the table below. For the per-image ImageReward comparison, TORE achieves the highest score among the accelerated methods on 260, 275, and 250 out of 1,000 images (26.0%, 27.5%, and 25.0%) for the three Table 1 configurations, respectively: ((N,M)=(3,4),(5,6),(8,5)).
| Method | ImageReward ↑ | CLIP ↑ | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|---|
| Full | 1.0414 (±0.1622) | 28.010 (±0.981) | — | — | — |
| 60% steps | 1.0069 (±0.1609) | 27.440 (±0.934) | 19.060 (±0.713) | 0.7401 (±0.0249) | 0.2477 (±0.0253) |
| 50% steps | 1.0126 (±0.1618) | 27.607 (±0.894) | 17.408 (±0.652) | 0.6947 (±0.0254) | 0.3102 (±0.0253) |
| 40% steps | 0.9418 (±0.1720) | 27.505 (±0.821) | 16.363 (±0.563) | 0.6604 (±0.0246) | 0.3599 (±0.0238) |
| TaylorSeer (N=2,O=1) | 1.0494 (±0.1557) | 27.873 (±0.909) | 22.142 (±0.846) | 0.8043 (±0.0186) | 0.1575 (±0.0175) |
| HiCache (N=2,O=2)† | 0.9913 (±0.1733) | 27.668 (±0.926) | 23.953 (±0.919) | 0.8109 (±0.0231) | 0.1394 (±0.0212) |
| TC-Padé (N=3,θ=1.0) | 0.9968 (±0.1628) | 27.608 (±0.951) | 23.035 (±0.838) | 0.8384 (±0.0218) | 0.1546 (±0.0156) |
| TAP (N=2) | 1.0224 (±0.1592) | 27.826 (±0.903) | 21.368 (±0.848) | 0.7928 (±0.0187) | 0.1817 (±0.0175) |
| TORE (N=3,M=4) | 1.0462 (±0.1622) | 28.043 (±0.947) | 24.450 (±0.977) | 0.8291 (±0.0196) | 0.1188 (±0.0163) |
| TaylorSeer (N=3,O=2) | 1.0266 (±0.1616) | 27.919 (±0.954) | 19.297 (±0.738) | 0.7345 (±0.0218) | 0.2391 (±0.0211) |
| HiCache (N=3,O=2) | 1.0275 (±0.1597) | 27.945 (±0.921) | 19.554 (±0.726) | 0.7384 (±0.0218) | 0.2319 (±0.0216) |
| TC-Padé (N=6,θ=1.0) | 0.9808 (±0.1679) | 27.509 (±1.011) | 22.034 (±0.808) | 0.8043 (±0.0246) | 0.2067 (±0.0194) |
| TAP (N=3) | 0.9781 (±0.1677) | 27.559 (±0.940) | 18.683 (±0.678) | 0.7281 (±0.0224) | 0.2588 (±0.0219) |
| TORE (N=5,M=6) | 1.0364 (±0.1628) | 28.011 (±0.974) | 24.387 (±0.971) | 0.8210 (±0.0209) | 0.1302 (±0.0174) |
| TaylorSeer (N=5,O=2) | 1.0528 (±0.1607) | 27.819 (±0.984) | 16.657 (±0.531) | 0.6546 (±0.0244) | 0.3600 (±0.0212) |
| HiCache (N=4,O=2) | 1.0252 (±0.1580) | 27.879 (±0.925) | 18.023 (±0.648) | 0.6893 (±0.0238) | 0.2942 (±0.0221) |
| TC-Padé (N=8,θ=1.4) | 0.9469 (±0.1680) | 27.430 (±0.989) | 20.644 (±0.729) | 0.7679 (±0.0249) | 0.2728 (±0.0217) |
| TAP (N=5) | 0.9431 (±0.1669) | 27.585 (±0.973) | 15.583 (±0.600) | 0.6410 (±0.0218) | 0.3998 (±0.0206) |
| TORE (N=8,M=5) | 1.0308 (±0.1678) | 28.020 (±0.947) | 21.699 (±0.824) | 0.7582 (±0.0236) | 0.2126 (±0.0215) |
For question 5: We thank the reviewer for this question. We conducted a sensitivity analysis on FLUX.1-dev by varying (M) from 0 to 6 while fixing (N=7). We evaluate generation fidelity, latency, FLOPs, and peak VRAM. As illustrated in Appendix Fig.10 (page 15), increasing introduces additional computational and memory overhead but also improves generation fidelity, with diminishing gains at higher orders. Importantly, only determines the maximum candidate order, while the actual prediction order is adaptively selected for each token. Therefore, increasing does not necessarily require higher-order Taylor extrapolation for all tokens. These observations highlight the trade-off between prediction fidelity and computational cost, motivating a moderate choice of that balances both.
I thank the authors for the detailed response and the additional experiments. However, several of my concerns remain unresolved, as detailed below.
[W1] Limited Scope of Novelty
Just to clarify, my concern in W1 was not whether TaylorSeer [1] provides a computable criterion for adaptive order selection. I agree that it doesn't. My concern was that the error decomposition itself, which the paper presents as one of its main contributions, already appears in TaylorSeer. This point was not addressed in the response, so let me elaborate.
The decomposition is already in TaylorSeer (Sec. 3.2, "Error Bounds Analysis", Eqs. (11)–(12)). TaylorSeer bounds the order- prediction error as a truncation term, , plus a finite-difference term, . It also discusses the same trade-off that motivates TORE (paper lines 85–92): a higher order reduces the truncation error but adds finite-difference error that grows with .
TORE's estimator is essentially TaylorSeer's bound, with the unknown constants filled in by cached finite-difference norms. The truncation estimate (Eq. (8)) is TaylorSeer's first term with . The finite-difference estimate (Eq. (10)) can be rewritten as , which is TaylorSeer's second term with .
However, the paper does not mention this connection. TaylorSeer is cited only as a forecasting method (paper lines 56 and 130), while the decomposition is claimed as the paper's own contribution (paper lines 101–104 and Sec. 3.3.2). The response itself says that TORE extends "this analysis," which suggests the analysis already exists in TaylorSeer.
So I see the error decomposition as prior work. In my view, TORE's contribution is narrower: a concrete estimate of TaylorSeer's constants, and its use as a token-wise selection score.
[1] Liu, Jiacheng, et al. "From reusing to forecasting: Accelerating diffusion models with TaylorSeers." ICCV, 2025.
[W2-a, Q2] Statistical Significance of Quality Gains
Looking at the new confidence intervals, the gains in ImageReward and CLIP Score do not seem statistically meaningful. The differences fall within the intervals, and TORE is behind TaylorSeer in ImageReward in two of the three settings. TORE also wins on only 25.0–27.5% of images, which is close to chance. This is not a minor issue of trailing on one metric: for T2I, ImageReward is the only reported metric that reflects image quality itself, whereas PSNR, SSIM, and LPIPS all measure similarity to the full model. As it stands, the evidence supports better fidelity to the full model, but the contribution to image quality remains unclear.
[Q4] Availability of Higher-Order Differences
I'm not sure how often TORE actually selects high orders. Using an order up to requires cached full steps, which takes up much of the trajectory when is large, and the response also mentions that candidate orders are restricted early on. In Fig. 8, the finite-difference error already exceeds the truncation error at order 1. If most selected orders are 0–2, the large values in the main tables are mostly nominal. Could the authors show the distribution of selected orders for the main configurations?
[Q5] Sensitivity to
The response presents results that were not actually measured. It says that fidelity, latency, FLOPs, and peak VRAM were evaluated, but according to the caption of Fig. 10, all curves other than peak VRAM "are illustrative and are not experimental measurements." Illustrative curves should not be presented as results in a rebuttal, and the conclusion drawn from them (fidelity improves with with diminishing returns) is therefore unsupported. The VRAM measurement does address my memory concern, but the other panels need to be either actually measured or removed.
Response to reviewer
We have uploaded a revised manuscript with additional experiments and analyses in the Appendix, as referenced in our responses below.
For weakness 1: We thank the reviewer for this suggestion. As illustrated in Fig. 2, TAP selects Taylor orders using shallow-layer proxy errors, which may not reflect the errors of the actual prediction target. This proxy–target mismatch can lead to suboptimal order selection, as demonstrated in Fig. 3. Our main contribution lies in the order-selection criterion. Specifically, starting from the Taylor forecasting formulation, we decompose residual prediction error into truncation error and finite-difference derivative approximation error. We then derive computable estimates of both components from cached full-model residuals (Eqs. 8–10). By minimizing their combined estimate, TORE selects the Taylor order for each token and prediction horizon without requiring future ground-truth residuals. Regarding TaylorSeer, it does not provide a computable error estimation criterion for adaptive order selection during inference. TORE extends this analysis into a practical selection criterion directly aligned with the residual prediction target. Detailed derivations are provided in Appendix B.
For weakness 2 (a) and question 3: We thank the reviewer for this suggestion. We would like to clarify that our original DrawBench evaluation uses 200 prompts with five random seeds per prompt, totaling 1,000 generated images per method. To further strengthen the evaluation, we conducted additional experiments on GenEval, following the evaluation protocol used by JiT (CVPR 2026), with 553 prompts and four random seeds per prompt, totaling 2,212 images per method. Together with DrawBench, our evaluation now covers 3,212 generated images per method across two benchmarks, providing broader evidence of generation quality and efficiency.
For weakness 2 (b) and question 4: We thank the reviewer for this question. To balance warm-up overhead with the availability of higher-order finite differences, TORE determines the number of initial full-computation steps based on M, using W=max(2,⌊M/2⌋+1). Similar to the short warm-up strategy used in TaylorSeer (ICCV 2025), TORE progressively accumulates residual history to support higher-order forecasting. During the early stages, candidate orders are restricted by the available finite differences. This strategy enables forecasting without waiting for all high-order differences to become available, while progressively expanding the candidate orders as the cache grows.
For question 1: We thank the reviewer for this suggestion. To directly evaluate the advantage of TORE's criterion over TAP's shallow-proxy criterion, we conducted controlled experiments where only the order-selection criterion differs. Specifically, both methods extrapolate from the same full-computation anchor along a shared generation trajectory, using identical input states, residual caches, candidate Taylor predictors, and prediction distances. We compare the true residual prediction MSE of TAP, TORE, and the residual oracle, which selects the lowest-error candidate order for each token. We also report the relative prediction error gap to this oracle. TORE achieves lower prediction errors and smaller oracle gaps across all three settings and the advantage becomes more pronounced under longer forecasting intervals, supporting the effectiveness of TORE's residual-oriented selection criterion.