- 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
-
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.
-
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
-
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.
-
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
-
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?
-
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.
-
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
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 computed steps, optimizing fidelity and quality (PLPIPS HPS), 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 , ECAD uses its best found schedule with cost at most full steps.
Table 4. Original ECAD vs ReCache FORA, FLUX.1-dev. All methods are evaluated at 1024$\times$1024.
| Method | LPIPS ↓ | HPS ↑ | IR ↑ | IR DB ↑ | |
|---|---|---|---|---|---|
| 7 | uniform FORA (N=7) | 0.567 | 0.270 | 0.808 | 0.628 |
| ECAD | 0.594 | 0.279 | 0.931 | 0.771 | |
| ReCache FORA | 0.492 | 0.292 | 1.025 | 0.981 | |
| 8 | uniform FORA (N=6) | 0.550 | 0.278 | 0.904 | 0.711 |
| ECAD | 0.586 | 0.281 | 0.940 | 0.758 | |
| ReCache FORA | 0.439 | 0.297 | 1.041 | 0.991 | |
| 9 | uniform FORA (N=5) | 0.530 | 0.285 | 0.968 | 0.826 |
| ECAD | 0.585 | 0.282 | 0.943 | 0.759 | |
| ReCache FORA | 0.370 | 0.297 | 1.037 | 0.987 |
Table 5. ECAD vs ReCache, both with Taylor-DP and our objective, FLUX.1-dev. All methods are evaluated at 1024$\times$1024.
| Method | Search res. | PLPIPS HPS ↓ | LPIPS ↓ | HPS ↑ | IR ↑ | IR DB ↑ | |
|---|---|---|---|---|---|---|---|
| 7 | ECAD | 256 | -1.503 | 0.532 | 0.289 | 1.012 | 0.865 |
| ECAD | 1024 | -1.568 | 0.537 | 0.299 | 1.063 | 0.924 | |
| ReCache | 1024 | -1.618 | 0.490 | 0.299 | 1.092 | 1.012 | |
| 8 | ECAD | 256 | -1.515 | 0.526 | 0.290 | 1.010 | 0.861 |
| ECAD | 1024 | -1.644 | 0.501 | 0.305 | 1.108 | 0.984 | |
| ReCache | 1024 | -1.729 | 0.418 | 0.305 | 1.126 | 1.045 | |
| 9 | ECAD | 256 | -1.516 | 0.527 | 0.290 | 1.015 | 0.860 |
| ECAD | 1024 | -1.709 | 0.446 | 0.306 | 1.092 | 0.989 | |
| ReCache | 1024 | -1.741 | 0.410 | 0.305 | 1.133 | 1.004 |
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.
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.
| k | Schedule | LPIPS ↓ | HPS ↑ | IR ↑ | IR DB ↑ |
|---|---|---|---|---|---|
| 7 | uniform | 0.561 | 0.273 | 0.851 | 0.642 |
| Spectrum | 0.642 | 0.198 | -0.359 | -0.059 | |
| ReCache | 0.476 | 0.295 | 1.031 | 0.980 | |
| 8 | uniform | 0.543 | 0.280 | 0.925 | 0.723 |
| Spectrum | 0.492 | 0.269 | 0.852 | 0.828 | |
| ReCache | 0.408 | 0.298 | 1.073 | 0.977 | |
| 9 | uniform | 0.522 | 0.287 | 0.984 | 0.828 |
| Spectrum | 0.453 | 0.291 | 1.070 | 0.992 | |
| ReCache | 0.332 | 0.300 | 1.085 | 0.987 | |
| 13 | uniform | 0.472 | 0.297 | 1.062 | 0.953 |
| Spectrum | 0.239 | 0.303 | 1.091 | 1.015 | |
| ReCache | 0.235 | 0.303 | 1.099 | 1.001 |
Spectrum's schedule produces substantially worse results at . ReCache achieves best LPIPS, HPS and IR at every budget; Spectrum's schedule achieves higher IR DB at .
Benefits of budget-conditional training
Thank you for the valuable suggestion! Below we compare three parameterizations on FLUX.1-dev with Taylor-DP () 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 ().
| Budget | Parameterization | LPIPS ↓ | HPS ↑ | IR ↑ | IR DB ↑ |
|---|---|---|---|---|---|
| 7 | Direct logits | 0.538 | 0.294 | 1.063 | 0.935 |
| Fixed-input MLP | 0.544 | 0.289 | 1.008 | 0.947 | |
| Budget-conditioned MLP (default) | 0.490 | 0.299 | 1.092 | 1.012 | |
| 8 | Direct logits | 0.467 | 0.301 | 1.108 | 1.004 |
| Fixed-input MLP | 0.514 | 0.294 | 1.037 | 0.952 | |
| Budget-conditioned MLP (default) | 0.418 | 0.305 | 1.126 | 1.045 | |
| 9 | Direct logits | 0.410 | 0.303 | 1.112 | 1.029 |
| Fixed-input MLP | 0.413 | 0.301 | 1.096 | 0.977 | |
| Budget-conditioned MLP (default) | 0.410 | 0.305 | 1.133 | 1.004 | |
| 13 | Direct logits | 0.287 | 0.306 | 1.117 | 1.017 |
| Fixed-input MLP | 0.222 | 0.306 | 1.111 | 1.025 | |
| Budget-conditioned MLP (default) | 0.244 | 0.306 | 1.123 | 1.034 |
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.
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⋅HPS 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=2).
Our final-image objective (PLPIPS −7⋅HPS) 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⋅HPS with exactly k computed steps.
Effectiveness under comparable settings. We compare on FLUX.1-dev with Taylor-DP (O=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=2).
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=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.