acceptodds
Under review as a conference paper at ICLR 2027

Rethinking What to Cache in Few-Step Diffusion Transformers: Solver-Aware Target Selection

Abstract

Diffusion Transformers (DiTs) can generate high-quality images and videos, but generating each sample requires multiple costly DiT forward passes. Two common ways to accelerate DiT sampling are step distillation, which reduces the number of sampling steps, and caching, which skips some DiT evaluations by reusing a tensor computed at an earlier step. Most caching methods decide in advance which tensor to reuse. After distillation, adjacent sampling steps are farther apart. Reusing a tensor across this larger gap introduces more error, so choosing what to cache becomes especially important. We therefore introduce AutoTarget, a method that chooses the cached tensor for a given model, solver, and reuse schedule. AutoTarget uses a small set of runs without cache reuse to measure the error caused by reusing each candidate tensor, then selects the candidate with the lowest error. We also analyze how an error at one reuse step affects the final sample. For Euler sampling, we identify cache targets that produce the same trajectory and show why a stored solver update may not. Experiments on distilled image and video DiTs show that the best cache target changes with the model, image resolution, and solver. AutoTarget reduces DiT evaluations and retained cache storage. Generation quality remains close to the corresponding uncached run. On the tested PixArt-LCM and FLUX.1-schnell settings, its calibration ranking matches the ranking from held-out cached runs. To help others reproduce the method, we provide its core implementation in an anonymous repository at

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Transformers, Few-Step Diffusion, Inference Acceleration, Cache Reuse, Cache Target Selection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bBEVOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The papers main idea appears to be simple and effective, however complex, over indexed, over use of different letters, or simply unclear notation makes it hard to understand. Let me note know that I did my best to understand it, however there could be parts which still not fully clear to me.

The goal of the paper is to accelerate the generation process of flow and diffusion models. The methods they suggest is to cache models outputs and trajectory states to skip some of the model evaluations along the generation trajectory. This is in contrast to other caching methods which suggest to cache the intermediates features of the model.

Given a pre-trained diffusion or flow matching model with parametrization yy (ϵ\epsilon-prediction, xx-prediction, or vv-prediction), AutoTarget set a time parametrization t0,t1,…,tTt_0,t_1,\ldots,t_T, and a solver Φ\Phi (e.g. Euler method). Denoting the states and model outputs along the trajectory by xjx_j and yjy_j (resp.), they propose a search algorithm to find a subset A⊂0,1,…,T−1\mathcal{A}\subset \\{ 0,1,\dots, T-1 \\} such that that trajectory can be approximated with xa,yax_a,y_a for a∈Aa\in\mathcal{A}.

The method is validate empirically on Flux.1, PixArt-LCM, and HunyuanVideo, showing improvements compared to previous caching methods.

Critical strengths

The proposed AutoTarget method does not require to change the weights of the models, and only optimizes the inference algorithm.

The efficacy of AutoTarget is demonstrated on large scale models, both text-to-image and text-to-video. Additionally, though it is not stated clearly, my understanding is that all baselines were reimplemented and reevaluated for a fair comparison.

AutoTarget performance is quite consistently better than the considered baselines.

Critical weaknesses

I have two main concerns for the paper. First the presentation of the paper makes it very hard to follow. Second the even though AutoTarget is referred as caching method for diffusion and flow models, it is actually much more close to solver learning methods[1,2,3,4] and the paper does not discuss these at all.

Regarding the presentation issue, I apologize if my writing is not 100% clear, but this is only as I had a hard time following the method section.

Presentation:

  1. The authors introduce the notation Φ\Phi to denote a general solver step. But the only provided example is Euler solver I found only one use of this notation in equation 5. If this is indeed the case, adding redundant notation makes it harder for the reader to follow. If it is not the case, then currently it is unclear what other solvers are used.

  2. The relation between \eps\eps-prediction, xx-prediction, and vv-prediction are well known and can be presented in a simple small table. Given the current state xjx_j, current time, tjt_j and one of the three prediction types the other two can be recovers, hence it is not clear why is it important which one is being cached. Additionally, the related model specific discussion at the beginning of the method section is hard to follow at this stage and it is also unclear. There is a use of both x^j\hat{x}_j and x~j\tilde{x}_j, the difference between them is still not clear to me. Model specific discussion should probably place in or just before the experiment section.

  3. While the index jj is clearly defined to to denote the times steps, between lines 162 to 264, there is a heavy use of the index aa even though its relation to the index jj is first defined in line 270.

  4. It still not clear to me what does the letter rr index and what is its domain. It is first introduced by "Let r index a solver-facing cache target considered by AutoTarget", and the meaning of "solver-facing cache target" is still unclear to me. This seems important as later it is the parameter that the method minimize over in equation 13.

  5. In equation 3 another two letters notations are introduced to the notation mm, and EE. The function EE is never explicitly defined, is it a conversion between the three prediction types? why not simply cache xa,ya,tax_a,y_a,t_a? Later on, the letter EE is used to denote accumulated error.

  6. In equation 4, the function DD is introduced, again this function not defined explicitly and I still do not understand it exactly what is does. The only thing clear is that the authors assume that the composition of EE and DD yields a an approximation to yjy_j. This seems rather important point which is not presented clearly.

  7. In equation 12, QQ is introduced and again it is never explicitly defined in the text, and it is still not clear to me what does QQ denotes. Again, it makes the text harder to follow.

  8. In line 269 the supscript ∗^* is introduce without any explanation, only later in line 285 in equation 13 the reader my understand that it is related to the minimizer w.r.t. rr. But this still does not make its meaning clear as it is already used inside the definition of the risk.

  9. in the risk it is not clear how does Jreuse\mathcal{J}_{\text{reuse}} is chosen.

learned solvers: Since the proposed method caches only the models outputs and states along the trajectory, and to my understanding the approximation for cached steps is a linear combination of previous states and model outputs, it is essentially a new solver. Thus it seems appropriate to include an extensive discussion about the relation of AutoTarget to such methods [1,2,3,4], and clearly state what its contribution on top of them. Additionally, it seems appropriate to include an empirical comparison with such methods.

[1] Zhao, Wenliang, et al. "Unipc: A unified predictor-corrector framework for fast sampling of diffusion models."

[2] Zheng, Kaiwen, et al. "Dpm-solver-v3: Improved diffusion ode solver with empirical model statistics."

[3] Shaul, Neta, et al. "Bespoke non-stationary solvers for fast sampling of diffusion and flow models."

[4] Frankel, Eric, et al. "S4s: Solving for a diffusion model solver."

Questions for authors

As the paper already present a good empirical results, I would appreciate if the author could address the weaknesses I have raised in the "Weakness" section.

Also could the authors please clarify for which tasks were AI tools were used and for which tasks AI tools were not used? The ICLR recommendation is to clearly state both and I find the provided statement somewhat vague. For example does statement "AI-assisted tools to support academic writing, language editing" means AI model essentially written a draft of the paper (including equations) and the authors have verified it or the authors have written most of the paper and an AI model just polished the text, fixed type grammar, typos, and so on? Another example, do not state whether AI model was used to develop the suggest method or not?

5 comments

We thank the reviewer for the detailed comments. We reorganized the method section, made the cache formulas explicit, and expanded the discussion of fast and learned sampling solvers.

1. Method inputs, outputs, and procedure

The revised method starts by specifying its inputs: a pretrained model, a solver, sampling coordinates, and fixed positions for full model evaluations. AutoTarget compares cache representations under these conditions and returns a target and an acceptance decision.

Calibration first records uncached trajectories. All candidates are then tested from the same recorded states and preceding refresh positions. During generation, full evaluations refresh the selected cache; the remaining steps reconstruct the input required by the original solver from that cache.

2. Notation and explicit formulas

We now define notation before its first use:

  • jj is the sampling-step index; a(j)a(j) is the most recent preceding full-evaluation step.
  • rr identifies a candidate cache representation in the finite set supported by the pipeline.
  • The initial state x0x_0 and the predicted clean endpoint xcleanx^{\mathrm{clean}} have distinct symbols.
  • The superscript ref\mathrm{ref} denotes the uncached reference trajectory; rselr_{\mathrm{sel}} denotes the selected target.
  • Φj\Phi_j denotes the original solver update, including its coordinates, coefficients, and required history.

A compact table now states what is stored and how it is reused for velocity, clean-endpoint, and noise-endpoint caching. Model-specific conversions are placed in the appendix. The native model prediction is itself included as a candidate.

We also explicitly define the scoring conversion QjQ_j. It converts each reconstructed prediction into the latent estimate supplied to the frozen VAE. Calibration averages decoded squared RGB error over prompts, all planned reuse steps, and RGB elements. These steps are fixed before scoring. The lowest-error candidate is accepted when the runner-up/best error ratio reaches the preset threshold of 1.10.

3. Why compare representations that can be converted into each other?

The familiar conversions apply at a common state and time. Cross-step reuse additionally requires specifying how an earlier stored tensor is used with the current state and time. The revised text gives these storage and reconstruction formulas explicitly.

For representations that yield the same Euler trajectory, we provide a proof and group them together, testing one representative per group. For a stored solver displacement, we explain that it includes the anchor step size; reusing it when the current step size differs can change the update. This makes clear which candidates are equivalent and which implement different reuse rules.

4. Relation to sampling solvers and additional experiments

We added UniPC, DPM-Solver-v3, BNS, and S4S to the related-work discussion. They share the goal of reducing model evaluations. UniPC constructs predictor-corrector updates; DPM-Solver-v3 uses empirical model statistics; BNS and S4S fit solver coefficients to reference outputs. AutoTarget fixes the solver, coordinates, and evaluation positions and selects which tensor to retain between evaluations. Its calibration selects a discrete candidate without fitting solver coefficients.

We added PixArt-LCM experiments with 16 calibration prompts and 100 independent test prompts. All methods use four full model evaluations; PSNR and SSIM use the native eight-step output as reference. AutoTarget achieves 14.5864 dB PSNR, 0.5559 SSIM, and 0.5676 ImageReward, compared with 12.8336 dB, 0.4989, and 0.5091 for native four-step LCM sampling.

The experiment also includes UniPC and DPM-Solver++. Direct scheduler replacement on the LCM-distilled checkpoint produces texture artifacts, motivating an additional comparison on native PixArt-XL with model-appropriate sampling settings.

5. Declaration of AI assistance

The research idea, core methodology, and overall research direction presented in this paper were independently conceived and developed by the authors.

AI tools were used to assist with code implementation, experimental development, and debugging under the guidance and supervision of the authors. During manuscript preparation, the authors independently determined the main arguments, technical content, and overall structure, while AI tools were employed to improve linguistic clarity, readability, and academic expression.

AI tools also assisted in mathematical derivations and formula development. The resulting derivations and equations were subsequently reviewed and verified by the authors for correctness and consistency.

The authors take full responsibility for the scientific content, experimental results, mathematical correctness, and conclusions presented in this paper.

Reviewer_bBEVOpenReview ↗

Reviewer response

In the authors' response they promised a revised version of the paper with improved presentation which was main concern at my initial review. Do the authors plan to upload this version before end of discussion period?

I want to stress that presentation was my main concern in the initial review because the paper is too hard to follow for me to properly assess the proposed method. Hence, I would not consider to increase the score before reviewing the revised version of the paper.

Moreover, the authors response regarding use of AI tools does not match the statement in paper: "We used AI-assisted tools to support academic writing, language editing, code development, debugging, and experiment implementation.". Following my inquiry, their response adds the use of AI tools for "mathematical derivations and formula development.". This suggest that indeed the Method section was heavily written by AI tools. While in itself, it is not a problem, I have strong concern that the authors did not verified the method section properly.

Yes—we have uploaded the revised manuscript to OpenReview prior to the end of the discussion period. Thank you for the candid feedback and for taking the time to re-evaluate our paper. Below, we summarize the presentation improvements made to Section 3 and detail how the authors verified the mathematical derivations and Method section.

1. Summary of Presentation Improvements in Section 3

To make the Method section substantially easier to follow, we restructured Sections 3.1–3.3 around concrete intuition, lighter notation, and clearer logical transitions:

  • Concrete Four-Step Running Example (Section 3.1): Before introducing the general storage and reuse operators (Storer,Reuser\mathrm{Store}_r, \mathrm{Reuse}_r), we added a self-contained four-step linear-flow Euler example (T=4,A={0,1,3},t4=0T=4, \mathcal{A}=\{0,1,3\}, t_4=0) showing directly why caching velocity v1v_1 or clean endpoint x1cleanx_1^{\mathrm{clean}} yields the exact same update at step 2, whereas caching step displacement u1=h1v1u_1 = h_1 v_1 shifts the next state by (h2−h1)v1(h_2 - h_1)v_1.
  • Reduced Subscript Overhead (Sections 3.1 & 3.3): We omitted the prompt index pp throughout Sections 3.1 and 3.2 when defining single-trajectory storage, reconstruction, and Euler equivalence, introducing p∈Pp \in \mathcal{P} only in Section 3.3 when averaging calibration risk across prompts.
  • Denominator Conditioning & Solver Dependence (Sections 3.1 & 3.2): Under the linear-flow conversion table, we explicitly stated the positivity requirements (tj>0t_j > 0 for clean endpoints and 1−tj>01 - t_j > 0 for noise endpoints) and noted that small denominators can amplify tensor and finite-precision perturbations. In Section 3.2, we also brought forward the solver-comparison finding from Table 8 (Appendix D.1)—showing that switching from multi-step DPM-Solver++ to DDIM on PixArt-XL causes clean-endpoint caching to produce non-finite values and reverses the lower-ranked targets while ϵ\epsilon remains best under both solvers.
  • Clearer Calibration Motivation & Theoretical Scope (Section 3.3): We clarified why candidates residing in different parameterizations and numerical scales are mapped through QjQ_j and the frozen VAE GG into a shared decoded RGB space for comparison. We also kept the scope of the error-propagation bound (Eqs. 7–8) explicit: it characterizes error accumulation in the candidate-specific complete solver state rather than directly guaranteeing decoded RGB rankings (Eq. 6), which motivates fixing the separation-ratio acceptance gate (γ≥1.10\gamma \ge 1.10) prior to held-out testing.

2. Consistency of AI Disclosure and Verification of the Method Section

We apologize for the inconsistency between our earlier response and the initial statement in the paper. We have updated the Declaration of AI Assistance (page 10) in the revised manuscript so that it explicitly and consistently states that AI tools assisted with mathematical derivations and formula development, in addition to code implementation, debugging, and linguistic editing.

Every derivation, formula, and algorithm in Section 3 and the appendices was independently checked, verified line-by-line, and cross-validated against our codebase by the authors:

  • Analytical Verification of Euler Equivalence and Step-Size Coupling (Proposition 1, Eq. 4, Appendix B.1): The authors manually verified the induction proof establishing xj=xa−(ta−tj)vax_j = x_a - (t_a - t_j)v_a across velocity, clean-endpoint, and noise-endpoint reconstructions (Eqs. 10–17), as well as the exact single-step discrepancy (hj−ha)va(h_j - h_a)v_a for stored solver displacement (Eq. 19).
  • Verification of Error-Accumulation Recursion and Scope (Eqs. 7–8, Appendices B.2–B.3 & E): We checked the complete-state local Lipschitz recursion Ej+1(r)≤L‾jEj(r)+Δj(r)E_{j+1}^{(r)} \le \overline{L}_j E_j^{(r)} + \Delta_j^{(r)} and explicitly documented its boundaries in Section 3.3, Appendix B.3, and Appendix E so that state-space error bounds are not conflated with decoded RGB risk rankings.
  • Code-Level Numerical Cross-Checks (Appendix A.1): All storage and reconstruction formulas (Storer,Reuser\mathrm{Store}_r, \mathrm{Reuse}_r) and pipeline conversions (QjQ_j) were cross-checked directly against our implementation, including pre-set numerical equivalence checks on recorded calibration traces (e.g., BF16 terminal latent relative L2≤10−6L_2 \le 10^{-6} and decoded RGB MSE ≤10−10\le 10^{-10} for PixArt-LCM's denoised and clean predictions).

We hope the revised Section 3 and these clarifications make the proposed method straightforward to follow and fully address your concerns. Please let us know if any remaining part of the derivation or presentation warrants further detail.

Reviewer_ud4vOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AutoTarget is a training-free procedure for choosing which solver-facing tensor to cache for a fixed model, solver, precision, and reuse schedule. It evaluates candidate targets from identical exact states, ranks them using decoded RGB error, and deploys the best target only when its error is sufficiently separated from the runner-up. The paper also analyzes equivalence among velocity, clean-endpoint, and noise-endpoint caching under Euler sampling.

Critical strengths
  1. The controlled target-selection evidence is strong. With model, solver, schedule, storage, prompts, and seeds fixed, PixArt-LCM at 1024 resolution obtains 23.155 dB using denoised/clean, versus 18.340 for velocity, 12.844 for denoised residual, and 10.007 for epsilon. Changing only the resolution reverses the selected target, and the selected target wins all 20 held-out prompts in both controlled settings.
  2. Calibration is carefully designed. All candidates are evaluated from the same exact state, avoiding accumulated-error confounding. The abstention rule correctly rejects the shifted LTX-Video case where candidates are insufficiently separated. The Euler equivalence and stored-displacement derivations are also correct under their stated assumptions.
Critical weaknesses
  1. The paper does not establish that solver-output caching adds value over removing the corresponding steps. Under the stated Euler equations, repeatedly reusing an anchor velocity reaches the same next anchor as one coarse Euler transition. Table 4 empirically confirms this: cached and anchor-only PSNR are 32.575 versus 32.577 for FLUX and differ by only 0.016 dB for PixArt. Anchor-only latency is not reported, and this control is missing for HunyuanVideo. Thus, the main Euler results may add redundant solver transitions rather than provide a useful caching approximation.
  2. The practical generality and compute fairness remain limited. AutoTarget must be recalibrated whenever the model, solver, resolution, precision, or schedule changes. Baselines use different reuse schedules, while Tables 1–2 do not report their full DiT-evaluation counts, so target and schedule effects remain conflated. The local decoded-MSE objective also has no theoretical guarantee of preserving final task-quality rankings; the paper itself shows that fidelity improvements do not correlate with CLIPScore improvements.
Questions for authors
  1. Does anchor-only sampling at the identical anchor coordinates produce the same Euler trajectory? Please report its latency and task-quality metrics for every main setting, especially HunyuanVideo.
  2. What are the exact DiT-call counts and anchor schedules for every baseline in Tables 1–2? Do the conclusions hold on matched-NFE quality-latency frontiers?
  3. On the same checkpoint and coordinate schedule, can changing only the solver change the selected winning target?
  4. Does the decoded-error calibration ranking remain stable on substantially larger and distribution-shifted prompt sets, and how does it compare with latent-error or propagation-weighted objectives?
3 comments

We thank the reviewer. We added direct step-reduction controls, matched-budget comparisons, and calibration-stability experiments.

Q1. Does direct sampling at the same anchors produce the same Euler trajectory?

For deterministic Euler sampling, holding the same velocity across several small steps is equivalent to one larger step between the same anchors, provided that model inputs and conditioning agree at the shared anchors. This explains the nearly identical FLUX results.

We added a HunyuanVideo comparison at 256x256 resolution with 33 frames and 20 held-out prompts. AutoTarget retains 20 solver updates, while direct sampling uses only the anchor-to-anchor transitions. Both methods evaluate the model at exactly the same positions, with budgets of 8, 10, or 12 calls.

Latency uses a separate L20X measurement with three prompts and five timed repeats. Results are AutoTarget / direct sampling:

  • 8 calls: PSNR 22.873 / 22.867 dB; SSIM 0.7690 / 0.7688; latency 1.515 / 1.553 s.
  • 10 calls: PSNR 23.081 / 23.079 dB; SSIM 0.7739 / 0.7741; latency 1.842 / 1.824 s.
  • 12 calls: PSNR 28.185 / 28.234 dB; SSIM 0.8902 / 0.8908; latency 2.097 / 2.134 s.

Video-quality scores are also close. At 8 calls, subject consistency is 95.617 / 95.698, background consistency is 96.779 / 96.876, and motion smoothness is 99.109 / 99.110. These results agree with the Euler equivalence.

We also added SANA-Sprint-1.6B results on 100 prompts. With the same three model calls and evaluation positions, AutoTarget retains four solver updates and achieves 27.65 dB PSNR and 0.8677 SSIM, compared with 22.839 dB and 0.7379 for direct three-call sampling. The benefit of retaining updates and reusing a suitable target therefore depends on the model and sampling configuration.

Q2. What are the actual call counts, and do differences remain at matched budgets?

We added the image-baseline configurations and measured call counts:

  • AutoTarget, TaylorSeer, and HiCache: full evaluations at steps {0,1,2}, with reuse or prediction at step 3; three calls.
  • FLUX TeaCache: four calls on every prompt.
  • FLUX DiCache: {0,1,3} on 196 prompts and all four steps on four prompts; 3.020 calls on average.
  • PixArt TeaCache: {0,1,3} on every prompt; three calls.
  • PixArt DiCache: {0,2,3} on 121 prompts, {0,1,3} on 50, and all four steps on 29; 3.145 calls on average.

Step indices are zero-based. These adaptive counts come from the supplemental profiling runs.

We additionally compared native-velocity caching, TaylorSeer, and HiCache on HunyuanVideo at 704x704 resolution with 129 frames and 20 test prompts, using four H800 GPUs. All methods retain 20 solver updates and share the same evaluation positions at each budget:

  • 8 calls: {0,1,2,5,9,13,17,19}.
  • 10 calls: {0,1,2,5,9,11,13,15,17,19}.
  • 12 calls: {0,1,2,3,5,7,9,11,13,15,17,19}.

PSNR values, in the order native-velocity caching / TaylorSeer / HiCache, are:

  • 8 calls: 27.53 / 25.27 / 26.67 dB.
  • 10 calls: 27.53 / 25.40 / 26.77 dB.
  • 12 calls: 33.34 / 32.90 / 33.46 dB.

Corresponding generation times are 62.91 / 64.42 / 64.15 s, 72.85 / 74.22 / 73.95 s, and 82.74 / 83.92 / 83.78 s. Thus, differences remain when call counts and evaluation positions are matched; HiCache is slightly better in PSNR at 12 calls.

Q3. Can changing only the solver change the selected target?

We fixed the PixArt-XL checkpoint, resolution, precision, prompts, seeds, 20 sampling coordinates, and refresh positions, and changed only the solver between DDIM and second-order DPM-Solver++.

The lowest-error calibration candidate is clean under DDIM and velocity under DPM-Solver++. Their separation ratios are 1.001394 and 1.004734, respectively. Both fall below the preset threshold of 1.10, so neither configuration enables caching; both fall back to all 20 model calls.

Q4. Is calibration stable on larger and shifted prompt sets? How does RGB scoring compare with alternatives?

We calibrated on 128 PartiPrompts and tested on 200 separate PartiPrompts and 200 GenEval prompts. Texts are disjoint across these sets, and choices are frozen before testing.

In the four-step configuration, the selected denoised target has the lowest final RGB error among candidates on 200/200 PartiPrompts and 199/200 GenEval prompts. In the eight-step repeated-reuse configuration, it is best on 200/200 prompts in both sets.

We also repeated calibration-subset sampling 1,000 times per subset size. In the eight-step configuration with four prompts per subset, the top-ranked candidate agrees with the full-calibration result in:

  • 99.3% of trials for RGB scoring.
  • 65.0% for clean-latent error.
  • 97.0% for next-transition latent error.
  • 82.9% for propagation-weighted scoring, which additionally uses 16 calibration prompts to estimate the weights.

These results show that RGB scoring finds the same top-ranked target more consistently with small calibration sets.

We apologize for the oversight. Due to an error on our part during the upload process, the revised PDF was not successfully uploaded earlier, so you could not access the updated version. We have now successfully uploaded the revised manuscript. Thank you for your patience and for considering the updated version.

Thank you for your time and the additional experiment, which addresses my key concern.

I will maintain my positive score.

Reviewer_xHa2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes AutoTarget, a training-free method for selecting which solver-facing representation to cache in diffusion transformers. For a fixed model, solver, and reuse schedule, it compares pipeline-compatible targets, such as velocity, denoised predictions, and solver updates, using local decoded RGB error on uncached calibration trajectories. The selected representation is then reused to skip full DiT forward passes. The paper also analyzes equivalence between cache representations under Euler sampling and the propagation of reuse errors. Experiments on FLUX.1-schnell, PixArt-LCM, and HunyuanVideo evaluate inference latency, generation quality, output fidelity, and retained cache storage. Additional experiments assess target selection on held-out prompts and compare caching with reduced-step sampling at the same DiT evaluation budget.

Critical strengths

The paper offers an insightful perspective by shifting the caching focus from block-wise intermediate features to solver-facing representations of a full forward pass, enabling full DiT evaluations to be skipped with a compact cache.

Critical weaknesses
  1. Limited motivation and practical contribution in the main few-step experiments. In the four-step PixArt-LCM and FLUX.1-schnell settings, AutoTarget saves only one DiT evaluation. PixArt-LCM explicitly reuses cached outputs only at the final step, while the reuse position for FLUX.1-schnell is not clearly specified. I find the practical value of this restricted reuse regime limited. Moreover, Table 4 shows nearly identical PSNR to a reduced-step baseline with matched anchor positions and DiT evaluation counts, leaving the additional contribution of cache-target selection unclear.

  2. Missing essential details in the main video experiment. The CFG-distilled, 20-step HunyuanVideo setting could demonstrate the method’s value under repeated cache reuse. However, Table 2 does not clearly specify AutoTarget’s cache schedule or actual DiT evaluation count, nor connect its results to the 8/10/12-evaluation settings reported elsewhere. These omissions hinder reproducibility and assessment of comparison fairness, weakening the evidence that could otherwise support the method’s broader practical contribution.

Questions for authors
  1. What cache schedule is used for HunyuanVideo, and how is the RGB-based selection criterion justified? Please specify which of the 20 steps in Table 2 reuse cached outputs and how many DiT forward passes are performed. Since subsequent sampling steps may amplify or correct errors, how do you verify that local RGB error reliably reflects final output error, especially during early, high-noise stages and when caching is applied at multiple steps?

  2. Can you evaluate higher cache reuse ratios on more recent few-step models, such as reducing DiT evaluations from 8 to 4 for Z-Image-Turbo or FLUX.2 Turbo LoRA? Reusing cached outputs only at the final step of the current four-step PixArt setting offers limited practical value.

  3. What are the specific baseline configurations in Table 1? Please report the caching parameters, such as TaylorSeer’s NN and OO, and the total MACs per complete generation to enable a comparison of computational cost.

  4. What additional benefits does AutoTarget provide over directly reducing the number of sampling steps? In Table 4, the two approaches achieve nearly identical PSNR when the number of DiT evaluations and the anchor positions are matched.

  5. How much does RGB-based scoring improve upon latent-space error scoring? Do the gains justify the additional VAE decoding and calibration costs, particularly for video models?

5 comments

We thank the reviewer for recognizing the perspective of selecting solver-facing cache targets. We address the concerns below and have added the corresponding configurations and results to the revision.

Q1. HunyuanVideo schedule and local RGB scoring.

Table 2 uses 20 solver updates with 8 full DiT evaluations at zero-indexed steps {0,1,2,5,9,13,17,19}; the remaining 12 steps reuse cached outputs. Latency decreases from 231.5 s to 150.4 s (1.539x).

We additionally evaluate whether local selection predicts final output fidelity on HunyuanVideo at 256x256 with 33 frames, using 16 calibration and 20 disjoint test prompts. For budgets of 8/10/12 full evaluations, local RGB scoring selects the native velocity target. Its final mean RGB MSE is 0.007349/0.007528/0.003202, the lowest among the candidate targets at each budget. It is also the best target on 20/20, 20/20, and 18/20 individual test prompts. This experiment measures final errors after the complete trajectory, including repeated reuse and subsequent error propagation. The first three steps remain full evaluations in these schedules; the experiment does not test reuse from the first high-noise step.

Q2. Higher reuse on a recent few-step model.

We added Z-Image-Turbo experiments at 512x512, reducing 8 solver steps to 4 full DiT calls at {0,1,3,5}, with 4 calibration and 50 test prompts. AutoTarget selects velocity with separation ratio 1.145, exceeding the fixed threshold 1.10.

Under fixed refresh positions, results are listed as latency / aggregate PSNR / ImageReward:

  • AutoTarget: 319.5 ms / 21.46 dB / 0.9366.
  • DBCache: 348.1 ms / 18.89 dB / 0.8158.
  • DBCache + first-order Taylor prediction: 348.3 ms / 20.28 dB / 0.9251.

The uncached latency is 577.1 ms, giving AutoTarget a 1.81x speedup. AutoTarget executes 50.0% of uncached DiT work, versus 54.7% for the two block-cache baselines because they retain probe-block computation. Retained cache storage is 0.125 MiB for AutoTarget, versus 15.47/38.67 MiB for the two baselines.

Q3. Table 1 configurations and computation.

We added the following FLUX/PixArt settings:

  • TeaCache: thresholds 0.4/0.55.
  • DiCache: thresholds 0.4/0.391, p=1; FLUX r=0.2.
  • TaylorSeer: first-order prediction (O=1), full calls at {0,1,2}, prediction at step 3. These four-step runs use an explicit refresh schedule, without a separate periodic N. The video configuration uses N=3, O=1.
  • HiCache: local scaled-Hermite implementation, orders 2/1, coefficient 0.5, full calls at {0,1,2}.
  • AutoTarget: velocity/denoised targets, full calls at {0,1,2}, reuse at step 3.

Total MACs per complete generation, including text encoding, sampling, and VAE decoding, are listed as FLUX/PixArt in TMAC:

  • Uncached: 145.3743 / 22.1500.
  • TeaCache: 145.3745 / 18.8962.
  • DiCache: 111.9376 / 19.4673.
  • TaylorSeer: 110.6491 / 18.8962.
  • HiCache: 110.6491 / 18.8952.
  • AutoTarget: 110.6410 / 18.8952.

Adaptive-method totals are averaged over 200 prompts. The appendix also reports the observed evaluation schedules.

Q4. Benefits over directly reducing sampling steps.

For Euler sampling, reusing velocity over an interval can be algebraically equivalent to a direct step between the same anchors. This explains the near-identical results in Table 4. The benefit of target selection depends on the model, solver, and candidate representations.

We added SANA-Sprint-1.6B results on 100 prompts, with all three methods using 3 full DiT calls. Results are PSNR / SSIM / ImageReward / CLIP:

  • Native 3-step sampling: 20.460 / 0.6672 / 1.0503 / 26.8051.
  • Same-anchor direct 3-call sampling: 22.839 / 0.7379 / 1.0858 / 26.7555.
  • AutoTarget, retaining 4 solver updates: 27.65 / 0.8677 / 1.0902 / 26.8522.

AutoTarget improves fidelity by 4.81 dB PSNR and 0.1298 SSIM over the same-anchor baseline. Target choice also matters in PixArt-LCM: denoised, velocity, residual, and epsilon caching yield 23.155, 18.340, 12.844, and 10.007 dB PSNR under the same schedule.

Q5. RGB versus latent scoring and calibration cost.

In the added HunyuanVideo experiment, RGB scoring and the two latent-space criteria select the same native target and produce identical final outputs. Thus, this experiment shows no additional selection benefit from RGB decoding, and latent scoring is cheaper.

For the measured four-prompt PixArt calibration, total offline calibration takes 10.457 s, including 1.528 s for reference VAE decoding and 7.533 s for candidate decoding. This cost is paid once for the fixed model/solver/schedule configuration, not per generated sample.

We apologize for the oversight. Due to an error on our part during the upload process, the revised PDF was not successfully uploaded earlier, so you could not access the updated version. We have now successfully uploaded the revised manuscript. Thank you for your patience and for considering the updated version.

Thank you for the additional experiments. They have helped me better understand the proposed method. However, after further examining the method and experimental results, I still have several concerns regarding the motivation, experimental design, and baseline fairness.

1. Motivation and Practical Significance of the Few-Step Experiments

  1. Although the paper emphasizes Few-step Diffusion Transformers in its title, I believe the motivation is not sufficiently established. AutoTarget is essentially a general Cache Target Selection Framework that can also be applied to standard diffusion inference, without introducing any specific designs for few-step or distilled diffusion models. Therefore, I do not fully understand why the authors specifically emphasize Few-step Diffusion as the research setting, or what particular challenges in few-step generation make this cache-target selection approach necessary.

  2. Both FLUX and PixArt use only four sampling steps, with caching applied only at the final step (three DiT forward passes and one reuse step). I find this setting insufficiently motivated, as caching only the final step provides limited practical insight into the benefits of the proposed method.

2. Fairness of the HunyuanVideo Baselines

I believe the baseline comparison on HunyuanVideo is still insufficient to establish the practical advantages of AutoTarget over existing methods. The current paper only uses TaylorSeer with N=3N=3 and O=1O=1 on the 20-step CFG-distilled HunyuanVideo model, which may not adequately represent the competitive performance of existing baselines. I would encourage the authors to further tune the configurations of TaylorSeer and other baseline methods and provide matched-latency comparisons. For example, what generation quality can these methods achieve at the same inference latency as AutoTarget? Such comparisons would provide a fairer assessment of AutoTarget's practical advantages over existing methods.

3. Prompt Evaluation and Generalization

  1. The image experiments primarily rely on DrawBench and ImageReward-related evaluations. I believe the authors should include additional compositional generation benchmarks, such as GenEval, to further evaluate generation quality. Furthermore, the current image experiments mainly rely on 200 DrawBench prompts, which I consider a relatively limited evaluation scale and insufficient to convincingly demonstrate generalization across different prompt distributions.

  2. The authors have not sufficiently demonstrated whether Cache Target Selection generalizes across prompts of different lengths and levels of complexity. In particular, it remains unclear whether cache targets selected using a small calibration set remain reliable for longer prompts involving multiple objects and complex semantic relationships. I would encourage the authors to provide additional experiments to investigate this.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.