acceptodds
Under review as a conference paper at ICLR 2027

Geoshrink: Accelerating Diffusion Transformers with Two Lines of Code

Abstract

Diffusion transformers incur substantial inference cost through repeated model evaluations along a sampling trajectory. We introduce GeoShrink, a training-free acceleration method that retains the original solver grid while evaluating the model only at a prescribed set of anchors. At skipped stages, GeoShrink predicts the solver-facing output by adding a geometrically retained fraction of the latest observed innovation to the most recent exact output. We derive this rule from chordal tangent transport and round-trip line projection, and establish a geometric anchor-spacing principle that minimizes the largest adjacent gap expansion under fixed coverage and first span. The analysis characterizes the geometric closure and propagation of prediction errors without assuming access to future model outputs. Experiments cover image, video, motion, and audio generation, together with adapted 3D backends. At approximately acceleration, GeoShrink improves FLUX PSNR by  dB over the strongest listed baseline. On HunyuanVideo, it achieves a reported speedup and improves ChronoMagic-Bench-150 PSNR by  dB over the strongest listed fidelity baseline. Comparisons at fixed evaluation budgets further show substantial gains on Motion, Audio, Music and 3d generation.

Primary area:
generative models
Keywords:
diffusion transformers, accelaration, generation models

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 3 reviews, above 85% of the others.
Reviewer_u6TnOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes GeoShrink, a training-free acceleration method for diffusion transformers that preserves the original solver grid while reducing model evaluations through sparse anchor queries. The method introduces a geometric retention rule to predict velocity fields from historical model outputs, along with a geometric anchor scheduling strategy to allocate exact model evaluations under a fixed computational budget. The authors provide theoretical analysis of the retention mechanism, anchor scheduling, and prediction error propagation. Extensive experiments across image, video, motion, audio, music, and 3D generation demonstrate strong acceleration performance and improved generation fidelity compared with existing training-free acceleration baselines.

Critical strengths
  1. The proposed solver-level velocity prediction is simple, training-free, and architecture-agnostic, offering an elegant alternative to model-specific feature caching methods.
  2. The method is supported by clear mathematical analysis and extensive experiments across multiple generative domains, demonstrating strong acceleration performance while preserving generation quality.
Critical weaknesses

The theoretical connection between geometric anchor scheduling and actual velocity-field prediction accuracy requires further clarification.

Questions for authors

Thank you for the solid work. The paper is well presented and supported by comprehensive experiments. My only question concerns the geometric anchor scheduling strategy. Could the authors further explain why geometric spacing is preferred and how it relates to velocity-field prediction accuracy? A clearer justification would strengthen the theoretical framework. I would like to raise my score to 4 if this concern be clearly clarified.

4 comments

Thank you for the positive assessment of GeoShrink's solver-level prediction rule and its experimental coverage. We address the question about the connection between geometric anchor spacing and velocity-prediction accuracy below.

Q1. Why is geometric spacing preferred, and how does it relate to velocity-field prediction accuracy?

The schedule controls how quickly the prediction horizon can grow after the three warm-up queries. This is the quantity that enters the geometric retention rule.

Let the anchors be ordered in an increasing progress coordinate ss,

\qquad h_j=s_{j+1}-s_j.$$ The first three queries are consecutive, so $h_1=h_2=1$. Define the maximum adjacent gap-expansion ratio $$\begin{aligned} \gamma &= \max_{2\le j\le B-2}\frac{h_{j+1}}{h_j}. \end{aligned}$$ For a skipped update in the cell following $h_j$, Lemma 1 gives $$0\le |\rho_t^{(s)}| \le \frac{h_{j+1}}{h_j} \le \gamma.$$ Since the geometric exposure is $R_t^{(s)}=(\rho_t^{(s)})^2\sin^2\theta_t$, the retention rule satisfies $$R_t^{(s)}\le \gamma^2, \qquad \omega_t^{(s)} =\frac{1}{1+R_t^{(s)}} \ge \frac{1}{1+\gamma^2}.$$ Thus, a smaller $\gamma$ limits the worst-case normalized horizon and gives a larger lower bound on the retained fraction of the latest innovation. The paper also makes explicit the distinction between the progress coordinate used to construct the schedule and the physical $\sigma$ coordinate used by the solver. If $-\sigma=\psi(s)$ and $$0<m\le \psi'(s)\le M_\psi, \qquad K=\frac{M_\psi}{m},$$ then the corresponding physical-coordinate guarantee is $$|\rho_t|\le K\gamma, \qquad \omega_t\ge \frac{1}{1+(K\gamma)^2}.$$ For an affine grid, $K=1$. For a nonuniform grid, the bounded-distortion factor is necessary; an index-space gap ratio alone does not certify the physical prediction horizon. ### Why the minimax solution is geometric Under a fixed number of exact queries, the total coverage is fixed: $$\sum_{j=1}^{B-1}h_j=T-1, \qquad h_1=h_2=1.$$ The schedule objective is $$\min_{\\{h_j\\}} \max_{2\le j\le B-2}\frac{h_{j+1}}{h_j} \quad \text{subject to}\quad h_1=h_2=1,\quad \sum_{j=1}^{B-1}h_j=T-1.$$ Let $\gamma$ be the maximum ratio of a feasible schedule. Recursively applying $h_{j+1}\le \gamma h_j$ gives $$h_j\le \gamma^{\max(j-2,0)}.$$ Therefore, $$T-1 =\sum_{j=1}^{B-1}h_j \le \sum_{j=1}^{B-1}\gamma^{\max(j-2,0)}.$$ The right-hand side is increasing in $\gamma$. Let $r_\star\ge 1$ be the unique solution of $$\sum_{j=1}^{B-1}r_\star^{\max(j-2,0)}=T-1.$$ Every feasible schedule must have $\gamma\ge r_\star$, and equality is attained only when every recursive inequality is tight: $$h_j=r_\star^{\max(j-2,0)}.$$ Hence, the unique continuous minimax solution is $$\begin{aligned} h_1,h_2,h_3,\ldots &= 1,1,r_\star,r_\star^2,\ldots. \end{aligned}$$ The solver uses discrete indices, so the continuous anchor locations are projected onto the original grid with order-preserving guards. The resulting integer gaps are approximately geometric; the reported $\gamma$ values are computed from those discrete gaps.

Experimental protocol and schedule comparison

We evaluate the schedules on SD3.5-Medium at 1024×10241024\times1024, using Pick-a-Pic-100, CFG 4.5, the original 28-step reference grid, and five random seeds. Each schedule uses seven exact model queries. The first three anchors and the terminal anchor are fixed; only the intermediate anchors change. All schedules use the same solver, sigma schedule, prediction rule, and sampling configuration.

The manuscript reports anchor indices in reversed order (28 to 1):

ScheduleAnchor indices (28 → 1)γ\gamma
Geometric28, 27, 26, 24, 21, 14, 12.333
Cosine28, 27, 26, 22, 14, 5, 14.000
Log-SNR28, 27, 26, 17, 5, 2, 19.000
Uniform-Sigma28, 27, 26, 14, 8, 4, 112.000
Karras28, 27, 26, 6, 3, 2, 120.000

To test whether the schedule-level bound is reflected in actual field prediction, we additionally evaluate the true model output at each skipped timestep using the accelerated trajectory's current latent, timestep, conditioning, and CFG. These diagnostic forward passes do not update the sampler and are excluded from the seven-query budget. We compare that output with GeoShrink's predicted field at the same state.

SchedulePSNR ↑SSIM ↑LPIPS ↓MSE ↓Cosine ↑Angle (°) ↓Relative L2 ↓
Geometric19.8510.73550.36060.13090.971612.430.2256
Cosine19.4840.65370.43860.14310.968913.130.2353
Log-SNR18.1080.70550.45230.13680.969913.080.2320
Uniform-Sigma17.7330.70230.41550.13150.971112.730.2266
Karras16.4960.61590.63800.13680.971012.730.2289

Geometric has the smallest gap-expansion ratio and the best result on each reported generation and same-state velocity metric. The MSE difference from Uniform-Sigma is small (0.1309 versus 0.1315); the evidence should therefore be read as a consistent but modest advantage in this five-seed diagnostic, not as a universal law that a smaller γ\gamma always produces lower empirical MSE.

Finally, Appendix E.12 makes the connection to the full solver trajectory explicit. For an augmented solver state and update maps satisfying the stated stability assumptions, Proposition 9 gives

\left\lVert \widetilde X_0-X_0^\star \right\lVert_2 &\le \sum_{n=1}^T \beta_n \left\lVert \epsilon_n \right\lVert_2 \qquad \cdot \prod_{k=1}^{n-1} \alpha_k. \end{aligned}$$ where $\epsilon_n$ is the field defect at stage $n$. This is a conditional perturbation bound: it connects local field errors to final solver-state deviation under explicit assumptions on the solver. It does not claim that $\gamma$ alone deterministically bounds final image error. The revision therefore clarifies the chain of reasoning at three levels: Section 2.3 (pages 4–5) derives the schedule-level exposure bound and the unique continuous minimax solution; Section 3.6 (Table 7, page 10) reports generation fidelity and same-state velocity diagnostics; and Appendix E.12 (pages 42–44) gives the full-grid error-propagation result. For completeness, the broader evaluation results in Appendix D.3 are organized as HPSv3 in Table 11 and the combined GenEval/Qwen-Image-Bench results in Table 12(a)–(b).
Reviewer_u6TnOpenReview ↗

Concerns Addressed

Thank you for the detailed response and additional experiments. The authors have fully addressed my concern regarding geometric anchor scheduling. The theoretical explanation and comparisons with alternative schedules provide convincing support for the proposed design. I have two minor suggestions for the final revision. First, I recommend including the new scheduling comparisons and velocity-field prediction results in the main paper, as they provide valuable evidence for the effectiveness of geometric anchoring. Second, it would be helpful to clarify the distinction between the theoretical bound on prediction horizon and the empirical improvement in prediction accuracy. But I still appreciate the authors' efforts and decide to raise my score to 4.

Thank you very much for your thoughtful follow-up. We are glad that our explanations and additional experiments have addressed your concerns.

Regarding your first suggestion, the table above includes both scheduling comparisons and velocity-field prediction results. Specifically, Cosine similarity, Angle, and Relative L2 measure the agreement between the predicted velocity and the reference velocity under different scheduling strategies, while PSNR, SSIM, and LPIPS measure the similarity between the generated images and their reference images.

We also appreciate your second suggestion. We will add a clearer discussion distinguishing the theoretical bound on the prediction horizon from the empirically observed improvement in prediction accuracy as soon as possible.

Thank you again for your constructive feedback and helpful suggestions.

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

This work proposes a training-free method for DiT acceleration by proposing GeoShrink, predicting the output of the solver on the original grid while evaluating the model at a set of anchors. The paper claimed that this work was derived from chordal tangent transport and round-trip line projection. Experiments show improved results over baselines for multiple applications on generation.

Critical strengths
  1. I think the main idea of GeoShrink looks interesting - while the underlying principles were not clear, using the exact nonuniform query, which are optimized over timesteps, and its prediction method with past velocities are interesting.
  2. This work shows strong experimental results over prior works including ResilPhase (while there are also missing recent works, though) on multiple applications with limited comparisons.
Critical weaknesses
  1. This work is very difficult to follow. A lot of symbols not defined well and I had to guess many of them. This writing issue really undermines the quality of this work seriously.
  2. This work proposes a number of interesting formulations, but without much explanation or insight. For example, Eq. (1) introduces omega without any explanation about it and then all of the sudden, omega, which is a line overlap, was introduced in Proposition 1. What is the justification of this? Moreover, there are a number of notations not defined for Lemma 1 and Theorem 1, even some important ones such as r_star. Thus, it is very hard to see any new exactly. I put my score for informativeness as 2, but it should be 1.5 unless the revision resolves all these issues.
  3. Some of the comparisons are not complete. For some applications, only the original and this work were compared, not other acceleration methods. Moreover, there are also other fast methods including caching, spatial acceleration and distillation. It will be especially interesting if the proposed method also works for distilled models.
Questions for authors

I think this work is interesting, but its writing, explanation and insufficient comparisons are my major concerns. See the above and address as much as possible. There are also a lot of propositions, but none of them were explained and used in the main text. I am not sure what is the point of having these propositions and the reader (like me) has to connect all of them to the main text. Note that the abstract claimed that this work was derived from chordal tangent transport and round-trip line projection, but these important parts were only scattered in the appendices. "chordal tangent transport" has never appeared in the main text. I also have a concern on Figure 8, which may be misleading. All other methods start from speedup x2 up to 6x, but the proposed method started from around 1.5x to 4x. Showing the results over fair ranges are important. Lastly, what is the meaning of this work over practical generative models distilled from the original models? Will this method outperform them? Will this method work on them? Please see below for some recent related works that were not discussed and compared.

  1. Lihui Gu et al., ScalingCache: Extreme Acceleration of DiTs through Difference Scaling and Dynamic Interval Caching, ICLR 2026
  2. Divya Jyoti Bajpai et al., FlowCast: Trajectory Forecasting for Scalable Zero-Cost Speculative Flow Matching, ICLR 2026
  3. Shikang Zheng et al., Forecast then Calibrate: Feature Caching as ODE for Efficient Diffusion Transformers, AAAI 2026
  4. Junwen Tan et al., VDE: Training-Free Accelerating Rectified Flow Model via Velocity Decomposition and Estimation, CVPR 2026
  5. Zhengyang Geng et al., Mean Flows for One-step Generative Modeling, NeurIPS 2025
  6. Zhengyang Geng et al., Improved Mean Flows: On the Challenges of Fastforward Generative Models, CVPR 2026
  7. Wongi Jeong et al., Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers, CVPR 2026
12 comments

Thank you for the careful reading and for identifying the parts of the paper that needed a clearer explanation or broader evaluation. We have revised the method exposition, added a global notation table, clarified the geometric motivation and its connection to the implemented predictor, added comparisons with recent inference-time methods, tested two few-step distilled models, and updated the speed–quality plot. We address the main points below.

Q1. What is the geometric meaning of the prediction rule, and how are the theoretical results used?

We have revised the exposition in Section 2.2 and added a global two-column notation table in the Appendix’s “Notation” section. The table collects the principal symbols used throughout the main text and technical appendices; auxiliary variables that occur only in local proofs are defined where they first appear.

Before Proposition 1, the revised text now gives an intuitive explanation of the predictor. GeoShrink does not attempt to infer an entirely new future velocity direction. Instead, it reuses the most recently observed innovation and estimates what fraction of that change should be retained at a skipped stage. In particular, ωt\omega_t is an innovation-retention coefficient, not a time-extrapolation factor: ωt=0\omega_t=0 reuses the latest exact output, whereas ωt=1\omega_t=1 retains the full latest innovation.

At a skipped stage, let vt(1)v_t^{(1)} be the latest exact output and define the latest and preceding observed innovations as

\Delta v &= v_t^{(1)}-v_t^{(2)},\\\\ \Delta v^- &= v_t^{(2)}-v_t^{(3)}. \end{aligned}$$ The prediction family is $$v_t(\omega)=v_t^{(1)}+\omega\Delta v,\qquad 0\leq\omega\leq1.$$ The angle $\theta_t$ measures the turning between the two innovation lines. The normalized horizon $\rho_t$ measures the distance from the current solver-update midpoint to the latest exact anchor, relative to the span between the two latest anchors. Their combination defines the geometric exposure $$R_t=\rho_t^2\sin^2\theta_t.$$ We have also clarified why the retention coefficient is defined through the **squared line overlap**. For unit representatives $u_t$ and $\widetilde u_t$ of the original and transported innovation lines, let $P_{u_t}=u_tu_t^\top$ and $P_{\widetilde u_t}=\widetilde u_t\widetilde u_t^\top$ be their rank-one orthogonal projectors. Then $$\operatorname{tr}(P_{u_t}P_{\widetilde u_t})=|\langle u_t,\widetilde u_t\rangle|^2.$$ This quantity is a parameter-free, sign-invariant measure of agreement between two one-dimensional subspaces: it is unchanged if either unit representative is multiplied by $-1$, equals one for coincident lines, and equals zero for orthogonal lines. Equivalently, it is the fraction of the transported unit-direction energy retained by projection back onto the original innovation line. We use this projector overlap instead of introducing an independently learned or manually tuned retention coefficient.

Proposition 1 makes this construction precise. It forms the chordal tangent component of the preceding innovation direction, transports the latest innovation direction by the normalized horizon, and projects the transported direction back onto the latest innovation line. The resulting coefficient is

\omega_t &=\operatorname{tr}(P_{u_t}P_{\widetilde u_t})\\\\ &=\frac{1}{1+\rho_t^2\sin^2\theta_t}. \end{aligned}$$ Thus, Proposition 1 establishes the innovation-retention coefficient used by GeoShrink. Substituting this coefficient into the prediction family $v_t(\omega)=v_t^{(1)}+\omega\Delta v$ gives the closed-form prediction rule in Eq. (5). Algorithm 1 applies that same Eq. (5) at each skipped stage, while evaluating the model exactly only at the selected anchors. We therefore make the connection from the geometric proposition to the implemented prediction rule explicit, without introducing an additional abstract predictor notation. Section 2.3 then connects the predictor to anchor placement. Lemma 1 bounds the normalized horizon by the adjacent anchor-gap ratio. Theorem 1 solves the corresponding fixed-budget minimax problem. If $h_j$ denotes the consecutive anchor gaps, the unique continuous solution is $$\begin{aligned} h_j &= r_\star^{\max(j-2,0)},\\\\ \sum_{j=1}^{B-1}r_\star^{\max(j-2,0)} &= T-1,\qquad r_\star\ge1. \end{aligned}$$ The paragraph immediately following Theorem 1 now emphasizes the meaning of this result: under the specified initial-gap constraints $h_1=h_2=1$ and fixed total trajectory coverage, the geometric schedule is the **unique continuous minimizer of the worst-case adjacent-gap expansion**. It is not selected empirically or by manually tuning an expansion ratio. The coverage equation uniquely determines $r_\star$, after which the continuous anchor locations are projected onto the discrete solver grid with order-preserving guards. We have also made the scope of the geometric argument more explicit. The rule is an observable chordal-transport-and-projection construction; it does not identify the unknown true future velocity direction from a finite history, nor does Theorem 1 claim to minimize the model-specific empirical velocity-prediction MSE. Appendix E.3 formalizes why smoothness and finitely many anchor observations alone cannot determine the future retention coefficient. Appendix E.9 distinguishes the retention coefficient from the amplitude of ordinary time-scaled linear extrapolation. Appendix E.12 gives a full-grid perturbation bound under explicit solver-stability assumptions. The core construction and schedule theorem are in the main text (Sections 2.2–2.3), while the detailed derivations and limitations are collected in Appendix E.

Q2. How does GeoShrink compare with recent acceleration methods?

We added the comparisons in Appendix D.1 (Table 9). GeoShrink is evaluated at approximately 2.8×2.8\times acceleration on SD3.5-Medium and approximately 5×5\times on FLUX.1-dev. The table reports generation-quality metrics. To make the comparison protocol explicit, latency excludes encoding and decoding, and each speedup is measured against the corresponding full-NFE reference using the same timing protocol. For each example and seed, the compared methods and the corresponding reference use identical inputs, conditioning, and initial noise; all other inference settings are held fixed except for the acceleration configuration.

BackboneMethodPSNR ↑SSIM ↑LPIPS ↓ImageReward ↑
SD3.5-MediumGeoShrink24.740.84500.20780.792
SD3.5-MediumVDE20.600.74190.23300.834
SD3.5-MediumScalingCache21.120.74090.27610.738
SD3.5-MediumFoCa20.260.72570.29180.715
SD3.5-MediumRALU11.250.42400.70570.781
FLUX.1-devGeoShrink21.620.79030.25520.947
FLUX.1-devVDE17.590.64760.41470.915
FLUX.1-devScalingCache17.510.70610.32750.923
FLUX.1-devFoCa16.970.68560.35540.906
FLUX.1-devRALU12.240.45820.68580.803

On SD3.5-Medium, GeoShrink gives the best PSNR, SSIM, and LPIPS, while VDE has a higher ImageReward score (0.834 versus 0.792). On FLUX.1-dev, GeoShrink gives the best result on all four reported metrics.

We also clarify the scope of the remaining comparisons:

  • ScalingCache, VDE, and FoCa are included using the available implementations evaluated in our setup.
  • FlowCast is discussed in the related-work section, but an official implementation was not available to us at the time of evaluation, so we do not report a quantitative comparison.
  • MeanFlow and Improved Mean Flows learn fast generators through dedicated training objectives. Their published one-step results are not plug-in inference configurations of the unchanged SD3.5-Medium or FLUX.1-dev checkpoints, so comparing those scores directly would not be a controlled same-checkpoint comparison.
  • RALU reduces spatial computation through mixed-resolution latent upsampling, whereas GeoShrink reduces the number of exact model evaluations while preserving the solver grid.

These distinctions are stated in Appendix B; the quantitative comparison is in Appendix D.1 (Table 9). The complementary benchmark evaluation is reported in Appendix D.3: HPSv3 appears in Table 11, while GenEval and Qwen-Image-Bench are presented together as panels (a) and (b) of Table 12, respectively.

Q3. Does GeoShrink work on distilled generative models?

We added experiments on two officially released eight-step distilled DiT models: Z-Image-Turbo and ERNIE-Image-Turbo. Direct sampling and GeoShrink use the same five-NFE budget and the same reference grid.

ModelMethodNFEPSNR ↑SSIM ↑LPIPS ↓ImageReward ↑
Z-Image-TurboDirect518.950.74110.19970.9976
Z-Image-TurboGeoShrink521.160.80010.14741.0278
ERNIE-Image-TurboDirect515.850.60480.27691.0425
ERNIE-Image-TurboGeoShrink518.160.70730.17111.0633

For Z-Image-Turbo, GeoShrink increases PSNR by 2.21 dB and SSIM by 0.0590, reduces LPIPS by 0.0523, and increases ImageReward by 0.0302. For ERNIE-Image-Turbo, the corresponding changes are +2.31 dB, +0.1025, −0.1058, and +0.0208.

The comparison keeps the NFE budget fixed at five, so the gains do not come from additional model evaluations. These results support compatibility with the two tested few-step distilled models. The experiments are reported in Appendix D.2 (Table 10).

Q4. How has the speed–quality comparison in Figure 8 been revised?

We agree that the previous plot did not make the operating ranges sufficiently clear. Figure 8 has been replaced with a speed–quality comparison using a common speedup axis. The updated GeoShrink curve covers approximately 1.4×1.4\times to 7×7\times speedup, and the figure reports four metrics: PSNR, SSIM, LPIPS, and ImageReward.

This broader range makes the overlapping operating regions easier to inspect and also shows the quality degradation as acceleration becomes more aggressive. The revised figure and its accompanying discussion are in Section 3.6 (page 10).

Reviewer_WXX8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a training-free acceleration method for diffusion sampling. The method keeps the original solver grids and only calls the model forward at limited points. Velocities in the middle timestep are extrapolated from adjacent model outputs with a geometry-inspired formula. The author conducted experiments across several domains and shows strong empirical results.

Critical strengths
  1. Geometric interpretation is interesting and novel. The extrapolation formula includes an angle similarity and distance-based shrink, which makes sense to me.

  2. The experimental results are rich. The authors include experiments in image generation, video generation, and 3d generation.

  3. The extrapolation formula is intuitive and easy to follow.

  4. The paper shows improvement across multiple domains.

Critical weaknesses
  1. Although the generation results are improved. There is a lack of analysis results to show the difference between the true velocity and the extrapolated results.

  2. ImageReward is an old human preference model. The results lack some of the latest reward evaluations. Other metrics are related to reconstruction results.

  3. The settings for ablation studies are vague.

  4. The anchor schedule is tilmestep-inspired, first lacks ablation studies to support this is better. In the first section, the extrapolation is based on previous velocity. I wonder whether the schedule should be designed according to the velocity distribution for different time steps. More experiments are needed for the scheduling part.

Questions for authors
  1. Can you provide some analysis of the true velocity and extrapolated results?

  2. More evaluations are needed. Can you add more evaluations on HPSv3, Qwen-Image-Bench, and GenEval?

  3. Provide detailed settings for the ablation studies.

  4. Can you compare with more timestep spacing methods? Only linear is not enough.

7 comments

Thank you for the careful review and for recognizing the geometric interpretation, intuitive prediction rule, and broad experimental evaluation of GeoShrink. We address the four questions below.

Q1. How accurately does GeoShrink predict the model velocity at skipped timesteps?

Final-sample metrics do not directly measure the accuracy of the intermediate solver-facing fields. We therefore added a same-state velocity diagnostic in Section 3.6 (Table 7).

At each skipped timestep, we take the current latent produced by the accelerated trajectory and make an additional forward pass of the original SD3.5-Medium model using the same latent, timestep, text conditioning, and CFG combination. This gives the reference field vttruev_t^{\mathrm{true}} at exactly the same state as GeoShrink's prediction v^t\hat v_t. The diagnostic pass does not update the sampler and is excluded from the seven-query operational budget.

We report four quantities and five-seed results in Table 7 are:

Anchor scheduleMSE ↓Cosine ↑Angle (°) ↓Relative L2 ↓
Geometric0.13090.971612.430.2256
Cosine0.14310.968913.130.2353
Log-SNR0.13680.969913.080.2320
Uniform-Sigma0.13150.971112.730.2266
Karras0.13680.971012.730.2289

Geometric scheduling gives the best value on all four diagnostics. The difference from Uniform-Sigma is small in MSE (0.1309 versus 0.1315), and we do not interpret this result as a universal monotonic relationship between the gap ratio and empirical prediction error. Rather, the diagnostic shows that, in this setting, the geometric schedule's smaller worst-case gap expansion is accompanied by slightly better same-state velocity estimates.

Q2. What do the additional preference and compositional evaluations show?

We added HPSv3 results in Appendix D.3 (Table 11). The same appendix now presents GenEval and Qwen-Image-Bench together in Table 12, as panels (a) and (b), respectively. For all three additional evaluations, the baseline settings are kept identical to those used in Table 2. These metrics complement the reconstruction-oriented measures used in the main experiments.

HPSv3

MethodSD3.5-Medium ↑FLUX.1-dev ↑
GeoShrink1.6822.736
ResilPhase1.5822.767
TaylorSeer1.6302.737
SpeCa1.4952.492
TeaCache1.4272.436
ToCa1.556-3.703

GeoShrink obtains the highest HPSv3 score on SD3.5-Medium.

GenEval (Table 12(a))

GenEval evaluates compositional text-to-image generation.

FLUX.1-dev

MethodOverall ↑Attribute ↑Color count ↑Counting ↑Spatial ↑Single ↑Two ↑
GeoShrink0.65720.51000.75530.71250.18000.98750.7980
TeaCache0.65490.44000.76600.75000.18000.97500.8182
SpeCa0.64710.46000.74470.71250.16000.98750.8182
ResilPhase0.64610.44000.76600.72500.16000.98750.7980
ToCa0.63530.44000.73400.70000.17001.00000.7677
TaylorSeer0.63160.47000.69150.65000.17001.00000.8081

SD3.5-Medium

MethodOverall ↑Attribute ↑Color count ↑Counting ↑Spatial ↑Single ↑Two ↑
GeoShrink0.66920.57500.83780.61250.21500.98440.7904
TaylorSeer0.65420.50000.82710.61250.21000.98750.7879
ResilPhase0.65400.49500.83240.61250.21000.98120.7929
SpeCa0.65290.52750.82450.60620.20000.98120.7778
ToCa0.64800.51750.81910.59380.20500.98750.7652
TeaCache0.64730.55250.79520.59060.21500.97810.7525

GeoShrink obtains the highest overall GenEval score on both backbones (0.6572 on FLUX.1-dev and 0.6692 on SD3.5-Medium).

Qwen-Image-Bench (Table 12(b))

SD3.5-Medium

MethodTotal ↑Quality ↑Aesthetics ↑Alignment ↑Real-world ↑Creative ↑
GeoShrink38.6842.0243.3138.4739.9429.65
TaylorSeer37.5541.2341.5737.2739.4828.18
ResilPhase37.5241.2641.8137.1739.3328.03
ToCa36.9539.3041.6236.2639.2628.33
SpeCa36.3937.9041.4336.1438.7327.77
TeaCache35.5435.9940.3935.3138.7727.25

FLUX.1-dev

MethodTotal ↑Quality ↑Aesthetics ↑Alignment ↑Real-world ↑Creative ↑
GeoShrink44.6449.0750.6444.0742.0237.42
ResilPhase44.3849.1050.1544.2641.5536.82
TeaCache44.3348.5750.3443.8342.1936.73
ToCa44.0948.8250.0444.0341.6135.96
TaylorSeer43.2147.9448.4643.2741.2435.16
SpeCa42.7047.8148.6742.5140.0034.53

GeoShrink obtains the highest Total score on both backbones: 38.68 on SD3.5-Medium and 44.64 on FLUX.1-dev. On SD3.5-Medium, the margin over TaylorSeer is 1.13 points; on FLUX.1-dev, the margin over ResilPhase is 0.26 points.

Q3. What are the exact settings of the ablation studies?

Section 3.6 now specifies the common ablation protocol: SD3.5-Medium, Pick-a-Pic-100, CFG 4.5, and a 28-NFE reference sampler. Fixed-budget comparisons use seven exact model queries.

We clarify the coordinate domain of the Fixed baseline, which was not explicit in the previous wording. Fixed uses uniform spacing in the original solver-index domain, not in the noise-level or log-SNR domain. Let the original solver stages be indexed by k=0,…,T−1k=0,\ldots,T-1, with the first three warm-up anchors 0,1,20,1,2 and the terminal anchor T−1T-1 fixed. For a budget of BB exact queries, the target locations of the B−4B-4 intermediate anchors are

k~r=2+rB−3(T−3),r=1,…,B−4.\tilde k_r=2+\frac{r}{B-3}(T-3),\qquad r=1,\ldots,B-4.

which are mapped to distinct, order-preserving integer stages on the original solver grid. Thus, Fixed is a Uniform-Index baseline. We will label it “Fixed (Uniform-Index)” in the text to avoid ambiguity.

This is distinct from Uniform-Sigma in Table 7. Uniform-Sigma first places the intermediate target values uniformly in the noise coordinate:

σ~r=σ2+rB−3(σT−1−σ2),r=1,…,B−4.\tilde\sigma_r=\sigma_2+\frac{r}{B-3}(\sigma_{T-1}-\sigma_2),\qquad r=1,\ldots,B-4.

and then maps each target noise level to the corresponding available solver stage while preserving anchor order. Log-SNR is defined analogously in the log-SNR coordinate.

For the component ablation, we vary two factors independently:

  1. Anchor schedule: Fixed (Uniform-Index) or Geometric.
  2. Skipped-field prediction: Linear or geometry-aware Shrink.

The linear baseline uses the two latest exact outputs:

\hat v_t^{\mathrm{lin}} &=v_t^{(1)}+\alpha_t^{\mathrm{lin}}\Delta v,\\\\ \alpha_t^{\mathrm{lin}} &=\frac{\sigma_t-\sigma_t^{(1)}}{\sigma_t^{(1)}-\sigma_t^{(2)}}. \end{aligned}$$ All four variants share the same checkpoint, prompts, solver grid, query budget, and sampling settings; only the two factors above change. | Method | Anchor schedule | Prediction rule | PSNR ↑ | SSIM ↑ | LPIPS ↓ | ImageReward ↑ | |---|---|---|---:|---:|---:|---:| | Fixed-Linear | Fixed (Uniform-Index) | Linear | 9.882 | 0.404832 | 0.702306 | -1.2556 | | Geo-Linear | Geometric | Linear | 15.457 | 0.621274 | 0.414758 | 0.278629 | | Fixed-Shrink | Fixed (Uniform-Index) | Shrink | 19.556 | 0.692327 | 0.388319 | 0.559383 | | **GeoShrink** | **Geometric** | **Shrink** | **19.851** | **0.735519** | **0.360551** | **0.619176** | The two components contribute separately. Geometric scheduling improves the linear variant, while replacing linear extrapolation with Shrink improves both schedules. Their combination gives the best result across the four reported metrics.

Comments on the paper

1 comment

Summary of Reviewer Concerns and Revisions

We sincerely thank all reviewers for their constructive feedback and time. Below, we summarize the main concerns raised during the review process and the corresponding revisions and additional experiments.

Reviewer WXX8

  • Accuracy of intermediate velocity predictions. We added same-state velocity-field diagnostics in Section 3.6 (Table 7), reporting MSE, cosine similarity, angular error, and relative L2 error across five anchor schedules. Geometric scheduling achieves the best results across all four metrics.
  • Insufficient generation-quality evaluations. We added HPSv3, GenEval, and Qwen-Image-Bench evaluations on both SD3.5-Medium and FLUX.1-dev, providing additional evidence beyond reconstruction-oriented metrics.
  • Unclear ablation settings. We specified the complete experimental protocol, including the backbone, dataset, CFG scale, reference sampling steps, and exact query budget. We also clarified the distinction between Uniform-Index and Uniform-Sigma scheduling and added a component-wise ablation of anchor scheduling and prediction rules.
  • Motivation for geometric scheduling. We clarified the theoretical objective and its relationship to normalized prediction horizons and geometric exposure, while distinguishing the theoretical worst-case guarantee from empirical velocity-prediction accuracy.

Following the revision, the reviewer acknowledged that most concerns had been addressed and increased the score to 3 (Weak Accept).

Reviewer DuCR

  • Unclear core methodology and geometric motivation. We explicitly introduced chordal tangent transport and round-trip projection in the abstract, introduction, and main methodology, explaining how they lead to the implemented innovation-retention coefficient.
  • Disconnected theoretical results. We added a unified Theoretical Roadmap in Appendix E, organizing all propositions and theorems around the geometric construction, anchor scheduling, prediction-loss analysis, theoretical limitations, implementation extensions, and solver-level error propagation.
  • Notation and terminology issues. We consolidated the principal notation into a compact global table and clarified key terms, including innovation, exact query, retention coefficient, geometric exposure, and oracle coefficient.
  • Insufficient comparisons and related-work discussion. We expanded comparisons with recent acceleration methods, including ScalingCache, FoCa, VDE, and RALU, and discussed FlowCast, MeanFlow, and Improved MeanFlow. We also added dedicated discussions of TACache (NeurIPS 2026) and TC-Padé (CVPR 2026) to clarify the methodological distinctions.
  • Applicability to distilled models. We added fixed-budget experiments on Z-Image-Turbo and ERNIE-Image-Turbo, demonstrating improvements over direct sampling at the same five-NFE budget.
  • Potentially misleading speed–quality comparison. We revised Figure 8 to use a common speedup axis and provide a broader, more comparable operating range.

In the latest discussion, the reviewer stated that they had no further questions or requests and would carefully evaluate the revision and discuss it with the other reviewers and AC before adjusting their score.

Reviewer u6Tn

  • Theoretical justification of geometric scheduling. We clarified why geometric spacing minimizes the maximum adjacent gap-expansion ratio under the specified fixed-budget constraints, and how this controls the normalized prediction horizon and geometric retention exposure.
  • Connection to actual velocity-prediction accuracy. We added comparisons against Cosine, Log-SNR, Uniform-Sigma, and Karras schedules under matched experimental settings, together with same-state velocity-field diagnostics and generation-quality metrics.
  • Distinction between theoretical and empirical guarantees. We explicitly clarified that the scheduling theorem establishes a worst-case geometric bound, rather than guaranteeing minimum empirical prediction MSE. We also connected local prediction errors to full-grid solver deviations through a perturbation bound under stated stability assumptions.

The reviewer confirmed that the concerns had been fully addressed and increased the score to 4 (Clear Accept).

Overall Summary

Across the review process, we have substantially improved the methodological exposition, theoretical organization, experimental coverage, and reproducibility of GeoShrink. In particular, the revised manuscript now makes the central geometric construction explicit, connects the theoretical results into a unified framework, provides additional quantitative evidence, and more clearly delineates the scope and limitations of the proposed method.

We sincerely appreciate the reviewers' feedback, which has helped us improve the clarity, rigor, and completeness of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.