acceptodds
Under review as a conference paper at ICLR 2027

ReCache: Learning Budget-Aware Caching Schedule for Diffusion Models via REINFORCE

Abstract

Modern diffusion models generate high-quality images and videos, but their iterative denoising process makes inference expensive. Feature caching accelerates sampling by reusing or predicting intermediate activations across diffusion steps, exploiting computational redundancy along the reverse trajectory. In this work, we align caching schedule selection with the goals of inference: preserving fidelity to uncached outputs and improving generation quality under a fixed budget of full model evaluations. We propose **ReCache**, a policy-gradient method that learns a budget-conditioned scheduling policy to select which diffusion steps to fully recompute. Training combines an output-fidelity objective with a generation-quality reward, requiring no labeled data or backpropagation through the full denoising trajectory. ReCache supports both feature reuse and feature forecasting; for each model and caching mechanism, a single trained policy adapts across computational budgets at inference time without retraining. On FLUX.1-dev, ReCache reduces LPIPS by 31% compared to DiCache at a FLOPs reduction. On Wan2.1, at a FLOPs reduction, it reduces LPIPS by 45% and improves VBench by 6.77 points over uniform HiCache.

Primary area:
generative models
Keywords:
diffusion models, inference acceleration, feature caching, budget-aware scheduling, reinforcement learning, REINFORCE, image and video generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_5pfPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReCache, an offline-trained, budget-conditioned policy that uses REINFORCE to select diffusion steps for full computation. Its objective combines fidelity to uncached outputs with a perceptual-quality reward. For each model and caching mechanism, one policy serves multiple budgets, and the resulting deterministic schedules are reused across generations. Experiments on FLUX.1-dev, Wan2.1, and HunyuanVideo show promising improvements, particularly under tight computational budgets.

Critical strengths
  1. ReCache provides a practical framework for optimizing caching schedules under explicit computational budgets. It enforces the target number of full model evaluations exactly, and a single trained policy serves multiple budgets for a given model and caching mechanism. Compatibility with both feature reuse and forecasting increases its practical value.

  2. The evaluations demonstrate substantial gains in challenging low-budget settings across image and video backbones. For example, on Wan2.1 with HiCache at k = 7, ReCache reduces LPIPS from 0.512 to 0.283 and raises VBench from 70.450 to 77.217. Appendix C.2 also reports mean fidelity improvements across four optimization seeds on FLUX.1-dev, lending further support to the results.

Critical weaknesses
  1. Limited evidence for the proposed explanation of the gains. The comparisons with DPCache and LeMiCa support the effectiveness of the learned schedules. However, the paper motivates complete-schedule optimization by interactions between errors introduced in different cached segments, while the ablations do not isolate the contribution of the final-output objective from differences in search strategy and quality reward. The results therefore support improved scheduling, but do not establish how much of the gain comes from avoiding independently estimated segment costs. This limits the support for the proposed explanation, rather than demonstrating a failure of the method.

  2. Insufficiently established contribution over closely related methods. BudCache (ICML 2026) is particularly relevant: it searches step-level caching schedules under a fixed budget by evaluating final-output fidelity, using simulated annealing and hill climbing. ECAD (ICLR 2026) also searches caching configurations using final-image quality, although at finer granularity and with a Pareto objective. Spectrum (CVPR 2026) includes a non-uniform schedule motivated by error accumulation, while its main contribution is spectral feature forecasting. These works are not discussed in the manuscript. ReCache's budget-conditioned policy and strict k-subset selection are clear design differences, but their added value over the closest alternatives is insufficiently established. This makes it difficult to assess the significance of the methodological contribution beyond the reported improvements.

Table 4 provides useful evidence that joint training across budgets improves on independently trained predictors. It does not, however, compare against a single budget-independent logit vector trained across the same budgets. The nested schedules in Section 4.3 motivate this simpler alternative. The remaining question is the added benefit of budget conditioning over shared training with a common ranking, rather than whether multi-budget training is useful.

Questions for authors
  1. How strongly do the authors intend to attribute the gains to interactions between caching errors? Is there evidence that distinguishes the benefit of the final-output objective from the search strategy and quality reward, such as a controlled comparison using the same caching mechanism and budget? Alternatively, can the explanatory claim be narrowed to what the current results establish?

  2. What is ReCache's added contribution relative to BudCache and ECAD in its objective, constraints, scheduling effectiveness, and offline optimization cost? Please prioritize evidence under comparable settings. Spectrum's scheduling rule could provide a simple non-uniform reference where applicable; its forecasting mechanism is a distinct contribution.

  3. Table 4 compares joint budget-conditioned training with independently trained per-budget predictors. Under the same multi-budget training protocol, what additional benefit does the budget-conditioned MLP provide over a single shared logit vector whose top-k entries define the schedule for each budget?

5 comments
AuthorsOpenReview ↗

Comparison with trajectory-level objective and BudCache

We thank the reviewer for the valuable comments and suggestions. Below we address weaknesses and questions.

Gains attributed to the final-output objective

To isolate the contribution of the final-output objective, we replace our PLPIPS −7⋅-7\cdotHPS loss with DPCache-inspired trajectory-level feature MAE. It compares final-block features of the actual cached rollout with the uncached reference, averaging over all dimensions and denoising steps. The rest of the setup is the same.

Table 1. Objective comparison under the same search procedure and caching budgets for FLUX.1-dev with Taylor-DP (O=2O=2).

BudgetObjectiveLPIPS ↓HPS ↑IR ↑IR DB ↑
7Feature MAE0.5820.2690.7510.544
Final-image (default)0.4900.2991.0921.012
8Feature MAE0.5820.2770.8450.658
Final-image (default)0.4180.3051.1261.045
9Feature MAE0.5270.2941.0420.895
Final-image (default)0.4100.3051.1331.004
13Feature MAE0.5040.2951.0350.895
Final-image (default)0.2440.3061.1231.034

Our final-image objective (PLPIPS −7⋅-7\cdotHPS) improves all reported metrics at every evaluated budget under the same search procedure. This supports the benefit of our chosen objective over the feature-based MAE alternative.

At the same time, we do not claim that every final-output objective will outperform any trajectory-level one. The related discussion in the Introduction was intended as a motivation behind our design choice; this is not a general claim about superiority of the final-output objectives. We will add this new ablation to the paper and refine our discussion of error interactions to better reflect what these experiments establish.

Contribution relative to BudCache

Objective and constraints. BudCache is a closely related baseline that also optimizes final-output fidelity under a fixed NFE budget. We use only BudCache’s Simulated Annealing (SA) and Hill Climbing (HC) schedule-search component, which directly addresses the schedule-optimization problem studied here. Its original objective is latent MSE, with a separate search for each budget. ReCache uses REINFORCE to learn one budget-conditioned policy across budgets, optimizing PLPIPS −7⋅-7\cdotHPS with exactly kk computed steps.

Effectiveness under comparable settings. We compare on FLUX.1-dev with Taylor-DP (O=2O=2) using the same 64 training prompt–noise pairs. We evaluate two variants: BudCache (original loss), using latent MSE, and BudCache (ReCache loss), using our objective. Both BudCache variants share the same search hyperparameters, preserving the original 10:1 SA/HC ratio with 20 SA iterations and 2 HC passes per budget.

Table 2. BudCache vs ReCache, FLUX.1-dev with Taylor-DP (O=2O=2).

kMethodLPIPS ↓HPS ↑IR ↑IR DB ↑
7BudCache (original loss)0.5100.3001.0640.940
BudCache (ReCache loss)0.5020.2981.0760.999
ReCache0.4900.2991.0921.012
8BudCache (original loss)0.4310.3031.1101.041
BudCache (ReCache loss)0.4450.3011.1021.023
ReCache0.4180.3051.1261.045
9BudCache (original loss)0.4360.3011.1001.009
BudCache (ReCache loss)0.3770.3031.1101.017
ReCache0.4100.3051.1331.004
13BudCache (original loss)0.3110.3061.1251.038
BudCache (ReCache loss)0.2840.3051.1201.042
ReCache0.2440.3061.1231.034

ReCache achieves superior quantitative results in most settings. Providing BudCache with ReCache's objective does not close the gap (the exception is LPIPS at k=9k=9). This comparison indicates that ReCache's gains cannot be generally attributed only to the loss function; REINFORCE-based optimization also has influence on the resulting quality.

Offline cost. BudCache used $\sim$16 GPU-h with its original loss and $\sim$22 GPU-h with ReCache’s loss for four separate searches at NFE 7, 8, 9, and 13. ReCache used approximately $\sim$16 GPU-h for one policy covering NFE 2–24, which already provides schedules for the remaining budgets. Extending BudCache to the same coverage within the same total time would require lowering the amount of computations for each budget on average.

AuthorsOpenReview ↗

Comparison with ECAD

Contribution relative to ECAD

Objective and constraints. ECAD runs a genetic search (NSGA-II) over a component-level FORA mask, optimizing ImageReward and MACs as two objectives. ReCache trains a step-level schedule with REINFORCE for an exact budget of kk computed steps, optimizing fidelity and quality (PLPIPS −7⋅-7\cdotHPS), and works with any caching mechanism.

Effectiveness under comparable settings. We compare ReCache with ECAD on FLUX.1-dev with the same prompts, noises, budgets and metrics. Specifically, in Table 4 we compare the original ECAD with its published search settings against ReCache with the FORA caching mechanism. Additionally, we isolate contribution of the RL-based optimization by comparing ReCache against ECAD restricted to ReCache's setting (learning only the caching schedule, Taylor-DP caching mechanism, ReCache objective), so that the only difference is the optimization procedure (Table 5). At budget kk, ECAD uses its best found schedule with cost at most kk full steps.

Table 4. Original ECAD vs ReCache FORA, FLUX.1-dev. All methods are evaluated at 1024$\times$1024.

kkMethodLPIPS ↓HPS ↑IR ↑IR DB ↑
7uniform FORA (N=7)0.5670.2700.8080.628
ECAD0.5940.2790.9310.771
ReCache FORA0.4920.2921.0250.981
8uniform FORA (N=6)0.5500.2780.9040.711
ECAD0.5860.2810.9400.758
ReCache FORA0.4390.2971.0410.991
9uniform FORA (N=5)0.5300.2850.9680.826
ECAD0.5850.2820.9430.759
ReCache FORA0.3700.2971.0370.987

Table 5. ECAD vs ReCache, both with Taylor-DP and our objective, FLUX.1-dev. All methods are evaluated at 1024$\times$1024.

kkMethodSearch res.PLPIPS −7⋅-7\cdotHPS ↓LPIPS ↓HPS ↑IR ↑IR DB ↑
7ECAD256-1.5030.5320.2891.0120.865
ECAD1024-1.5680.5370.2991.0630.924
ReCache1024-1.6180.4900.2991.0921.012
8ECAD256-1.5150.5260.2901.0100.861
ECAD1024-1.6440.5010.3051.1080.984
ReCache1024-1.7290.4180.3051.1261.045
9ECAD256-1.5160.5270.2901.0150.860
ECAD1024-1.7090.4460.3061.0920.989
ReCache1024-1.7410.4100.3051.1331.004

Here, ReCache outperforms ECAD on most metrics at every budget in both settings. Both ECAD variants were given the same time as the corresponding ReCache training (19 and 16 GPU-h for Tables 4 and 5). With optimizer being the only difference in Table 5, the results indicate that in this problem at equal training time REINFORCE generally finds more suitable schedules than the genetic search.

AuthorsOpenReview ↗

Comparison with Spectrum; budget-conditional training,

Comparison with Spectrum

Spectrum suggests a hand-crafted non-uniform schedule with dense early steps and growing intervals afterwards. Below we train ReCache on Spectrum's forecaster with its original hyperparameters and compare the three schedules.

Table 6. Comparison of different schedules for the Spectrum forecaster, FLUX.1-dev.

kScheduleLPIPS ↓HPS ↑IR ↑IR DB ↑
7uniform0.5610.2730.8510.642
Spectrum0.6420.198-0.359-0.059
ReCache0.4760.2951.0310.980
8uniform0.5430.2800.9250.723
Spectrum0.4920.2690.8520.828
ReCache0.4080.2981.0730.977
9uniform0.5220.2870.9840.828
Spectrum0.4530.2911.0700.992
ReCache0.3320.3001.0850.987
13uniform0.4720.2971.0620.953
Spectrum0.2390.3031.0911.015
ReCache0.2350.3031.0991.001

Spectrum's schedule produces substantially worse results at k=7k = 7. ReCache achieves best LPIPS, HPS and IR at every budget; Spectrum's schedule achieves higher IR DB at k≥9k \ge 9.

Benefits of budget-conditional training

Thank you for the valuable suggestion! Below we compare three parameterizations on FLUX.1-dev with Taylor-DP (O=2O=2) using a common multi-budget training setup:

  • Direct logits: one trainable logit vector shared across budgets;
  • Fixed-input MLP: a trained MLP with a fixed random input, producing budget-independent logits;
  • Budget-conditioned MLP: the same MLP architecture with a budget embedding as input.

Table 7. Policy parameterization comparison on FLUX.1-dev with Taylor-DP (O=2O=2).

BudgetParameterizationLPIPS ↓HPS ↑IR ↑IR DB ↑
7Direct logits0.5380.2941.0630.935
Fixed-input MLP0.5440.2891.0080.947
Budget-conditioned MLP (default)0.4900.2991.0921.012
8Direct logits0.4670.3011.1081.004
Fixed-input MLP0.5140.2941.0370.952
Budget-conditioned MLP (default)0.4180.3051.1261.045
9Direct logits0.4100.3031.1121.029
Fixed-input MLP0.4130.3011.0960.977
Budget-conditioned MLP (default)0.4100.3051.1331.004
13Direct logits0.2870.3061.1171.017
Fixed-input MLP0.2220.3061.1111.025
Budget-conditioned MLP (default)0.2440.3061.1231.034

Budget conditioning consistently improves all reported metrics over both alternatives at budgets 7 and 8, and remains strongest on most metrics at larger budgets. Direct logits perform better only on IR DrawBench at budget 9, while the fixed-input MLP performs better only on LPIPS at budget 13. These results support budget conditioning as the default under joint multi-budget training, with particularly consistent gains under tight budgets. We've added this comparison to the revised version of the paper.

Reviewer_8tYFOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors address the scheduling problem in diffusion feature caching, arguing that cache-step selection should be optimized for final output quality rather than uniform intervals, online error heuristics, or decomposed surrogate costs. The authors formulate budget-aware k-subset selection as a reinforcement learning problem and train a budget-conditioned MLP policy using Plackett–Luce sampling with Gumbel-Top-k and a REINFORCE leave-one-out estimator.

Critical strengths
  • The paper is well structured and easy to follow, with clear motivation, method presentation, and experimental organization.

  • The proposed method achieves strong empirical performance.

  • The experiments cover multiple generation model,FLUX.1-dev, HunyuanVideo, and Wan2.1, which demonstrates the generality of ReCache.

Critical weaknesses
  • Many workd have explored using Reinforcement Learning to control the cache mechanism in diffusion models. For example, "RAPID^ 3: Tri-Level Reinforced Acceleration Policies for Diffusion Transformer" (ICLR 2026). The authors should clearly discuss the difference and the superiority of the propsoed method over them.

  • The evaluation only considers budgets down to k=7 full-computation steps. Could the authors further reduce the budget to 4 or 5 steps? I recommend including a budget–performance curve to demonstrate how ReCache behaves as the number of full-computation step is reduced beyond the current range , and to compare it against naive step reduction.

Questions for authors

See weakness

3 comments

We thank the reviewer for the valuable discussion and comments. Below we address the weaknesses.

RL in concurrent works

We agree that prior works, including RAPID$^3$ [1], BlockDance-Ada [2], and OnlineCache [3], have used RL for caching. What distinguishes ReCache is its explicit budget constraint: a single trained policy provides inference-time control over the exact number of recomputed steps. These methods instead make successive binary caching decisions and regulate computation through soft penalties in the loss function. In Table 12 (Appendix C.4) we have compared our Gumbel-TopK design with a Bernoulli policy using a soft budget penalty: the latter struggles to meet the prescribed step count and often yields worse quantitative results.

Comparing ReCache specifically to RAPID$^3$, the papers pursue complementary goals. RAPID$^3$ jointly optimizes caching, sparse attention, and step skipping to learn a multi-component acceleration strategy. ReCache performs only caching schedule selection under a fixed compute budget, learning which steps to recompute and how to maximize their benefit. Our study delivers practical gains (as we quantitatively show in the paper), while enabling a focused isolated analysis of the learned schedules and their qualitative properties. ReCache could also be jointly trained with other acceleration techniques within a broader framework such as RAPID$^3$ when a more granular control over computations is desired.

We will adapt this discussion to the paper in the camera ready revision.

Further budget reduction

Thank you for the valuable suggestion! Below we extend the FLUX.1-dev Taylor-DP (O=2O=2) evaluation to k∈{3,4,5}k \in \lbrace3, 4, 5\rbrace, comparing ReCache with uncached kk-step inference.

Table 1.

kMethodLPIPS ↓HPS ↑IR ↑IR DB ↑
3Step reduction0.6640.171−1.106-1.106−0.875-0.875
ReCache0.6900.2200.1370.108
4Step reduction0.6160.2190.0890.047
ReCache0.6110.2450.5750.472
5Step reduction0.5850.2500.5500.420
ReCache0.6180.2670.8480.656

ReCache improves all preference metrics at these budgets, while LPIPS results are mixed. We have also added the suggested budget–performance curves for LPIPS and HPS at k∈{3,4,5,7,8,9,13}k \in \lbrace 3,4,5,7,8,9,13 \rbrace in Figure 7, Appendix C.3 (Further Analysis) of the revised paper. The figure shows that ReCache consistently outperforms step reduction in preference score (HPS), while its fidelity to full-inference samples generally improves as the number of full-computation steps increases.

 

References:

[1] RAPID$^3$: Tri-Level Reinforced Acceleration Policies for Diffusion Transformer;

[2] BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers;

[3] OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference.

Thank you. My concerns have been well addressed. However, I still recommend that the authors incorporate the newly added discussions and experiments into the manuscript, as ICLR policy allows such revisions and doing so would make the revised manuscript more complete.

Thank you for your response. We have uploaded the revised manuscript, which addresses the comments of all reviewers. The changes related to your comments are:

  • a discussion of concurrent works that use RL (Appendix A);
  • low-budget performance curves (Appendix C.3, Figure 7).
Reviewer_eqK8OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ReCache, a new framing and method for learning when diffusion models should perform full computation vs when they can cache/reuse or forecast cached features. A lightweight, budget-conditioned policy is learned, that selects a fixed number of full-computation steps and is trained with REINFORCE. The objective used combines fidelity/similarity to the full-inference output with an HPSv2.1 quality/human preference-standin reward. Training treats generation as a black box and does not backpropagate through the denoising trajectory. For a given backbone and caching mechanism, one trained policy supports multiple computation budgets. Experiments on FLUX.1-dev, Wan2.1 and HunyuanVideo combine ReCache with several existing caching mechanisms and report improved output fidelity and generally competitive or better quality than existing scheduling baselines, particularly at tight computation budgets.

Critical strengths

1: Substantial gains from better scheduling at the same inference cost.

ReCache learned schedules are modified while holding the caching mechanism and budget fixed, which helps isolate the scheduling contribution. For example, the paper reports FLUX LPIPS improving from 0.456 to 0.316 versus DiCache at the same FLOPs. Similarly, on Wan at seven full evaluations, ReCache improves over DPCache from 0.435 to 0.303 LPIPS and 70.898 to 76.462 VBench. The evaluation covers both image and video models, several caching mechanisms, and multiple budgets. This is the strongest empirical support for the paper’s usefulness.

2: A practical formulation with explicit budget control.

ReCache optimizes complete schedules against final-output fidelity and quality, accounting for the combined effects of caching decisions. One policy per backbone/mechanism supports multiple budgets. Schedules are computed once and reused across generations, so deployment requires no per-sample policy evaluation. The framework also accommodates existing reuse and forecasting mechanisms.

Critical weaknesses

1: Missing ECAD comparison.

The paper does not compare against ECAD. Both appear to occupy the same space and address the same problem, but with slightly different algorithims and approaches. While ECAD uses a genetic algorithim, ReCache uses REINFORCE. Additionally, ECAD learns the caching jointly with the schedule, while ReCache doesn't. In addition to comparison with ECAD, I would have liked to see if jointly learning the caching via REINFORCE would have been more effective than using existing caching methods. A direct explanation of the methodological differences and a matched comparison would help establish the contribution beyond changing the search procedure from genetic optimization to RL + the above caching note. Without a matched ECAD comparison, it is unclear how ReCache compares in generation quality at the same inference budget, or in the offline compute required to obtain its schedules.

2: The benefit of RL is not isolated.

The experiments establish that learned schedules can improve generation, but do not establish whether RL is responsible for the advantage or whether the same formulation and objective would work as well with direct differentiable optimization. A comparison using the same harness, with an appropriate relaxation and the same hard inference budget, would clarify this, including the quality and training-cost tradeoff. Alternatively, the authors could explain a concrete feasibility barrier.

Questions for authors
  1. Could you compare ReCache against/with ECAD, HyCa, SpeCa and ClusCa? In particular, compute-matched, with results on human preference standins such as Image Reward, HPSv2.1, and existing LPIPS-style metrics. See citations / links at the end. Given the short rebuttal period, if comparing to all is not possible, a comparison on FLUX 1.dev to ECAD should be prioritized as it is most simlar to this method. I believe HyCa and ClusCa could potentially be utilized as the caching method as part of ReCache, whlie SpeCa likely would not be easy to stack. If the authors could clarifiy which of these can be combined with ReCache, that would strength ReCache's headline numbers, should these methods + ReCache yield more impressive caching.

  2. Using the same harness, objective and inference budget, could you compare RL against direct differentiation through a suitable relaxation of the schedule decisions? This would help isolate the value of RL from the value of the problem formulation. If such a comparison is infeasible, what specifically prevents it? If not feasible in the rebuttal timeframe, this would make a strong addition prior to camera ready. And if differntiation isn't possible, using something like ES (https://arxiv.org/abs/1703.03864). In particular, I would be interested to know if the same amount of offline compute leads to stronger results with a differentiable optimization of the same objective.

  3. How does ReCache compose with graph compilation? Does the caching logic introduce graph breaks or recompilation, and do the practical speedups hold against a compiled base model? It would be nice to have a comparison of the compiled base model (at the various step-level budgets, after warmup+compilation) vs ReCache + compilation if applicable.

  4. How does ReCache compare with few-step distilled models, and can it compose with them for additional acceleration?

  5. Could you provide the learned schedules or trained policies, prompts, seeds and commands needed to reproduce the Wan/Hunyuan video examples? Recognizing that code is optional, could you briefly clarify which code/configurations produced the reported results, given the startup/evaluation issues, missing video evaluation assets and differences from the paper's training recipe (e.g. 120 versus 50 epochs)? I could not reproduce the ReCache-accelerated video examples because the learned schedules/ policy checkpoints weren't included.

  6. Appendix C.2 describes a held-out validation subset, but the released code places all 64 FLUX validation prompts in the test set with different noise seeds. Was this overlap present in the reported experiments, and how was validation used for model/hyperparameter selection?

If the above weaknesses and questions are responded to, I am happy to raise my score.

SpeCa: https://arxiv.org/abs/2509.11628
ClusCa: https://arxiv.org/abs/2509.10312
Both above have the same code base here: https://github.com/Shenyi-Z/Cache4Diffusion
HyCa: https://arxiv.org/abs/2510.04188
ECAD: https://arxiv.org/abs/2506.15682

6 comments
AuthorsOpenReview ↗

Comparison with ECAD

We thank the reviewer for the valuable discussion and suggestions. Below we address the weaknesses and questions.

Comparison with ECAD

Thank you for highlighting the connection between ReCache and ECAD. Methodologically, while ECAD jointly searches where and when to cache and constructs a quality–compute Pareto frontier, ReCache isolates the question: given a caching mechanism and a prescribed budget, which steps should be fully recomputed? It enforces exactly kk full-computation steps and learns a single budget-conditioned policy. Holding the caching mechanism fixed enables controlled investigation of scheduling and its properties across different mechanisms. Jointly optimizing the caching mechanism and schedule is a valuable extension; our focus here is to establish the benefits and properties of budget-constrained scheduling itself.

We support this design choice with the comparisons below. At equal offline cost, ReCache achieves better quantitative results than ECAD's genetic search, both against original ECAD (Table 1) and against the ECAD's search restricted to our setting, where only the optimizer differs (Table 2).

Original ECAD. We run ECAD with its published search settings (candidates per generation, IR objective, 256$\times$256 resolution) on our 64 training prompts. At budget kk, we take the best found schedule with cost at most kk full steps.

Table 1. Original ECAD vs ReCache FORA on FLUX.1-dev; ECAD search is matched to ReCache FORA training in offline time ($\sim$19 GPU-h). All methods are evaluated at 1024$\times$1024.

kMethodLPIPS ↓L1 ↓HPS ↑IR ↑IR DB ↑AQ ↑IQ ↑
7uniform FORA (N=7)0.5670.1620.2700.8080.6286.2866.26
ECAD0.5940.1570.2790.9310.7716.4267.61
ReCache FORA0.4920.0800.2921.0250.9816.3870.99
8uniform FORA (N=6)0.5500.1580.2780.9040.7116.3167.96
ECAD0.5860.1580.2810.9400.7586.4268.36
ReCache FORA0.4390.0760.2971.0410.9916.4370.83
9uniform FORA (N=5)0.5300.1530.2850.9680.8266.3269.06
ECAD0.5850.1580.2820.9430.7596.4268.33
ReCache FORA0.3700.0730.2971.0370.9876.4070.45

ReCache FORA outperforms ECAD on all metrics at every budget except AQ on k=7,9k = 7, 9.

Isolating the search algorithm. Here we restrict ECAD to ReCache's setting: step-level schedules, the Taylor-DP cache and our PLPIPS −7⋅-7\cdotHPS objective, and run its search at 256$\times$256 and 1024$\times$1024 for ReCache's training time ($\sim$16 GPU-h). This comparison is intended to isolate the effect of RL without additional changes in the objective.

Table 2. ECAD vs ReCache, both with Taylor-DP and our objective, FLUX.1-dev. All methods are evaluated at 1024$\times$1024.

kkMethodSearch res.LPIPS ↓L1 ↓HPS ↑IR ↑IR DB ↑AQ ↑IQ ↑
7ECAD2560.5320.1540.2891.0120.8656.4669.17
ECAD10240.5370.1380.2991.0630.9246.4370.49
ReCache10240.4900.1250.2991.0921.0126.4770.31
8ECAD2560.5260.1500.2901.0100.8616.4768.97
ECAD10240.5010.1340.3051.1080.9846.4571.25
ReCache10240.4180.0910.3051.1261.0456.5070.59
9ECAD2560.5270.1500.2901.0150.8606.4768.98
ECAD10240.4460.1090.3061.0920.9896.4671.25
ReCache10240.4100.0910.3051.1331.0046.4970.85

With the optimizer as the only difference, ReCache outperforms ECAD on most metrics, which indicates that at equal training time REINFORCE generally finds more suitable schedules than the genetic search.

AuthorsOpenReview ↗

Comparison with SpeCa and ClusCa

SpeCa is an adaptive scheduler built on TaylorSeer, thus, is an alternative to ReCache's learned schedule rather than a component to stack, so we compare them directly with the same TaylorSeer caching mechanism (order 2, same cached layers). For each kk, we calibrate SpeCa's threshold on ReCache's 64 training prompts to compute kk steps on average.

Table 3. SpeCa vs ReCache, both with TaylorSeer (O=2O=2), FLUX.1-dev.

kkScheduleLPIPS ↓L1 ↓HPS ↑IR ↑IR DB ↑AQ ↑IQ ↑
7uniform0.5570.1610.2810.9330.7676.3868.48
SpeCa0.6150.1600.2850.9160.7756.2469.45
ReCache0.5500.1460.2901.0220.9766.3169.67
8uniform0.5420.1570.2881.0020.8316.4369.45
SpeCa0.5720.1470.2941.0080.8876.3570.88
ReCache0.5150.1390.2971.0580.9756.4469.89
9uniform0.5200.1510.2931.0460.9156.4470.46
SpeCa0.5410.1370.2981.0470.9516.4071.19
ReCache0.4300.0990.3041.1061.0086.4570.63
13uniform0.4690.1320.3001.0921.0016.4871.09
SpeCa0.3990.0970.3041.1031.0376.4771.34
ReCache0.1950.0400.3061.1261.0416.4771.34

ReCache matches the quality or outperforms SpeCa on all metrics at every budget except IQ at k=8,9k = 8, 9.

HyCa and CLusCa as caching mechanism for ReCache

HyCa and ClusCa are caching mechanisms, orthogonal to schedule selection, so they stack with ReCache. HyCa has neither public code nor the hyperparameters needed to reproduce it, so we implemented ClusCa within ReCache and compare three schedules: uniform, the one from the official code (three warm-up steps, then uniform), and ours.

Table 4. ClusCa on FLUX.1-dev; rows differ only in the schedule of kk computed steps.

kkScheduleLPIPS ↓L1 ↓HPS ↑IR ↑IR DB ↑AQ ↑IQ ↑
7uniform0.5580.1610.2810.9300.7676.3768.49
official code0.5900.1500.2890.9020.8086.2669.18
ReCache0.5240.1230.2941.0310.9266.3770.91
8uniform0.5430.1570.2871.0020.8266.4369.45
official code0.5290.1310.2991.0360.9786.4170.93
ReCache0.4720.1100.3011.0931.0166.4570.94
9uniform0.5200.1510.2931.0440.9086.4470.45
official code0.4960.1190.3021.0801.0116.4370.76
ReCache0.4230.0940.3031.0881.0246.4671.04

The table indicates that combining ClusCa with ReCache is beneficial for the method's quality compared to its original and uniform schedules.

AuthorsOpenReview ↗

REINFORCE vs Relaxation and ES; Compilation

Value of RL compared to relaxation and Evolution Strategies

Thanks for the valuable suggestion! Below we compare REINFORCE, differentiable relaxation, and Evolution Strategies (ES) on FLUX.1-dev with Taylor-DP (O=2O=2) using the same setup on a single H100.

For differential relaxation, we use the exponential-family entropy top-k relaxation from [1]. Gumbel-perturbed step scores produce soft recomputation gates whose sum matches the target budget. These gates interpolate between fully computed features and cached predictions, with differentiable cache updates. We backpropagate the image loss through the full denoising trajectory into the policy, keeping the backbone frozen. Since training requires full computations at every step as well as backpropagation, we train it for 18 epochs versus 50 for REINFORCE and ES to match methods' training times. As for ES, we use three Gaussian directions with both positive and negative perturbations, which matches ReCache's 6 samples per training example.

Table 5. Generation quality on FLUX.1-dev with Taylor-DP (O=2O=2).

BudgetMethodLPIPS ↓HPS ↑IR ↑IR DB ↑
7ES0.6340.2700.7060.578
Relaxation0.6310.2320.1340.057
ReCache0.4900.2991.0921.012
8ES0.7330.182-0.859-0.617
Relaxation0.6300.2310.1330.053
ReCache0.4180.3051.1261.045
9ES0.6300.2670.7510.696
Relaxation0.6370.2570.6130.440
ReCache0.4100.3051.1331.004
13ES0.5960.2800.8550.717
Relaxation0.6590.2730.7180.564
ReCache0.2440.3061.1231.034

ReCache achieves better fidelity and preference scores than both alternatives at every tested budget. The tested relaxation requires more GPU memory on training, as shown in Table 6.

Table 6. Training epochs and GPU memory on a single H100.

MethodEpochsGPU memory (GiB)
ES5039.50
Differentiable relaxation1868.35
ReCache (default)5039.53

Graph compilation

We compare kk-step full inference and ReCache with a budget of kk (k∈{7,9,13}k \in \lbrace 7, 9, 13 \rbrace), using the learned schedules for Taylor-DP (O=2O=2) with batch size 1 on one NVIDIA H100, after warm-up and compilation. Due to the fixed caching schedule for each budget, the caching logic causes no graph breaks or recompilations and allows to achieve the same speed-up.

Table 7. Wall-clock time per image (s) on FLUX.1-dev. kk is the number of computed steps: full inference with kk steps vs ReCache with a budget of kk out of 50 steps. In parentheses: speedup from torch.compile.

MethodCompilek=7k=9k=13
Full inference with kk steps×\times1.571.962.78
✓\checkmark1.20 (1.31$\times$)1.51 (1.30$\times$)2.12 (1.31$\times$)
ReCache×\times1.592.022.84
✓\checkmark1.21 (1.31$\times$)1.53 (1.32$\times$)2.15 (1.32$\times$)

References

[1] Gradient Estimation with Stochastic Softmax Tricks

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.