acceptodds
Under review as a conference paper at ICLR 2027

Beyond the History Span: State-Guided Span Expansion for Efficient Diffusion Inference

Abstract

Diffusion Transformers require repeated model evaluations across many sampling steps, resulting in substantial computational overhead. A recent line of methods reduces this cost by replacing selected model evaluations with linear forecasts. These methods represent the Transformer velocity at each skipped step as a linear combination of historical velocities. However, forecasting errors grow over consecutive forecast steps because such forecasts are confined to the linear span of historical velocities and cannot represent velocity components outside this span. To overcome this limitation, we propose state-guided span expansion (SGSE), which expands the historical velocity span using a complementary direction derived from the sampling state. Specifically, we forecast a sampling state and map it through the inverse sampler update to obtain a state-implied velocity. The out-of-span component of this state-implied velocity is directionally similar to the corresponding component of the full-Transformer velocity, providing complementary information unavailable from historical velocities alone. SGSE incorporates this complementary direction into velocity forecasting, expanding the representational space beyond the historical span and thus reducing forecasting error. Experiments on image and video generation demonstrate that SGSE consistently outperforms existing forecasting-based acceleration methods.

Primary area:
generative models
Keywords:
Feature Caching, Diffusion Acceleration, Flow Matching Acceleration, Text-to-Image Generation, Linear Predictor

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2GXqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies training-free acceleration of flow-matching diffusion transformers, where full-Transformer (FT) evaluations are performed only at selected timesteps and the velocity at the remaining steps is forecast as a calibrated linear combination of recent velocities (as in L2P). The authors argue that error accumulates because forecasted velocities never leave the span of the historical velocities. They propose SGSE, which also linearly forecasts the next sampling state, maps it through the inverse sampler update to a “state-implied velocity”, projects it onto the orthogonal complement of the history span, and adds it with a per-timestep calibrated scale γ_i. All coefficients are calibrated offline on 50 FT trajectories. A projection argument shows that the minimum achievable error decreases by ‖g_i‖²A_i². Experiments on FLUX.2 Klein 4B (DrawBench) and Wan2.1-1.3B (VBench) show modest fidelity gains over L2P, TaylorSeer, VDE and MeanCache at matched FT budgets, with ablations on FLUX.2 under the 11-FT schedule.

Critical strengths
  1. The diagnosis is clear and well supported. The paper decomposes the forecasting error into in-span and out-of-span parts (Eq. 6, Fig. 2) and shows that linear forecasting is limited by the history span. This gives a useful lens for the growing family of forecasting-based caching methods.
  2. The method is simple and cheap. It adds only one calibrated scalar per timestep, requires no retraining, and yields consistent fidelity gains (SSIM/PSNR) over the strongest baseline L2P at all FT budgets on both an image and a video model, with negligible overhead.
Critical weaknesses
  1. The “state forecasting” narrative obscures what the method actually adds, and the key baseline is missing. Appendix E shows that under Euler, SGSE reduces exactly to vˉi=v^i+ηi(I−Pi)zi\bar v_i=\hat v_i+\eta_i(I-P_i)z_i, and directly calibrating η_i performs almost identically (Table 7). Since ziz_i equals the initial noise plus the accumulated past velocities, (I−Pi)zi(I-P_i)z_i is essentially the component of the initial noise, together with velocities older than the window, that lies outside the history span. In the “All” window setting (Table 6), it reduces to the out-of-span component of the initial noise alone. Under rectified flow, the velocity is closely tied to the initial noise and the current state, so this is a natural extra basis vector. The paper should therefore compare against the straightforward baseline of adding the initial noise ε and/or the current state ziz_i as additional regression bases. Without this comparison, both the role of state forecasting and the novelty relative to L2P are unclear.
  2. Gains are small, measured only as fidelity to FT, and the evaluation is narrow. Over L2P, SSIM improves by about 0.01, LPIPS is slightly worse at 9 FT on FLUX.2, and VBench is lower than L2P at both budgets on Wan2.1. A random Gaussian direction with FT calibration already recovers part of the gain (0.8294 vs. 0.8265 without guidance and 0.8358 for SGSE; Table 5), and the alignment A_i is small (about 0.05–0.1) for most intermediate timesteps (Fig. 2d). The coefficients are shared across all samples and calibrated on only 50 prompts, yet no robustness to prompt distribution, CFG scale, resolution or step count is tested. No variance across seeds or per-sample worst-case errors are reported, and all ablations are on FLUX.2 with 11 FT only.
Questions for authors
  1. How does SGSE compare with simply adding the initial noise ε, or the current state ziz_i, as an extra basis vector in the calibrated linear forecaster, on both FLUX.2 and Wan2.1? If these baselines perform comparably, what does the state forecast contribute?
  2. Can you show the calibrated γ_i values across timesteps and repeat the key ablations (Tables 3 and 5) on Wan2.1 with UniPC? Fixing γ_i = 1 performs much worse than no guidance at all (0.7693 vs. 0.8265), which suggests the direction is unreliable without per-step rescaling.
5 comments
AuthorsOpenReview ↗

Response to W1

W1: The paper should compare against the straightforward baseline of adding the initial noise ε and/or the current state as additional regression bases.

Thank you for your comment. State forecasting provides a general method to construct the state-guided direction under different samplers (like Flow-UniPC), without relying on a state update rule specific to the Euler sampler. Compared with L2P, which forecasts velocities using linear combinations of historical velocities, our contribution is to identify the representational limitations of this approach and expand the historical velocity span with complementary information from sampling states.

We added two baselines that separately use the current state ziz_i and initial noise ϵ\epsilon as additional basis vectors. As shown in Table R3.1, both direct baselines achieve performance close to SGSE on FLUX.2, consistent with our analysis. However, on Wan2.1 with Flow-UniPC, SGSE outperforms both direct baselines across all four metrics. Moreover, both direct baselines perform worse than velocity-only forecasting, whereas SGSE achieves better results. These results show that directly using the current state or initial noise as an additional direction does not improve performance under Flow-UniPC, while SGSE consistently improves performance.

Table R3.1: Comparison of velocity-only forecasting, direct current-state and initial-noise span expansion, and SGSE. Velocity-only results are from Tables 3 and 4 (page 10, lines 487–503), while direct-state and direct-noise results are from Tables 12 and 13 (page 20, lines 1027–1041).

ModelMetricVelocity onlyDirect stateDirect noiseSGSE
FLUX.2SSIM ↑0.82650.83470.83450.8358
FLUX.2PSNR ↑20.854521.026221.013321.0309
FLUX.2LPIPS ↓0.14690.14420.14370.1432
FLUX.2CLIPScore ↑0.82180.82150.82190.8228
Wan2.1SSIM ↑0.64380.63520.63540.6450
Wan2.1PSNR ↑17.750917.696117.641317.8253
Wan2.1LPIPS ↓0.32460.32770.33000.3223
Wan2.1VBench ↑84.137883.900284.072684.2584
AuthorsOpenReview ↗

Response to W2 part1

W2: ① Gains are small, ② measured only as fidelity to FT, and ③ the evaluation is narrow.

We appreciate your detailed comments on the performance gains, evaluation, and robustness of SGSE. Below, we provide point-by-point responses to your eight concerns.

① We conducted additional evaluations using different generation seeds (page 15, lines 795–809), providing further evidence of the stability of SGSE's performance gains. As shown in Table R3.2, SGSE achieves higher mean SSIM and PSNR than L2P across all five evaluated configurations on FLUX.2 and Wan2.1, along with lower mean LPIPS in four of them. The evaluation metrics also exhibit relatively limited variation across generation seeds. These results further highlight the value of state-guided span expansion: leveraging complementary information beyond the historical velocity span at minimal computational cost offers a valuable direction for further improving generation quality in efficient diffusion inference.

Table R3.2: Multi-seed evaluation (mean ± standard deviation over three seeds). Results are from Tables 7 and 8 (page 15, lines 757–775).

ModelFTL2P SSIM ↑SGSE SSIM ↑L2P PSNR ↑SGSE PSNR ↑
FLUX.2180.8728 ± 0.00460.8791 ± 0.004422.2788 ± 0.220122.5673 ± 0.2139
FLUX.2110.8141 ± 0.00380.8262 ± 0.004120.2964 ± 0.038420.5455 ± 0.0778
FLUX.290.7640 ± 0.00350.7723 ± 0.003118.3504 ± 0.066218.5022 ± 0.0817
Wan2.1180.7536 ± 0.00300.7628 ± 0.002220.4336 ± 0.041320.6845 ± 0.0387
Wan2.1110.6401 ± 0.00400.6445 ± 0.003017.6226 ± 0.050517.7511 ± 0.0870

② Following prior work L2P, we adopted SSIM, PSNR, and LPIPS to evaluate fidelity, ensuring a consistent comparison. In addition, we introduced CLIPScore and VBench to evaluate generation quality. We also measured inference latency and floating-point operations (FLOPs) to analyze the trade-off between quality and efficiency, as shown in Fig. 4 (page 8, lines 378–390). Furthermore, we provided qualitative visual comparisons of representative samples in Figs. 5 and 6 (page 9, lines 432–466) and Appendix D (pages 16–17, lines 858–863). Thus, our evaluation covers fidelity, generation quality, inference efficiency, and qualitative results, rather than focusing solely on fidelity.

③ In the original manuscript, we evaluated SGSE on both text-to-image and text-to-video generation in terms of fidelity, generation quality, inference efficiency, and qualitative results. We also examined performance across different FT budgets and conducted ablations on key components, method mechanisms, and history-window size. In the revised manuscript, we further extended our evaluation by adding comparisons with reduced-step FT on both models (Tables 1 and 2, page 7, lines 325–366), component and mechanism ablations on Wan2.1 with Flow-UniPC (Tables 4 and 11, pages 10 and 17, lines 496–503 and 892–902), multi-seed evaluations and per-sample worst-case analysis (Tables 7–9, page 15, lines 756–790), and direct current-state and initial-noise span expansion baselines under Euler and Flow-UniPC (Tables 12 and 13, page 20, lines 1026–1041). These additional experiments provide further evidence of SGSE's effectiveness across different generation tasks, FT budgets, random seeds, and samplers, while offering a more comprehensive evaluation of state-guided span expansion.

AuthorsOpenReview ↗

Response to W2 part2

W2: ④ A random Gaussian direction with FT calibration already recovers part of the gain (0.8294 vs. 0.8265 without guidance and 0.8358 for SGSE Table 5), and ⑤ the alignment A_i is small (about 0.05–0.1) for most intermediate timesteps (Fig. 2d). ⑥ The coefficients are shared across all samples and calibrated on only 50 prompts, yet no robustness to prompt distribution, CFG scale, resolution or step count is tested.

④ A random Gaussian direction with FT calibration can yield performance gains, which is consistent with our theoretical analysis that expanding the historical velocity span can reduce the approximation error even when the additional direction exhibits weak alignment with the missing velocity component. Specifically, a random direction may have a nonzero projection onto the missing velocity component, which can be exploited through calibration. In the FT-calibrated Gaussian, we replace the current state ziz_i in Eq. (10) with a random Gaussian vector while retaining the subsequent inverse sampler update, orthogonal projection, and FT calibration. The timestep-specific coefficient γi\gamma_i is learned to adjust the sign and magnitude of the resulting correction direction, allowing the baseline to exploit its limited complementary information. However, state guidance achieves better results. The SSIM for velocity-only forecasting, FT-calibrated Gaussian guidance, and SGSE are 0.8265, 0.8294, and 0.8358 on FLUX.2, and 0.6438, 0.6440, and 0.6450 on Wan2.1, respectively (Tables 10 and 11, page 17, lines 879–902). These results show that calibration can partially benefit from an additional direction, but the complementary information provided by sampling states leads to better performance.

⑤ AiA_i measures the directional alignment between the state-guided direction and the missing velocity component. According to Eq. (19) (page 6, lines 304–310), the reduction in the minimum achievable squared approximation error is ∣∣gi∣∣22Ai2.||g_i||_2^2 A_i^2. Thus, the theoretical improvement depends on both the alignment AiA_i and the magnitude of the missing velocity component gig_i. Even when AiA_i is small, nonzero alignment indicates that the state-guided direction can capture part of the missing velocity component beyond the historical velocity span.

⑥ We followed prior work L2P in using shared calibration coefficients and a small calibration set. L2P specifically studied the effects of calibration set size. We adopted its 50-sample setting for a fair comparison. Moreover, our calibration and evaluation sets are disjoint, providing evidence of cross-dataset generalization. Specifically, FLUX.2 was calibrated on 50 PartiPrompts samples and evaluated on DrawBench, while Wan2.1 was calibrated on 50 FETV prompts and evaluated on VBench, as described in Section 4.1 (page 8, lines 398–402). We followed the evaluation settings of image generation used in L2P, including CFG scale, resolution, and total sampling steps, and evaluated all methods under the same configurations. We agree that further robustness evaluations of the calibration coefficients across different prompt distributions, CFG scales, resolutions, and sampling-step counts would be valuable. However, we cannot complete all these additional experiments within the rebuttal period. We will investigate these settings and report more detailed results in future work.

Reviewer_6QjtOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes state-guided span expansion (SGSE) to improve velocity forecasting when Transformer evaluations are skipped during diffusion sampling. Its central observation is that linear combinations of cached velocities cannot recover components outside their historical span. SGSE forecasts the next sampling state, converts it into a state-implied velocity through the inverse sampler update, and adds its orthogonal component to the velocity forecast using a calibrated coefficient. Experiments on FLUX.2 and Wan2.1 show improved fidelity to full-Transformer outputs over existing forecasting methods at similar computational costs.

Critical strengths
  1. The geometric formulation identifies a concrete limitation of linear velocity forecasting. The decomposition into in-span and out-of-span errors clearly separates coefficient prediction errors from representational limitations. This provides a useful explanation for why improving forecasting coefficients alone may be insufficient, and motivates a targeted correction with limited computational overhead.

  2. The fidelity improvements extend across image and video generation. SGSE improves SSIM and PSNR over the compared methods across the reported evaluation budgets, with corresponding latency comparisons. Calibrating L2P on the same datasets also strengthens the comparison against the closest baseline. These results support the practical usefulness of incorporating state information into velocity forecasting.

Critical weaknesses

The central role attributed to state forecasting is weakened by the paper’s own Euler equivalence. Appendix E shows that, under Euler sampling, the complete correction reduces to vˉi=v^i+ηi(I−Pi)zi.\bar v_i=\hat v_i+\eta_i(I-P_i)z_i. Thus, the additional direction is available directly from the current state; forecasting the next state determines only a scalar that can be absorbed into the learned scaling coefficient. Table 7 further shows that directly calibrating this simpler formulation nearly reproduces SGSE’s performance. This substantially narrows the conceptual contribution to a calibrated current-state augmentation of velocity forecasting. The manuscript acknowledges the equivalence but does not adequately reconcile it with the main explanation that accurate next-state forecasting supplies the complementary geometric information. A substantive benefit of the more elaborate construction remains insufficiently established.

Questions for authors
  1. Clarification of the contribution: How should the claimed role of state forecasting be interpreted given Eq. (33) and the near-equivalent results in Table 7? What evidence establishes a consequential benefit of the state-forecasting construction beyond directly calibrating the projected current-state correction, particularly for Flow-UniPC?

  2. Clarification of the diagnostic protocol: In Figs. 2(b–d) and 3, are the histories obtained from actual recursive accelerated sampling, and are reference velocities evaluated at those accelerated states? Or are some quantities taken from full-Transformer trajectories?

  3. Additional evidence needed for the current practical claim: At matched end-to-end latency, how does SGSE compare with reducing the number of steps of the underlying sampler? In that comparison, does its advantage extend to generation-quality metrics, or primarily to resemblance to the 50-step reference?

3 comments
AuthorsOpenReview ↗

Response to W & Q1

W & Q1: The central role attributed to state forecasting is weakened by the paper’s own Euler equivalence. & How should the claimed role of state forecasting be interpreted given Eq. (33) and the near-equivalent results in Table 7?

We appreciate your attention to the role of state forecasting. The role of state forecasting is to provide a general way to construct the state-guided direction under different samplers, rather than limiting SGSE to a simplified formulation that is specific to Euler sampling.

As derived in Eqs. (28)–(33) (pages 19–20, lines 1020–1056), under Euler sampling, the state-guided correction can be equivalently expressed using the current state:

vˉi=v^i+ηi(I−Pi)zi.\bar{v}_i=\hat{v}_i+\eta_i(I-P_i)z_i.

However, this equivalence relies on the Euler update rule and does not directly extend to Flow-UniPC, whose updates also depend on the solver history.

To verify the practical benefit of the state-forecasting construction, we added comparisons with direct current-state span expansion. The results are shown below.

Table R2.1: FLUX.2 uses Euler, and Wan2.1 uses Flow-UniPC. Velocity-only results are from Tables 3 and 4 (page 10, lines 487–503), while direct-state results are from Tables 12 and 13 (page 20, lines 1027–1041).

ModelMetricVelocity onlyDirect stateSGSE
FLUX.2SSIM ↑0.82650.83470.8358
FLUX.2PSNR ↑20.854521.026221.0309
FLUX.2LPIPS ↓0.14690.14420.1432
FLUX.2CLIPScore ↑0.82180.82150.8228
Wan2.1SSIM ↑0.64380.63520.6450
Wan2.1PSNR ↑17.750917.696117.8253
Wan2.1LPIPS ↓0.32460.32770.3223
Wan2.1VBench ↑84.137883.900284.2584

Under Euler sampling, direct current-state span expansion achieves performance close to SGSE, consistent with the theoretical equivalence. In contrast, under Flow-UniPC, direct current-state span expansion performs worse than velocity forecasting without state guidance, whereas SGSE achieves the best results across all four metrics. This demonstrates the practical benefit of using state forecasting together with the inverse sampler update.

Moreover, compared with direct current-state span expansion, SGSE introduces only 0.000024% additional TFLOPs on FLUX.2 and 0.000039% on Wan2.1 (pages 20–21, lines 1075–1077 and 1087–1089). Thus, the sampler-aware state-forecasting construction provides practical benefits with negligible additional computational cost.

Finally, we do not claim that state forecasting is the only way to construct complementary directions. One of our contributions is to expand the historical velocity span with a complementary direction inferred from the sampling state, rather than limiting SGSE to a simplified formulation specific to a particular sampler. We do not rule out the possibility of similar equivalent simplifications for other samplers. We believe that our work provides a useful direction for studying the relationship between sampler update rules and the construction of complementary directions, as well as for developing more effective velocity forecasting methods under different samplers.

AuthorsOpenReview ↗

Response to Q2

Q2: In Figs. 2(b–d) and 3, are the histories obtained from actual recursive accelerated sampling, and are reference velocities evaluated at those accelerated states? Or are some quantities taken from full-Transformer trajectories?

Thank you for your question. The histories in Figs. 2(b–d) and 3 come from actual recursive accelerated sampling. The reference velocities are evaluated at those accelerated states. In the revised manuscript, we clarified the sources of the histories and reference velocities in the captions of Fig. 2 (page 4, lines 173–175) and Fig. 3 (page 6, lines 280–284), with the added text highlighted in blue.

AuthorsOpenReview ↗

Response to Q3

Q3: At matched end-to-end latency, how does SGSE compare with reducing the number of steps of the underlying sampler? In that comparison, does its advantage extend to generation-quality metrics, or primarily to resemblance to the 50-step reference?

Thank you for your suggestion to compare SGSE with reducing the number of sampling steps. We introduced reduced-step FT, which covers the full sampling interval with fewer, more widely spaced timesteps and performs a full-Transformer evaluation at every step (page 8, lines 393–396). Fig. 4 (page 8, lines 378–390) shows that SGSE achieves higher SSIM than reduced-step FT at comparable end-to-end latency on both FLUX.2 and Wan2.1.

As shown in Tables R2.2 and R2.3, SGSE consistently achieves higher fidelity and better generation quality than reduced-step FT across all evaluated configurations. These results demonstrate that SGSE is more effective than simply reducing the number of steps of the underlying sampler.

Table R2.2: Comparison between reduced-step FT and SGSE on FLUX.2 using DrawBench. Results are reported under the same FT budgets. Bold values indicate better results.

FTSSIM ↑ (Reduced-step FT / SGSE)PSNR ↑ (Reduced-step FT / SGSE)LPIPS ↓ (Reduced-step FT / SGSE)CLIPScore ↑ (Reduced-step FT / SGSE)
180.7991 / 0.886919.1799 / 23.17290.1923 / 0.09590.8234 / 0.8252
110.7333 / 0.835817.1567 / 21.03090.2643 / 0.14320.8192 / 0.8228
90.7116 / 0.775416.9387 / 18.51760.2860 / 0.21690.8147 / 0.8170

Table R2.3: Comparison between reduced-step FT and SGSE on Wan2.1 using VBench. Results are reported under the same FT budgets. Bold values indicate better results.

FTSSIM ↑ (Reduced-step FT / SGSE)PSNR ↑ (Reduced-step FT / SGSE)LPIPS ↓ (Reduced-step FT / SGSE)VBench ↑ (Reduced-step FT / SGSE)
180.4784 / 0.760013.3015 / 20.58220.4720 / 0.185883.1690 / 83.8665
110.4147 / 0.635411.9771 / 17.60080.5648 / 0.329281.2128 / 82.9168
Reviewer_FC8vOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose State-Guided Span Expansion (SGSE). SGSE is a training-free acceleration method for diffusion models. SGSE identifies that linear velocity forecasting is restricted to the span of historical velocities, which can lead to accumulated prediction errors in steps. SGSE forecasts the next sampling state and converts it into a state-implied velocity, whose out-of-span component is used to improve velocity prediction. Experiments on FLUX.2 and Wan2.1 demonstrate improved reconstruction fidelity relative to full-Transformer outputs and compared with existing baseline methods.

Critical strengths
  1. The paper provides an interesting analysis of the limitations of historical velocity forecasting. Figures 1 and 2 provide useful insights into error accumulation during diffusion inference.

  2. The proposed method achieves consistent improvements in reconstruction fidelity compared with existing baseline methods. Tables 1 and 2 demonstrate improvements in SSIM and PSNR under comparable numbers of full-Transformer evaluations. The extension of the method appears to be useful in both image and video generation.

Critical weaknesses
  1. SGSE gains over L2P are relatively small, especially on video generation. PSNR only increases from 20.3725 to 20.5822, while VBench slightly decreases from 83.8872 to 83.8665 (Table 2). I am not fully convinced that the improvements are sufficiently significant.

  2. Appendix E shows that, under Euler sampling, SGSE can be equivalently expressed as a correction using the current sampling state and a single calibrated coefficient (Eq. 33). Furthermore, Table 7 shows that directly calibrating this simplified formulation achieves nearly identical results. This raises questions about whether the additional state forecasting mechanism is fundamentally necessary.

  3. The paper shows that expanding the historical velocity span reduces the theoretical minimum approximation error. However, this result does not necessarily imply that. The ablation in Table 3 shows only a relatively modest improvement from adding state guidance, with SSIM increasing from 0.8265 to 0.8358.

Questions for authors

Please see the weakness section.

3 comments
AuthorsOpenReview ↗

Response to W1

W1. SGSE gains over L2P are relatively small, especially on video generation.

Thank you for your comment. Although the absolute performance gains over L2P are not very substantial, SGSE provides a simple, broadly applicable, and theoretically grounded approach to improving velocity forecasting with negligible additional computational overhead. To further verify the stability of these improvements, we conducted additional experiments using three generation seeds. The results are presented in Table R1.1 below and Tables 7 and 8 of the revised manuscript (page 15, lines 757–775), with the experimental setup described in Appendix C (page 15, lines 795–809). SGSE achieves higher mean SSIM and PSNR than L2P across all five evaluated configurations. For video generation quality, although the original single-seed VBench scores were slightly lower than those of L2P, the additional three-seed evaluations show higher mean VBench scores for SGSE at both FT budgets.

We believe that identifying the representational limitations of historical velocity forecasting and revealing the complementary information available in sampling states provide a valuable perspective for advancing efficient diffusion inference and inform future research toward high-quality one-step generation.

Table R1.1: SSIM and PSNR (mean ± standard deviation over three seeds). Results are reported on DrawBench for FLUX.2 and VBench-200 for Wan2.1.

ModelFTL2P SSIM ↑SGSE SSIM ↑L2P PSNR ↑SGSE PSNR ↑
FLUX.2180.8728 ± 0.00460.8791 ± 0.004422.2788 ± 0.220122.5673 ± 0.2139
FLUX.2110.8141 ± 0.00380.8262 ± 0.004120.2964 ± 0.038420.5455 ± 0.0778
FLUX.290.7640 ± 0.00350.7723 ± 0.003118.3504 ± 0.066218.5022 ± 0.0817
Wan2.1180.7536 ± 0.00300.7628 ± 0.002220.4336 ± 0.041320.6845 ± 0.0387
Wan2.1110.6401 ± 0.00400.6445 ± 0.003017.6226 ± 0.050517.7511 ± 0.0870
AuthorsOpenReview ↗

Response to W2

W2. Question about whether the additional state forecasting mechanism is fundamentally necessary.

Thank you for your interest in simplifying SGSE. Under Euler sampling, the simplified velocity forecast can be expressed as

vˉi=v^i+ηi(I−Pi)zi.\bar{v}_i = \hat{v}_i + \eta_i (I-P_i)z_i.

This simplification follows from Eq. (28): zj=zi−∑ℓ=ji−1Δtℓvℓ,z_j = z_i - \sum_{\ell=j}^{i-1}\Delta t_\ell v_\ell, which relies on the Euler update rule. For other samplers, Eq. (28) does not necessarily hold and the same simplification does not directly apply (Appendix G.2, pages 19–20, lines 1020–1056).

We evaluated this simplified formulation on FLUX.2 (Euler) and Wan2.1 (Flow-UniPC). The results are shown in Table R1.2. Under Euler sampling, direct current-state span expansion achieves performance close to SGSE, consistent with our analysis. In contrast, under Flow-UniPC, this baseline performs worse than velocity forecasting without state guidance, while SGSE achieves the best results across all four metrics. Moreover, compared with direct current-state span expansion, the total computational cost of SGSE increases by only 0.000024% on FLUX.2 and 0.000039% on Wan2.1, respectively. These results show that SGSE provides a more general approach applicable to different samplers, achieves better practical performance, and requires negligible additional computation compared with the simplified formulation.

Table R1.2: Comparisons under the 11-FT schedule. FLUX.2 uses Euler and DrawBench; Wan2.1 uses Flow-UniPC and VBench-200. Results are from Tables 3, 4, 12, and 13 of the revised manuscript (page 10, lines 487–503 and page 20, lines 1027–1041).

ModelMetricVelocity onlyDirect stateSGSE
FLUX.2SSIM ↑0.82650.83470.8358
FLUX.2PSNR ↑20.854521.026221.0309
FLUX.2LPIPS ↓0.14690.14420.1432
FLUX.2CLIPScore ↑0.82180.82150.8228
Wan2.1SSIM ↑0.64380.63520.6450
Wan2.1PSNR ↑17.750917.696117.8253
Wan2.1LPIPS ↓0.32460.32770.3223
Wan2.1VBench ↑84.137883.900284.2584
AuthorsOpenReview ↗

Response to W3

W3. The paper shows that expanding the historical velocity span reduces the theoretical minimum approximation error. However, this result does not necessarily imply that.

Thank you for your attention to the relationship between our theoretical analysis and practical performance gains. As derived in Eqs. (16)–(19) (page 6, lines 287–307), the minimum achievable squared approximation error after expanding the historical velocity span is

=||g_i||_2^2(1-A_i^2),$$ where $g_i$ is the missing velocity component outside the historical velocity span, and $A_i$ measures the directional alignment between $g_i$ and the state-guided direction. When $A_i$ is nonzero, the theoretical minimum squared approximation error decreases from $||g_i||_2^2$ to $||g_i||_2^2(1-A_i^2)$. Although the practical gains achieved are not very substantial, the improvements are consistently observed. We believe that identifying the representational limitations of historical velocity forecasting and validating the effectiveness of span expansion through theoretical analysis and experiments provide a solid foundation for future research on finding better-aligned complementary directions and developing more effective methods for efficient diffusion inference.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear Area Chair and Reviewers,

We sincerely appreciate your careful reviews and constructive feedback.

In this work, we identify the representational limitations of velocity forecasting for efficient diffusion inference and demonstrate that complementary information from sampling states can improve forecasting accuracy. As highlighted by the reviewers, our work provides useful insights into the limitations of historical velocity forecasting (FC8v, 6Qjt, 2GXq), introduces a targeted correction with low computational overhead (6Qjt, 2GXq), and achieves consistent fidelity improvements across image and video generation (FC8v, 6Qjt, 2GXq).

We have carefully revised the manuscript to address your concerns. For ease of review, the changes are highlighted in blue in the revised manuscript. The main revisions are summarized as follows:

  • Sampler-specific analysis: Clarified the Euler-specific simplification and explained why it does not directly apply to Flow-UniPC (Appendix G.2).
  • Simplified baseline comparisons: Added direct current-state and initial-noise span expansion baselines under Euler and Flow-UniPC (Appendix G.3, Tables 12 and 13).
  • Reduced-step FT comparisons: Added reduced-step FT comparison method (Sec. 4.1, Tables 1 and 2, Fig. 4).
  • Multi-seed and worst-case analysis: Reported the mean and standard deviation of evaluation metrics across three generation seeds and analyzed worst-case performance (Appendix C, Tables 7–9).
  • Additional ablation studies: Extended the key component and mechanism ablations to Wan2.1 (Sec. 4.4, Appendix E, Tables 4 and 11).
  • Calibration analysis: Reported calibrated γi\gamma_i values across timesteps and clarified the role of learned scaling through the γi=1\gamma_i=1 ablation (Sec. 4.4, Appendix F, Fig. 10).
  • Experimental clarifications: Clarified the sources of the historical trajectories and reference velocities used in Figs. 2 and 3 (figure captions).

We look forward to your further feedback.

Sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.