acceptodds
Under review as a conference paper at ICLR 2027

DARTCache: Dual-Stage Temporal Reuse Planning for Efficient Inversion-Based Editing

Abstract

Inversion-based image editing requires repeated model evaluations during both inversion and denoising, making inference computationally expensive. Training-free caching can reduce this cost, but applying caching policies independently to the two stages leaves a shared-budget planning problem unresolved: where to refresh predictions within each stage and how to allocate computation between stages. We propose **DARTCache**, a training-free framework that jointly plans these two decisions using a single trajectory-based surrogate objective. We introduce Temporal Variation Exposure (TVE), which quantifies cumulative prediction variation between refresh locations. A quadratic TVE objective promotes exposure-balanced refresh placement and guides stage-wise budget allocation. **DARTCache** estimates temporal variation profiles from full-computation trajectories on a small calibration set and constructs fixed refresh schedules without candidate-level replay or online schedule search. On RF-Inversion and OT-RF, **DARTCache** achieves end-to-end speedups of 2.80 and 2.67, with full-image LPIPS of 0.107 and 0.137 relative to the full-inference outputs, yielding favorable efficiency–fidelity trade-offs against evaluated acceleration baselines. Further experiments on U-Net-based MasaCtrl and EDICT support the applicability of the planning framework beyond FLUX-based editing pipelines.

Primary area:
generative models
Keywords:
Inversion-Based Image Editing, Training-Free Acceleration, Prediction Caching, Temporal Scheduling, Computation Budget Allocation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7zpqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Inversion-based image editing consists of inversion and denoising stages, both of which require many network evaluations. Existing caching methods typically handle these stages separately, without jointly optimizing their shared computational budget, potentially leading to inefficient resource allocation. This paper proposes DARTCache, a training-free framework for jointly planning caching across both stages. It extends conventional single-stage caching optimization to joint computational resource planning, using a unified metric to determine both when to refresh cached features and how many network evaluations to allocate to each stage. Among the acceleration methods evaluated, the results demonstrate a favorable trade-off between efficiency and fidelity. The paper further validates its applicability to U-Net-based editing frameworks.

Critical strengths

The method requires neither image-quality evaluation through replay for each candidate nor an online predictor. The calibration curves computed on the given test dataset provide a good balance between quality and speed.

Critical weaknesses

Generalizability remains insufficiently validated. Although the paper reports that the calibration curves remain relatively stable as coverage varies across the 700 test cases, this does not establish what regularities govern these curves. Their behavior across different models and datasets has not been verified.

Questions for authors

Could you provide more evidence supporting generalizability?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7zpq: Overview of Generalizability Evaluation

We thank the reviewer for highlighting the need to establish the generalizability of DARTCache beyond calibration-coverage stability.

We agree that demonstrating stable calibration curves within a single dataset is insufficient to establish a generalizable planning principle. To address this concern systematically, we conduct additional experiments across different datasets, editing pipelines, and model architectures, organized around four complementary questions:

  1. Temporal Regularity: Do temporal variation profiles exhibit common properties across settings, and which characteristics are dataset- or host-dependent?

  2. Objective Generalization: Even when temporal profiles differ, does Quadratic TVE remain predictive of actual cache-induced fidelity degradation?

  3. Planning Effectiveness: Does TVE-based refresh placement and stage allocation consistently translate into practical fidelity improvements across editing frameworks?

  4. Schedule Transfer: Can a schedule calibrated on one dataset remain effective on another without recalibration?

These evaluations progress from characterizing temporal dynamics to validating the scheduling objective, its practical effectiveness, and its transferability.

Our central finding is that generalizability does not require identical temporal profiles across models. Instead, temporal variation exhibits a common nonuniform structure, while its detailed distribution is host-dependent. Quadratic TVE remains informative for scheduling across the evaluated settings, supporting a host-adaptive planning principle rather than a universal fixed schedule.

We present the detailed evidence in two parts:

  • Part 1: Temporal regularities and cross-setting TVE objective predictiveness.
  • Part 2: Planning effectiveness across editing frameworks and cross-dataset schedule transfer.

The corresponding experiments have also been incorporated into Appendix D.5 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 7zpq(Part 1/2): Temporal Regularities and TVE Predictiveness

We thank the reviewer for highlighting the need to characterize temporal variation regularities beyond calibration-coverage stability. We conduct new experiments on MagicBrush, additional profile and objective-ranking analyses on SD1.5-based U-Net editors, and revisit the U-Net component ablations reported in our original submission. Together, these experiments examine temporal variation structure, TVE predictiveness, and practical planning effectiveness across datasets and editing frameworks.

1. Temporal Variation Regularities Across Datasets and Editing Frameworks

Cross-dataset consistency. We independently extract temporal variation profiles from PIE-Bench (70 calibration cases) and MagicBrush (50 cases) using FLUX.1-dev with RF-Inversion. The stage-normalized profiles exhibit high Spearman correlations of 0.9515 (inversion) and 0.9719 (denoising). Their denoising-to-inversion variation ratios are also similar (3.98 vs. 3.84), with both independently selecting a 6/14 allocation at B=20B=20.

Cross-framework characteristics. We further analyze archived 10-case calibration profiles from four editing configurations: FLUX.1-dev with RF-Inversion/OT-RF and SD1.5 U-Net with MasaCtrl/EDICT. Even within the same backbone, different editing pipelines exhibit substantially different temporal profiles. For example, the inversion-profile correlation between MasaCtrl and EDICT is −0.5738, while the denoising-profile correlation between RF-Inversion and OT-RF is −0.0353. Nevertheless, across all four configurations, the highest-variation 20% of intervals account for approximately 36%–58% of total variation.

These descriptive results reveal a common empirical property: temporal variation is strongly nonuniform, while its detailed structure is host-dependent. DARTCache therefore does not assume a universal temporal profile, but instead adapts its planning decisions through host-specific calibration.

2. Cross-Dataset and Cross-Architecture TVE Objective Validation

To investigate whether TVE remains predictive of actual caching degradation, we extend our original objective-ranking evaluation to 200 candidate schedules on 50 held-out MagicBrush edits and 30 schedules on 30 SD1.5/EDICT cases. Both experiments use independently calibrated profiles and fixed stage allocations (10/10 and 15/17, respectively), isolating refresh-placement ranking.

Table R2-1: Spearman correlation between offline objectives and empirical LPIPS degradation.

ObjectivePIE-BenchMagicBrushSD1.5 / EDICT
Endpoint0.64330.53550.6609
Integrated0.76950.72660.7070
Lexi-Minimax0.93160.64280.4843
Quadratic TVE0.96370.86240.8061

Quadratic TVE achieves the highest observed correlation among the four objectives in all three settings. It also yields the lowest Top-1 LPIPS regret on MagicBrush (0.0094) and EDICT (0.00307).

On EDICT, Quadratic TVE achieves a correlation of 0.8061 (95% CI [0.7374, 0.8661]), compared with 0.7070 [0.6209, 0.7722] for Integrated. We report this as an observed advantage rather than a statistically established superiority.

Since candidate pools and evaluation sizes differ across settings, these correlations demonstrate within-setting predictive usefulness, rather than a controlled comparison of architecture-induced differences. Importantly, the original Quadratic TVE objective retains substantial schedule-ranking ability despite different datasets and temporal profile structures.

AuthorsOpenReview ↗

Response to Reviewer 7zpq (Part 2/2): Planning Effectiveness and Schedule Transfer

3. Generalization of TVE Allocation and Refresh Placement

We further examine whether these scheduling signals translate into practical fidelity improvements across editing frameworks.

Cross-dataset ablation. On MagicBrush, we evaluate 450 session-disjoint held-out edits using FLUX.1-dev with RF-Inversion and OT-RF, independently calibrated on 50 cases. Both pipelines use the original TVE formulation and the same B=20B=20 online budget.

Table R2-2: Component ablation on MagicBrush (LPIPSAll ↓).

Allocation + PlacementRF-InversionOT-RF
Equal + Uniform0.320240.35093
TVE + Uniform0.185310.28883
Equal + TVE0.130610.20776
TVE + TVE (DARTCache)0.088880.19919

DARTCache achieves the lowest mean LPIPSAll on both hosts, with measured speedups of 2.56× and 2.72×, respectively (measured on a fixed 10-case timing subset). Compared with EEdit SLoC+ISS under its native configuration, DARTCache also achieves lower LPIPSAll (0.08888 vs. 0.11421 on RF-Inversion; 0.19919 vs. 0.35138 on OT-RF) and higher speedups (2.56× vs. 2.28×; 2.72× vs. 2.35×). This external comparison does not assume matched fresh-NFE budgets.

Cross-architecture ablation. We also revisit the complete 2×2 component ablations on the SD1.5-based MasaCtrl and EDICT frameworks from our original Appendix D.

Table R2-3: TVE component ablation on U-Net editors.

Allocation + PlacementMasaCtrl I/DLPIPS ↓PSNR ↑EDICT I/DLPIPS ↓PSNR ↑
Equal + Uniform30/300.486611.0816/160.232623.09
Equal + TVE30/300.463911.9916/160.169424.43
TVE + Uniform40/200.134619.8715/170.227323.29
TVE + TVE40/200.036528.8415/170.151725.12

Under identical TVE-selected stage allocations, TVE-guided placement reduces LPIPS by 72.9% on MasaCtrl and 33.3% on EDICT relative to uniform placement. The complete configurations achieve reported speedups of approximately 1.72× and 2.48×, respectively.

Host-dependent component contributions. The relative benefit of allocation and placement varies substantially across editing frameworks. With TVE placement fixed, introducing TVE-guided allocation reduces LPIPS by 92.1% on MasaCtrl, compared with 10.4% on EDICT. Similar differences appear on MagicBrush: 31.9% on RF-Inversion versus 4.1% on OT-RF. These are conditional mean improvements; notably, the additional masked-LPIPS benefit of allocation on OT-RF is not statistically established.

Thus, TVE-guided placement consistently improves mean fidelity across the four evaluated hosts, while the incremental benefit of stage allocation is host-dependent. These findings support jointly adapting placement and allocation to host-specific temporal dynamics rather than relying on fixed scheduling rules.

4. Frozen-Schedule Transfer Without Target-Dataset Recalibration

Finally, we directly transfer the PIE-Bench-calibrated schedule to all 450 held-out MagicBrush edits, without target-dataset calibration. Both schedules use the same 6/14 allocation and differ at four refresh positions, primarily testing cross-dataset refresh-placement transfer.

The frozen schedule achieves LPIPSAll = 0.08947, compared with 0.08889 after MagicBrush recalibration and 0.32024 for uniform scheduling. The paired LPIPSAll difference is +0.00058 (95% CI [−0.00077, +0.00194]).

However, frozen transfer incurs a small but detectable masked-PSNR degradation (−0.1645 dB, 95% CI [−0.2847, −0.0421]), and some individual edits exhibit larger errors. Thus, the results support practical placement transfer under the evaluated configuration, but do not establish statistical equivalence or universally lossless transfer.

Conclusion. Our additional experiments identify an empirical regularity underlying DARTCache: temporal variation is consistently nonuniform but structurally host-dependent. Quadratic TVE retains substantial predictive ability across the evaluated datasets and backbone families, while the component ablations demonstrate practical scheduling benefits with host-dependent contributions. These results support the applicability of the original TVE planning principle across the evaluated settings, without assuming universal temporal profiles or performance guarantees.

Reviewer_uN8GOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes DARTCache, a training-free offline planner for prediction reuse in inversion-based image editing. Temporal Variation Exposure (TVE), estimated from full-computation calibration trajectories, guides both refresh placement and computation allocation between inversion and denoising. Experiments report 2.80× and 2.67× end-to-end speedups on RF-Inversion and OT-RF, respectively, with favorable fidelity to full inference. Additional analyses evaluate the surrogate objective, component contributions, calibration robustness, and extensions to U-Net editors.

Critical strengths
  1. Substantial validation of the planning criterion. Schedule-ranking experiments, controlled profile ablations, and comparisons with exact surrogate optimization support the usefulness of TVE. Component ablations separately establish benefits from refresh placement and stage allocation.

  2. Low-cost offline planning with negligible online scheduling overhead. The method avoids candidate-level replay and reports fidelity comparable to the evaluated LeMiCa-style planner at substantially lower offline cost. Calibration robustness tests and quality–speed curves further support its practical utility within the evaluated settings.

Critical weaknesses
  1. The incremental value over simpler allocation strategies remains unclear. EEdit already exploits unequal computation requirements between inversion and denoising, while offline budget-constrained cache planning is established by LeMiCa. Although the ablations support TVE, a stronger comparison against simple refresh intervals and stage-budget ratios tuned on the same calibration set would better establish the value of the proposed planner.

  2. The outcome-aware comparison supports a narrower conclusion than a general planning advantage. The LeMiCa-style adaptation measures inversion-segment costs at the inversion terminal latent, without propagating them through subsequent denoising. Its comparison also fixes the calibration sample count rather than matching calibration cost or achieved quality, leaving the broader quality–planning-cost advantage insufficiently characterized.

Questions for authors
  1. How does DARTCache compare with jointly tuned fixed refresh intervals and stage-budget ratios using the same calibration set?

  2. Does the LeMiCa-style comparison change when inversion-segment costs account for subsequent denoising and final-output deviation?

  3. How do the two planners compare when calibration cost is matched or the calibration sample count is varied?

Reviewer_S5ojOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DARTCache, a training-free cache schedule searching framework for dual-stage inversion-based image editing. DARTCache mainly uses Temporal Variation Exposure for schedule calibration: Typically a 70-sample calibration set is leveraged for (27 inversion + 28 denoing) 55 * 70 step function call, and then Temporal Variation Exposure is calculated for adjacent steps. Then DARTCache generate a budget assignment given a fixed total budget (e.g. inversion + denoising = 20 step calls) using the TVE ratio (e.g. 6 steps for inversion & 14 steps for denoising). Finally, DARTCache outputs caching step schedules which minimizing the total TVE difference gap for each intervals. The method is efficient and achieves competitive results against previous works.

Critical strengths
  1. The paper is well-written and easy to follow. The idea of evenly separated TVE for each trajectory is reasonable and well-motivated.
  2. The proposed method, DARTCache, is a training-free and computationally efficient caching searching strategy, which requires only a small set of calibration set.
  3. Extensive experiments demonstrate that the DARTCache provide favorable efficiency-fidelity tradeoff on various editing pipelines.
Critical weaknesses

1. Limited inversion-specific novelty

Although motivated by inversion-based image editing tasks, DARTCache still treats inversion and denoising stages as two independent trajectories (unless for the budget assignment stage). No inter-trajectory information is leveraged for the inversion-denoising procedure. Also, the inequality in inversion-denoising budget assignment is already proposed in early work like EEdit, which also limits the novelty of the paper.

2. Sensitivity to temporal discretization.

DARTCache estimates TVE from trajectories that comes from a fixed discrete timestep grid. There is no further analysis on finer timestep grid (e.g. 2x or 4x more steps). Will this improves the performance at the expense of longer calibration time?

3. Comparison with timestep scheduling

Works like (Timestep Rescheduling in Diffusion Inversion) and (Entropic Time Schedulers for Generative Diffusion Models) are dedicated timestep scheduling methods for inversion and denoising step. Since DARTCache directly caches the model outputs, its behavior is similar to a timestep schedule bounded to a discrete grid. Does DARTCache perform better than these schedulers (under same budget assignment)?

Questions for authors

I'm still confused since works like DARTCache directly caches model outputs, its main algorithmic contribution lies in selecting the denoising steps at which model evaluations are performed, rather than introducing a new caching operator. Which is more similar to a timestep scheduling method. I think this connection should be discussed more explicitly. Does DARTCache works with attention-based or intermediate states caching strategies for further acceleration?

4 comments
AuthorsOpenReview ↗

Response to Reviewer S5oj (Part 1)

Response to Reviewer S5oj (Part 1)

Thank you for the constructive review. We appreciate the reviewer's recognition of DARTCache's efficiency and motivation. We clarify its methodological contributions and provide additional evidence addressing temporal discretization sensitivity.

1. Generality and Inversion-Specific Novelty (W1)

We agree that DARTCache does not explicitly model inversion-to-denoising trajectory propagation, and that asymmetric stage budgets have previously been explored by EEdit. Our contribution is not stage asymmetry itself, but a unified TVE-based planning principle that jointly determines stage allocation and refresh placement without candidate-level replay.

Why avoid explicit cross-stage trajectory modeling? Different inversion-based editing methods employ distinct solvers, state representations, and editing controls. Explicitly modeling inversion-to-denoising error propagation would require knowledge of these host-specific dynamics, potentially limiting applicability across different editing methods.

Instead, DARTCache provides a unified, host-adaptive planning criterion. It extracts temporal prediction variation directly from each host's full-computation trajectories and applies the same TVE objective without requiring host-specific transition models, cross-stage state mappings, or error-propagation formulations. Different hosts may exhibit different temporal variation profiles while sharing the same planning principle.

The unified quadratic TVE objective preserves the native relative variation magnitudes across stages, guiding both exposure-balanced refresh placement and shared-budget allocation. Its continuous relaxation yields bs−1∝Vsb_s-1\propto V_s, where VsV_s denotes the total temporal variation of stage ss. Thus, stage asymmetry emerges from calibrated prediction variation rather than manually specified budget ratios.

Why full-model output reuse? We deliberately adopt a coarse-grained caching operator that skips the entire Transformer forward at reuse positions, offering high potential computation savings per reuse decision. Our methodological focus is on scheduling these evaluations effectively rather than introducing architecture-specific internal caching mechanisms.

Experimental evidence. At B=20B=20, DARTCache outperforms jointly tuned uniform scheduling (LPIPSAll: 0.1071 vs. 0.1673). Compared with our LeMiCa-E2E adaptation, which explicitly evaluates downstream error propagation, DARTCache requires only 3,850 vs. 966,770 offline Transformer evaluations, while achieving competitive fidelity (LPIPSAll: 0.1071 vs. 0.1036).

Together with evaluations on RF-Inversion, OT-RF, MasaCtrl, and EDICT, these results support a unified, host-adaptive, calibration-efficient planning framework that does not require explicit host-specific cross-stage propagation models.

2. Offline Calibration-Grid Sensitivity (W2)

We evaluate the sensitivity of TVE calibration to temporal discretization on FLUX.1-dev + RF-Inversion using 70 calibration and 630 disjoint held-out PIE-Bench cases.

Experimental setup. We vary the offline calibration resolution (1×, 2×, 4×), while keeping the online grid (27 inversion + 28 denoising positions), B=20B=20, cache operator, model, seeds, and editing parameters fixed. Full-computation fine-grid trajectories are collected, and fine-grid prediction variation is accumulated within the original coarse intervals before applying the same TVE planner.

Table R3-1: Offline calibration-grid sensitivity (B=20).

GridI/DLPIPSAll ↓LPIPSM ↓PSNRAll ↑Cal. Cost (GPU-h) ↓
1×6/140.10710.077726.89820.1345
2×6/140.10710.077726.89820.2310
4×6/140.10680.078126.83450.3880

Results. All three resolutions select the same 6/14 allocation. The 2× calibration produces the same refresh schedule as 1×. The 4× calibration changes one inversion and five denoising refresh positions, with only marginal differences in LPIPSAll (−0.0003), LPIPSM (+0.0004), and PSNRAll (−0.064 dB).

Meanwhile, offline calibration cost increases by 1.72× at 2× and 2.88× at 4×.

Conclusion. Increasing offline calibration resolution does not yield consistent fidelity improvements under the evaluated fixed online-grid protocol. The original calibration resolution provides a favorable fidelity–calibration-cost trade-off. This experiment addresses offline calibration resolution rather than arbitrary changes to online solver discretization.

AuthorsOpenReview ↗

Response to Reviewer S5oj (Part 2)

Response to Reviewer S5oj (Part 2)

3. Comparison with Timestep Scheduling (W3)

We agree that DARTCache and timestep scheduling both select model evaluation positions. However, prediction reuse and naive sparse timestep integration are not equivalent under the evaluated RF-Inversion editing updates.

Experimental setup. We evaluate four configurations on 630 held-out PIE-Bench cases using FLUX.1-dev + RF-Inversion, varying scheduling (Uniform vs. TVE) and execution (Cache vs. Sparse Euler). All configurations use B=20B=20 and the same 6/14 stage allocation.

Within each scheduling strategy, Cache and Skip use identical evaluation positions and are evaluated against the same Full Inference outputs. This controls the refresh-position selection when examining execution differences.

Execution mechanism. For simple Euler integration with constant velocity over an interval, repeated small steps and a single large step are mathematically equivalent. However, RF-Inversion contains additional state- and time-dependent editing operations:

  • Inversion: Cache reuses the raw Transformer prediction utu_t, but recomputes the conditional control (y1−Yt)/(1−t)(y_1-Y_t)/(1-t) and corrected velocity at every original timestep, with γ=0.7\gamma=0.7.
  • Denoising: Cache reuses the raw Transformer prediction while recomputing (y0−xt)/(1−t)(y_0-x_t)/(1-t) and the corrected velocity at every original timestep. The control uses η=0.7\eta=0.7 over steps 0–5, and background/mask reconstruction is performed at each step i<18i<18.
  • Sparse integration: The control-corrected velocity is evaluated only at selected anchors and integrated directly to the next anchor, without intermediate control updates or mask reconstruction.

Both implementations determine control activation using the original timestep indices, without renumbering skipped positions. Neither uses per-step stochastic updates.

Thus, their execution difference arises from the treatment of state-dependent RF editing operations, rather than finer Euler integration alone.

Table R3-2: Prediction reuse versus sparse timestep integration (630 held-out edits, B=20).

MethodLPIPSAll ↓LPIPSM ↓LPIPSBG ↓PSNRAll ↑PSNRM ↑PSNRBG ↑Latency (s) ↓
Uniform-Cache0.16730.14270.076523.850919.899627.722.035
Uniform-Skip0.61020.19490.459114.6217.5011.581.977
TVE-Cache0.10710.07770.038726.898224.057631.942.018
TVE-Skip0.57410.16310.451315.4818.7412.041.979

Additional absolute quality evaluation.

MethodImageReward ↑CLIP-T ↑CLIP-I ↑
Uniform-Cache0.399225.601485.8012
Uniform-Skip−1.096721.335368.6574
TVE-Cache0.450025.8685.55
TVE-Skip−0.986421.451469.2064

Results. With the same 20-evaluation budget, TVE-Cache achieves substantially higher reported fidelity than naive sparse integration (LPIPSAll: 0.1071 vs. 0.5741; PSNRAll: 26.90 vs. 15.48 dB), with approximately 1.97% higher latency.

TVE also improves evaluation-position selection under both execution modes. Relative to Uniform, the reported LPIPSAll is 36.0% lower under Cache and 5.9% lower under Skip. The different effect sizes suggest that scheduling effectiveness depends on the execution mechanism.

Stage attribution. On 630 cases, replacing only inversion caching with sparse integration yields a final latent MAE of 0.0429 relative to Cache–Cache, whereas replacing only denoising caching yields 0.4208. The observed divergence is substantially larger when sparse execution is applied to denoising than to inversion, consistent with the importance of retaining per-step denoising controls and mask reconstruction.

Regarding the cited schedulers. TRDI was developed for DDIM-based inversion, whereas Entropic Time Schedulers investigate conditional-entropy-based time reparameterization, with published experiments primarily on EDM generation. Faithfully adapting their original scheduling criteria to RF-Inversion editing requires additional derivation and validation.

We therefore do not claim superiority over their original published implementations. Our comparison instead clarifies the relationship between evaluation-position scheduling and prediction-reuse execution under a fixed computational budget.

Conclusion. DARTCache shares the position-selection problem with timestep scheduling, but its full-model prediction reuse retains the original RF editing updates between refreshed predictions. This distinction is important in editing pipelines with state-dependent controls and mask reconstruction.

AuthorsOpenReview ↗

Response to Reviewer S5oj (Part 3)

Response to Reviewer S5oj (Part 3)

4. Compatibility with Intermediate-State Caching (Q1)

We thank the reviewer for this suggestion. DARTCache is compatible with FirstBlockCache-style intermediate-state caching, since the two mechanisms operate at different computational granularities.

DARTCache performs temporal-level full-model prediction reuse, whereas intermediate-state caching reduces computation within the remaining Transformer evaluations.

Experimental setup. We implement a FirstBlockCache-style block-residual reuse mechanism within the FLUX Transformer and evaluate FBCache-only, DARTCache, and their Hybrid on 630 held-out PIE-Bench cases.

The Hybrid preserves DARTCache's 20 outer refresh positions and applies internal block-residual reuse with a residual-difference threshold of 0.08. FBCache states are reset between inversion and denoising and across editing cases.

Table R3-3: Hierarchical caching compatibility (630 held-out edits).

MethodOuter CallsFBCache HitLatency (s) ↓Speedup ↑LPIPSAll ↓PSNRAll ↑
FBCache-only5534.76%3.1831.514×0.085428.04
DARTCache20—1.8322.630×0.107126.8982
Hybrid209.94%1.7002.835×0.125225.9769

Full no-cache latency is 4.819 s/case. Outer calls count Transformer invocations; FBCache hits execute only part of the network. FBCache hit rates are calculated over eligible internal caching decisions.

Results. The Hybrid successfully combines both caching mechanisms, achieving 1.078× additional acceleration over DARTCache (latency: 1.832 → 1.700 s). However, additional intermediate-state reuse introduces a fidelity trade-off: LPIPSAll increases from 0.1071 to 0.1252, while PSNRAll decreases from 26.90 to 25.98 dB.

Why are the additional gains limited? FBCache's hit rate decreases from 34.76% in standalone execution to 9.94% in the Hybrid, with 0% inversion and 14.21% denoising hit rates.

A plausible explanation is that DARTCache's sparse refresh schedule increases temporal and state differences between successive evaluated positions, reducing intermediate-feature similarity. This interpretation is consistent with the observed hit rates but is not claimed as the uniquely established cause.

Conclusion. DARTCache can be combined with the evaluated intermediate-state caching mechanism, but their acceleration benefits are not automatically additive. The Hybrid provides modest additional acceleration with a non-negligible fidelity cost.

These findings support our choice to focus on full-model output reuse with TVE-guided temporal planning, while leaving joint cost-aware optimization with finer-grained caching as a promising future direction.

Overall Conclusion

DARTCache contributes a unified, host-adaptive, calibration-efficient dual-stage reuse planning framework.

By using temporal prediction variation as a common planning signal, it jointly determines computation allocation and refresh placement across heterogeneous inversion-based editing hosts without requiring explicit host-specific cross-stage propagation models.

The additional experiments characterize its calibration-grid sensitivity, clarify its execution semantics relative to sparse timestep integration, and demonstrate the feasibility and limitations of hierarchical caching.

Thank you for your replies, which clear my doubts. I will keep my rating of weak acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.