acceptodds
Under review as a conference paper at ICLR 2027

HASTE: Hierarchical Adaptive Spatio-Temporal Reuse for Efficient Image Decomposition

Abstract

Text-to-layer (T2L) diffusion models achieve editable image decomposition by jointly denoising multi-layer RGBA stacks, but suffer from prohibitive inference latency. Existing diffusion accelerators overlook this multi-layer structure, treating images as flat, spatially contiguous domains. We observe that computational redundancy in T2L models is inherently hierarchical, governed by inter-layer temporal heterogeneity (layers stabilize at disparate denoising stages) and intra-layer spatial sparsity (active synthesis is strictly confined to a sparse alpha support, outside of which latent states quickly freeze). To exploit this insight, we propose HASTE, the first dedicated online caching framework for T2L generation. HASTE infers alpha support online via an ultra-lightweight token-level regressor and deploys a two-stage hierarchical scheduler: converged layers bypass Transformer execution entirely via flow-matching endpoint extrapolation, while active layers restrict fresh evaluation strictly to tokens within the alpha support. To prevent cross-layer context fragmentation, HASTE executes a sparse DiT scheme powered by persistent Key-Value caches. On the Crello benchmark, HASTE accelerates Qwen-Image-Layered sampling by with negligible quality loss. Furthermore, it generalizes training-free across four diverse T2L backbones, delivering consistent – speedups. Code is available at https://anonymous.4open.science/r/HASTE-0434.

Primary area:
generative models
Keywords:
Text-to-Layer Generation, Diffusion Acceleration, Hierarchical Reuse, Sparse DiT Execution

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_a9v1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes HASTE, an inference acceleration framework for diffusion-based image decomposition into editable RGBA layers. The approach exploits differences in convergence across output layers and spatial sparsity within each layer’s alpha support. A lightweight AlphaNet estimates opacity from cached clean latent predictions. An online scheduler uses support-restricted feature changes and layer-specific sensitivity calibration to determine which layers require fresh computation. Within selected layers, computation is restricted to the estimated alpha support. Skipped tokens bypass the Transformer and advance using cached clean endpoints, while persistent KV caches allow freshly evaluated tokens to access context from skipped positions. Warmup, periodic full refreshes, and final refinement mitigate accumulated errors

Critical strengths

Alpha support provides a task-specific cue for allocating computation. Restricting scheduling statistics to this support is particularly sensible for small foreground layers, where full-canvas averaging could dilute meaningful changes.

The diffusion backbone remains frozen, and the auxiliary opacity predictor is small. The additional pipeline experiments provide useful evidence of portability within layered image generation.

Critical weaknesses

The distinctive contribution of the hierarchical policy is not sufficiently isolated. Table 3 reports 3.07× speedup for temporal-only scheduling and 3.19× for complete HASTE. Adding spatial selection reduces latency from 31.74 s to 30.49 s, an incremental reduction of approximately 3.9%. However, Appendix D.5 states that temporal-only timing includes execution-audit overhead, whereas complete HASTE uses normal sampling times.

Can you compare global scheduling, per-layer scheduling without alpha-restricted scoring, and the proposed support-aware scheduling using the same endpoint/KV implementation and matched computation budgets? Please also measure temporal-only and complete HASTE with identical timing instrumentation.

How sensitive is the final decomposition to missed alpha-support regions? An oracle-support control would help separate support-estimation error from endpoint and KV staleness. Please also specify the full-update guard conditions and report their trigger frequencies and runtime costs.

Questions for authors

Please see weakness part.

4 comments
AuthorsOpenReview ↗

Response (Part I)

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript.

Weakness 1&Weakness 2: The distinctive contribution of the hierarchical policy is not sufficiently isolated. Table 3 reports 3.07× speedup for temporal-only scheduling and 3.19× for complete HASTE. Adding spatial selection reduces latency from 31.74 s to 30.49 s, an incremental reduction of approximately 3.9%. However, Appendix D.5 states that temporal-only timing includes execution-audit overhead, whereas complete HASTE uses normal sampling times. Please also measure temporal-only and complete HASTE with identical timing instrumentation.

Response: We thank the reviewer for the careful observation. We clarify two points.

1. Timing instrumentation is identical in Table 3.
Temporal-only and complete HASTE in Table 3 were measured with the same execution-audit instrumentation, and its overhead is included in both latencies. Appendix D.5 incorrectly stated that complete HASTE used normal sampling times. This was a documentation error, not a measurement inconsistency, and it has been corrected. The 3.07× vs. 3.19× comparison is therefore based on identical timing instrumentation.

2. Table 3 is not the right setting to isolate the spatial contribution.
Table 3 uses the long-trajectory T=50T=50 setting, where temporal redundancy dominates. Layer-level/temporal skipping already removes most redundancy, leaving little spatial redundancy. The modest +3.94%+3.94\% gain is thus expected and does not undermine the two-layer design.

The independent value of Alpha-Guided Spatial Selection is clear when spatial redundancy is high. Under matched quality, adding spatial selection yields an additional 9.32%–16.35% latency reduction over layer-level scheduling. On Ming, spatial selection alone (1.41×) even outperforms layer-level scheduling alone (1.27×). Temporal and spatial sparsity target complementary redundancies; neither dominates universally.

Backbone & ScheduleFull (s)Layer-level scheduling only (s)Spatial-selection only (s)HASTE (Layer-level + Spatial) (s)Incremental Spatial Gain ↑
Qwen (T=50T=50, Original)97.3131.74 (3.07×)62.95 (1.55×)30.49 (3.19×)+3.94%
Qwen (T=15T=15, Few-step)46.1830.89 (1.50×)33.13 (1.40×)27.78 (1.66×)+10.07%
RevealLayer141.5292.32 (1.53×)127.51 (1.11×)83.72 (1.69×)+9.32%
Ming140.88111.31 (1.27×)100.05 (1.41×)93.11 (1.51×)+16.35%

These clarifications and ablations have been incorporated into the revision.

Weakness 2: Can you compare global scheduling, per-layer scheduling without alpha-restricted scoring, and the proposed support-aware scheduling using the same endpoint/KV implementation and matched computation budgets?

Response: We thank the reviewer for this valuable suggestion. Under nearly identical FLOPs(~ 4.41P) and query budgets (~ 29.44%), our support-aware scheduling consistently outperforms global and per-layer scheduling in both efficiency and decomposition quality. All three strategies share the same endpoint-reuse and KV-cache implementation, demonstrating the contribution of our proposed support-aware scheduling policy.

Scheduling strategyFLOPs(P)Query budgetLatency (s) ↓Speedup ↑LPIPS ↓GT αIoU ↑GT RGB L1 ↓
Proposed support-aware scheduling(HASTE)4.40629.44%30.723.22×0.015270.884790.06990
Global scheduling4.40629.44%37.452.64×0.017260.877660.07175
Per-layer scheduling4.40929.45%34.582.86×0.017390.873660.07015
AuthorsOpenReview ↗

Response (Part II)

Weakness 3: How sensitive is the final decomposition to missed alpha-support regions? An oracle-support control would help separate support-estimation error from endpoint and KV staleness. Please also specify the full-update guard conditions and report their trigger frequencies and runtime costs.

Response: Thanks for the catch! Our controlled experiments show that HASTE is reasonably robust to missed alpha-support regions, while support-estimation achieves near-oracle-support fidelity at lower computational cost.

(1) Robustness to Missed Alpha-Support Regions.

Even with 20% additional missed support, HASTE's decomposition quality remains stable. Moreover, alpha-support achieves comparable fidelity to oracle-support while reducing latency by 23.3%. These findings suggest that moderate support-estimation errors have limited impact under HASTE's recovery mechanisms.

Support SourceLatency (s) ↓GT αIoU ↑GT RGB L1 ↓LPIPS ↓
Alpha-support (HASTE)30.720.887380.050710.01341
Oracle-support37.230.886530.051210.01297
Oracle + 10% missed37.380.887300.050860.01334
Oracle + 20% missed37.110.887030.050740.01346

(2) Full-update guards frequencies and runtime costs are little.

Across 500 images (25,000 sampling steps), full-update guard triggered conditions occur only 0.68 times per image, accounting for 5.20% of total sampling time. In contrast, selective/cached updates cover 77.3% of sampling steps while consuming only 20.96% of runtime. These results demonstrate that HASTE achieves substantial computational savings with modest full-update guard overhead.

ComponentTotal stepsMean steps /imageMean time/imageRuntime share
Warmup3000 steps6.0012.786 s41.76%
Periodic refresh1830 steps3.667.683 s25.09%
Final refinement500 steps1.002.124 s6.94%
Selective / cached updates19330 steps38.666.416 s20.96%
Others——0.018 s0.06%
Full-update Guard refresh340 steps0.681.592 s5.20%
Total sampling25000 steps50.0030.617 s100.00%

These findings suggest that HASTE's recovery mechanisms help limit the impact of missed support, while maintaining manageable runtime overhead.

Reviewer_a9v1OpenReview ↗

Thank you for the reply.

Thank you for the detailed responses and additional experiments. The matched-budget comparisons, oracle-support analysis, and runtime breakdown substantially address my concerns about the effectiveness and efficiency of HASTE. I appreciate the clarification regarding the timing instrumentation. These results strengthen the paper's empirical evidence.

We are happy to hear that our response addressed your concerns. It was a pleasure to discuss with you! Wish you all the best and continued success in your scientific pursuits!

Reviewer_8ZsuOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HASTE, a hierarchical adaptive diffusion inference acceleration framework for multi-layer Text-to-Layer Generation. The core observation is that different layers have different temporal convergence rates, and effective computation for a single layer is primarily concentrated in non-transparent regions. Based on this, HASTE utilizes a lightweight AlphaNet to predict effective regions and combines layer-level dynamic skipping, token-level sparse computation, and KV cache reuse to reduce redundant computations. Experiments show that this method achieves a 3.22× speedup on Qwen-Image-Layered while largely preserving decomposition quality. It achieves speedups ranging from 1.52× to 4.26× on four other models, and provides ablation experiments and quality analysis.

Critical strengths
  1. This paper designs a reasonable adaptive acceleration mechanism tailored to the characteristics of multi-layer image generation. The paper identifies differences in convergence rates between layers and the sparsity of Alpha Support within layers, and integrates these two redundant structures into a unified inference scheduling framework. This design has clear practical significance.

  2. The method achieves significant acceleration, and the experimental validation is thorough. HASTE outperforms baselines such as PAB, ToCa, and Jano on Qwen-Image-Layered and demonstrates transferability across four additional T2L models. The paper also provides component ablation studies, hyperparameter sensitivity analysis, and hierarchical quality metrics, lending considerable credibility to the performance conclusions.

Critical weaknesses
  1. The innovation of the core method is relatively limited, and the additional benefits of the layered mechanism are insufficient. HASTE primarily combines existing diffusion caching, token-sparse computation, and KV cache reuse techniques. Although Alpha Support provides reasonable domain-specific prior knowledge for T2L scenarios, it remains unclear how much new contribution this combination brings that is independent of existing mechanisms. In particular, as shown in Table 3, using only layer-level time scheduling already achieved a 3.07× speedup. While adding the spatial sparsity mechanism only increased this to 3.19×, and layer-level fidelity decreased slightly. This undermines the necessity of the two-layer hierarchical design emphasized in the paper.​

  2. There is uncertainty regarding the method’s applicability. ​The main experiments used only 1,000 layouts, and primarily covered images with 2–6 layers. Meanwhile, AlphaNet was trained solely on a small number of trajectories derived from four development datasets. There is currently insufficient evidence to determine whether prediction errors would lead to irrecoverable content loss when dealing with more complex layer structures.

Questions for authors
  1. Given that time scheduling alone achieved a 3.07× speedup, while the full HASTE method achieved 3.19×, could the authors further elaborate on the independent benefits of Alpha-Guided Spatial Selection under the same quality constraints? For example, compared to an approach that uses only layer-level scheduling, does the additional spatial sparsity mechanism offer sufficient practical advantages to make the core method’s contribution more prominent?

  2. Could the authors provide additional information on the peak VRAM requirements of the current method when applied to more layers or higher resolutions. It can expand the analysis of the method’s applicability across different scenarios.

3 comments
AuthorsOpenReview ↗

Response (Part I)

We sincerely appreciate your responsible review and valuable suggestions!

Weakness 1: The innovation of the core method is relatively limited, and the additional benefits of the layered mechanism are insufficient. HASTE primarily combines existing diffusion caching, token-sparse computation, and KV cache reuse techniques.

Response: HASTE only draws on general caching and sparse-computation ideas, but its core ideas are fundamentally new: HASTE is the first to identify the hierarchical redundancy unique to image-to-layer decomposition and exploit this property to achieve layer-level and alpha-guided sparse execution for the first time. As Reviewer nXou also noted: "This is the first caching work targeting this domain, which provides a starting point distinct from general text-to-image caching."

Specifically, our contributions differ from existing methods in two key aspects:

  1. Layer-aware temporal scheduling. HASTE is the first to identify heterogeneous convergence rates across output RGBA layers and to exploit them for layer-level caching. Unlike conventional diffusion caching approaches that operate at the network-block or token level, HASTE selectively skips entire stabilized output layers based on their convergence behavior.

  2. Alpha-guided sparse execution. Rather than relying on generic token-importance scores or fixed refresh schedules, HASTE leverages the task-specific alpha support to identify tokens that require fresh computation. Its context-preserving sparse DiT execution further combines fresh queries with persistent cached K/V, maintaining full cross-layer attention despite selective computation.

Crucially, these are not merely independent techniques combined together: the same alpha-support prior guides both temporal convergence estimation and spatial token selection, enabling coordinated computation reuse at two complementary granularities. This task-specific hierarchical design constitutes the central methodological contribution of HASTE, rather than an incremental combination of existing caching and sparsity techniques.

Weakness 1&Q1: Although Alpha Support provides..., it remains unclear how much new contribution this combination brings that is independent of existing mechanisms. Given that time scheduling alone achieved a 3.07× speedup in Table 3, while the full HASTE method achieved 3.19×, could the authors further elaborate on the independent benefits of Alpha-Guided Spatial Selection under the same quality constraints?

Response: We clarify that Table 3 is not the right evaluation for isolating the spatial contribution. It uses the long-trajectory T=50T=50 setting, where temporal redundancy dominates: layer-level skipping already removes most of the redundancy, leaving little spatial redundancy for the spatial mechanism to exploit. The small +3.94%+3.94\% gain in Table 3 is therefore expected and does not undermine the two-layer design.

The independent value of Alpha-Guided Spatial Selection becomes clear when spatial redundancy is high. As shown below, under matched quality, spatial selection yields an additional 10.07% latency reduction over layer-only scheduling on Qwen (T=15T=15) and 9.32% on RevealLayer. On Ming, spatial selection alone (1.41×) even outperforms layer-only scheduling alone (1.27×), and spatial selection achieves a 16.35% incremental gain. These results show that temporal and spatial sparsity target complementary redundancies, and neither dominates universally.

Backbone & ScheduleFull (s)Layer-level scheduling only (s)Spatial-selection only (s)HASTE (Layer-level+Spatial) (s)Incremental Spatial Gain ↑
Qwen (T=50T=50, Original)97.3131.74 (3.07×)62.95 (1.55×)30.49 (3.19×)+3.94%
Qwen (T=15T=15, Few-step)46.1830.89 (1.50×)33.13 (1.40×)27.78 (1.66×)+10.07%
RevealLayer141.5292.32 (1.53×)127.51 (1.11×)83.72 (1.69×)+9.32%
Ming140.88111.31 (1.27×)100.05 (1.41×)93.11 (1.51×)+16.35%

The Alpha Support ablation shows that the Alpha guidance is what makes both layer-level scheduling and spatial selection efficient and quality-preserving. Specifically, our Alpha Support outperforms both whole-canvas and random Alpha Support.

VariantLatency (s) ↓Speedup ↑LPIPS ↓GT αIoU ↑GT RGB L1 ↓
HASTE with Approximated Alpha Support (ours)30.723.22×0.015270.884790.06990
HASTE w/o Alpha Support (Whole canvas)34.582.86×0.017390.873660.07015
HASTE with Random Alpha Support34.662.85×0.020160.863660.07215
HASTE with Oracle Alpha Support37.232.65×0.015250.885170.06930

In short, the two-layer design is not a marginal add-on. It is what makes HASTE robust across redundancy regimes and architectures, where the dominant redundancy shifts between temporal and spatial dimensions.

AuthorsOpenReview ↗

Response (Part II)

Weakness 2&Q2: There is uncertainty regarding the method’s applicability. The main experiments used only 1,000 layouts, and primarily covered images with 2–6 layers. Could the authors provide additional information on the peak VRAM requirements of the current method when applied to more layers or higher resolutions. It can expand the analysis of the method’s applicability across different scenarios.

Response: Thanks for raising these important concerns. Our experiments demonstrate that HASTE remains effective on more complex layer structures and higher resolutions, achieving consistent acceleration while preserving decomposition quality. We further report peak GPU memory usage to clarify its practical requirements.

(1) Generalization to more layers (N=7–15). We extended our evaluation to 2,000 Crello layouts on 4×A100 40GB GPUs, preserving aspect ratios with a 512-pixel short side.

Paired values denote Full / HASTE. peak VRAM is measured per GPU; GT metrics use LayerD at edit budget k=5.

Layers(N)Peak VRAM (GiB)Latency (s) ↓Speedup ↑GT αIoU ↑GT RGB L1 ↓
721.32 / 25.87106.11 / 33.223.19×0.78588 / 0.782220.10211 / 0.11664
821.83 / 27.41132.57 / 42.243.14×0.78587 / 0.802880.13968 / 0.13262
921.82 / 27.38136.10 / 44.103.09×0.61178 / 0.617480.17140 / 0.18949
1022.37 / 29.06157.94 / 49.553.19×0.60949 / 0.585430.18768 / 0.18927
1122.65 / 29.89207.39 / 65.483.17×0.41681 / 0.388750.28429 / 0.28965
1221.65 / 26.85200.72 / 63.983.14×0.45211 / 0.482580.22299 / 0.20808
1321.75 / 27.15209.75 / 65.673.19×0.37383 / 0.396720.26136 / 0.24711
1424.93 / 36.86277.14 / 84.283.29×0.44392 / 0.438260.22626 / 0.22683
1523.31 / 31.93228.59 / 71.113.21×0.35840 / 0.345950.23341 / 0.22669
All (7–15)24.93 / 36.86183.07 / 57.443.19×0.54080 / 0.540680.20263 / 0.20242

(2) Scalability to higher resolutions.

Paired values denote Full / HASTE; peak VRAM is measured per GPU.

ResolutionHardwarePeak VRAM (GiB)Latency (s) ↓Speedup ↑GT αIoU ↑GT RGB L1 ↓
512×5122×A6000 48GB38.63 / 41.6066.52 / 23.312.85×0.68416 / 0.697010.24346 / 0.24243
768×7682×A6000 48GB39.33 / 45.74181.42 / 57.123.18×0.76314 / 0.768920.20045 / 0.19754
1024×10244×A100 40GB24.93 / 36.861007.80 / 287.623.50×0.19761 / 0.190380.30805 / 0.30106

Thanks for providing the additional results and analysis. I don't have any further questions for the authors. I recommend weak acceptance.

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

This paper proposes HASTE, a training-free online caching framework for accelerating layered image decomposition models that jointly denoise multiple RGBA layers. Motivated by the observation that layers stabilize at different denoising steps and that content within each layer is confined to a sparse alpha support, HASTE uses a lightweight AlphaNet to estimate alpha support online. It then skips converged layers via endpoint extrapolation and recomputes only tokens within the alpha support of active layers, while preserving cross-layer context through persistent KV caches. On Crello, HASTE accelerates Qwen-Image-Layered by 3.22× with negligible quality loss and achieves 1.52–4.26× speedups on four other decomposition models.

Critical strengths
  1. The problem setting is timely and the observations are clear. Since the sequence length of layered decomposition models grows proportionally with the number of layers, there is a real demand for acceleration. To my knowledge, this is the first caching work targeting this domain, which provides a starting point distinct from general text-to-image caching.

  2. The transfer across multiple backbones and the measurement protocol are thorough. The method achieves 1.52–4.26× speedups on four models with different samplers and step counts, including Qwen-based, FLUX-based, RevealLayer, and Ming. The speedup is computed as the ratio of paired total sampling times rather than an average over layouts, and it includes online overhead.

  3. The validity of the priority score is quantitatively verified. Figure 3 shows a strong correlation between the online score and the actual endpoint change, which supports the core assumption that the evolution of skipped layers can be estimated without observing them.

Critical weaknesses
  1. The readability is poor. The figures contain an excessive amount of text, and even content that could be explained in the main text is placed in the figures, making it difficult to understand what each figure is intended to convey. Figures should be designed so that the motivation and method can be understood intuitively, but attempting to include nearly all of the paper’s content in the figures significantly reduces readability.

  2. The baseline comparisons are insufficient. The most basic baseline, reducing the number of sampling steps, is missing. The method should also be compared with the original model using a reduced number of steps that yields a latency similar to that of the proposed method, rather than only with T=50. Performing this task with T=50 is inherently inefficient in the first place. Since the proposed method already performs slightly below the T=50 baseline, I expect its quality to be similar to that of a reduced-step baseline.

  3. The task name is inaccurate. The paper refers to the task as “Text-to-Layer” throughout, but all evaluated models are image decomposition models that take an input image as a condition. This is also inconsistent with “Image Decomposition” in the title. In text-only layer generation methods (e.g., LayerDiff and DreamLayer), the input image does not constrain the structure, so whether the same redundancy holds needs to be verified separately.

  4. The contribution of the alpha-support prior, which is the core claim, is not isolated through ablation. In Table 3, temporal selection alone yields 3.07×, and adding spatial token selection only slightly increases the speedup to 3.19×. This means that most of the acceleration comes from layer-level skipping. In that case, the main role of the alpha support is to define the pooling region in the temporal stage. This is precisely where the transparent canvas disturbance and layer-scale disparity, which the paper uses as motivation, become relevant. However, there are no comparisons that replace support-restricted pooling with whole-canvas pooling, remove the κ calibration, or use an oracle alpha or a random mask instead of AlphaNet. Without these experiments, the contribution is not demonstrated beyond applying TeaCache-style skipping at the layer level. The motivation figure is also based on a single example (the Halloween poster).

Questions for authors
  1. Could you compare HASTE with the original model using a reduced number of sampling steps that yields similar latency, in terms of both decomposition quality and speed?

  2. Could you provide ablations that isolate the contribution of the alpha-support prior in the temporal stage, such as replacing support-restricted pooling with whole-canvas pooling, removing the κ calibration, and using an oracle alpha or a random mask instead of AlphaNet?

8 comments
AuthorsOpenReview ↗

Response (Part I)

Thanks for the responsible review and valuable suggestions.

Weakness 1: The readability of figures is poor.

Response: Constructive suggestion! We have revised the figures to make it easier to understand. Please see Figures 2 and 4.

Weakness 2&Q1: The baseline comparisons are insufficient. The most basic baseline, reducing the number of sampling steps, is missing. The method should also be compared with the original model using a reduced number of steps that yields a latency similar to that of the proposed method, rather than only with T=50. Could you compare HASTE with the original model using a reduced number of sampling steps that yields similar latency, in terms of both decomposition quality and speed?

Response: Good observation! We show that, at comparable latency/speed to the reduced-step baselines, simply reducing sampling steps would lead to significant quality losses (e.g., LPIPS) whereas our HASTE achieves near-lossless quality relative to the 50-step baseline:

MethodStepsLatency (s) ↓Speedup ↑GT αIoU ↑GT RGB L1 ↓LPIPS ↓SSIM ↑PSNR ↑
Qwen50 (baseline)99.251.00×0.88740.066201+∞+\infty
Qwen2039.432.52×0.87640.06870.02030.972632.37
Qwen1529.563.36×0.87090.06920.02890.965031.08
HASTE5030.813.22×0.88710.06780.00810.987235.50

We highlight that the reduced-step baselines are orthogonal to, not competitors of, our method; we can further accelerate them by 1.62×–1.94×:

MethodStepsLatency (s) ↓Speedup ↑GT αIoU ↑GT RGB L1 ↓LPIPS ↓SSIM ↑PSNR ↑
Qwen2039.431.00×0.87640.068701+∞+\infty
HASTE2019.951.94×0.88730.06740.01320.981833.70
Qwen1529.561.00×0.87090.069201+∞+\infty
HASTE1517.921.62×0.87050.07170.05160.945236.48

Weakness 3: The task name is inaccurate. The paper refers to the task as “Text-to-Layer” throughout, but all evaluated models are image decomposition models that take an input image as a condition. In text-only layer generation methods (e.g., LayerDiff and DreamLayer), whether the same redundancy holds needs to be verified separately.

Response: Good suggestion! We have revised the manuscript to consistently use "Image-to-Layer (I2L) " instead of "Text-to-Layer(T2L)" to align with the paper's task. Meanwhile, we confirm that the same redundancy holds in text-only layer generation. Hence, HASTE's spatio-temporal reuse can also benefit text-only layer generation methods, beyond image decomposition.

Since LayerDiff and DreamLayer have not released their implementations, we conduct experiments on two open-source Text-only Layer generation models, ART+[1] and LayerDiffuse[2]. HASTE achieves a 1.27-2.02× speedup while maintaining high fidelity.

T2L ModelFull Latency (s)HASTE Latency (s)SpeedupLayer LPIPS ↓Layer SSIM ↑Composite LPIPS ↓
ART+22.0510.902.023×0.038430.966640.10250
LayerDiffuse2.9842.3551.267×0.051440.926540.05701

Ref.:

[1] Junwen Chen et al. PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models. arXiv preprint arXiv:2505.22523, 2025.

[2] Lvmin Zhang and Maneesh Agrawala. Transparent Image Layer Diffusion using Latent Transparency. ACM Transactions on Graphics, 43(4), Article 100, 2024.

AuthorsOpenReview ↗

Response (Part II)

Weakness 4: The contribution of the alpha-support prior, which is the core claim, is not isolated through ablation. In Table 3, temporal selection alone yields 3.07×, and adding spatial token selection only slightly increases the speedup to 3.19×. This means that most of the acceleration comes from layer-level skipping.

Response: We clarify that Table 3 is not the right evaluation for isolating the spatial token selection contribution. It uses the long-trajectory T=50T=50 setting, where temporal redundancy dominates: layer-level skipping already removes most of the redundancy, leaving little spatial redundancy for the spatial mechanism to exploit. The small +3.94%+3.94\% gain in Table 3 is therefore expected and does not undermine the alpha-support prior / two-layer design.

The independent value of Alpha-Guided Spatial Selection becomes clear when spatial redundancy is high. As shown below, under matched quality, spatial selection yields an additional 10.07% latency reduction over layer-only scheduling on Qwen (T=15T=15) and 9.32% on RevealLayer. On Ming, spatial selection alone (1.41×) even outperforms layer-only scheduling alone (1.27×), and spatial selection achieves a 16.35% incremental gain. These results show that temporal and spatial sparsity target complementary redundancies, and neither dominates universally.

Backbone & ScheduleFull (s)Layer-level scheduling only (s)Spatial-selection only (s)HASTE (Layer-level+Spatial) (s)Incremental Spatial Gain ↑
Qwen (T=50T=50, Original)97.3131.74 (3.07×)62.95 (1.55×)30.49 (3.19×)+3.94%
Qwen (T=15T=15, Few-step)46.1830.89 (1.50×)33.13 (1.40×)27.78 (1.66×)+10.07%
RevealLayer141.5292.32 (1.53×)127.51 (1.11×)83.72 (1.69×)+9.32%
Ming140.88111.31 (1.27×)100.05 (1.41×)93.11 (1.51×)+16.35%

Weakness 4&Q2: There are no comparisons that replace support-restricted pooling with whole-canvas pooling, remove the κ calibration, or use an oracle alpha or a random mask instead of AlphaNet. Could you provide ablations that isolate the contribution of the alpha-support prior in the temporal stage, such as replacing support-restricted pooling with whole-canvas pooling, removing the κ calibration, and using an oracle alpha or a random mask instead of AlphaNet?

Response: The empirical ablation results below substantiate that Alpha Support prior is crucial for effective temporal scheduling. Our approximated Alpha Support outperforms whole-canvas, random Alpha Support, and κ calibration removal. Oracle Alpha is slower (37.23s / 2.65×) because it requires decoding the ground-truth alpha values. Thus Alpha Support prior is what makes both layer-level scheduling and spatial selection efficient and quality-preserving.

VariantLatency (s) ↓Speedup ↑LPIPS ↓GT αIoU ↑GT RGB L1 ↓
HASTE with Approximated Alpha Support (ours)30.723.22×0.015270.884790.06990
HASTE w/o Alpha Support (Whole canvas)34.582.86×0.017390.873660.07015
HASTE with Random Alpha Support34.662.85×0.020160.863660.07215
HASTE with Oracle Alpha Support37.232.65×0.015250.885170.06930
HASTE w/o κ calibration33.552.95×0.019200.874600.07180

In short, the two-layer design is not a marginal add-on. It is what makes HASTE robust across redundancy regimes and architectures, where the dominant redundancy shifts between temporal and spatial dimensions. We have incorporated these clarifications and ablations into the revision.

AuthorsOpenReview ↗

The authors are looking forward to your feedback. Let's discuss.

Dear Reviewer nXou,

We sincerely appreciate your time and constructive feedback on our submission. We hope that our detailed responses, particularly the matched-latency reduced-step comparisons, the ablations demonstrating the independent contributions of spatial selection and alpha-support guidance, and the new experiments on text-conditioned layer generation, have addressed your main concerns. We have also revised the figures and task terminology as suggested.

We understand that you may be handling multiple submissions and have a busy schedule. However, with less than two days remaining in the author–reviewer discussion period, we would greatly appreciate any feedback or remaining concerns you may have, so that we can address them before the discussion closes.

Thank you again for your time and consideration. We look forward to hearing your thoughts.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the reviewers for carefully reviewing this initial manuscript and are encouraged by the positive assessment of the first caching work targeting layered image decomposition; a starting point distinct from general text-to-image caching (Reviewer nXou), empirical thoroughness (Reviewer nXou, 8Zsu), and meaningful Alpha support for small foreground layers ( Reviewer a9v1).

The reviewers raised several common concerns: (i) the independent contribution of the alpha-support prior and the two-level hierarchical design is not sufficiently isolated; (ii) reduced-step and matched-latency baselines are missing; (iii) figure readability and task naming need improvement; (iv) timing instrumentation and full-update guard costs require clarification; and (v) applicability to more layers, higher resolutions, and VRAM should be analyzed. In response, the revised manuscript adds the following analyses and revisions:

  • We added ablations isolating the alpha-support prior, including whole-canvas pooling, random support, oracle support, and removing κ\kappa calibration. The proposed AlphaNet-based support achieves the best efficiency–quality trade-off.

  • We added matched-FLOPs and matched-query-budget comparisons among global, per-layer, and support-aware scheduling under the same endpoint/KV implementation, where support-aware HASTE consistently achieves better latency and decomposition quality.

  • We added reduced-step and matched-latency comparisons. At comparable latency, HASTE achieves near-lossless quality relative to the 50-step baseline, whereas simply reducing sampling steps incurs significant quality loss; HASTE also further accelerates 20-step and 15-step baselines by 1.94× and 1.62×.

  • We revised the figures to improve readability and consistently renamed the task as Image-to-Layer (I2L). We further verify that the same spatio-temporal redundancy holds for text-only layer generation, where HASTE achieves 1.27–2.02× speedup on ART+ and LayerDiffuse.

  • We clarified that Table 3 uses identical timing instrumentation and corrected the documentation error in Appendix D.5. We further show that the small +3.94% gain at T=50T=50 is expected because temporal redundancy dominates; when spatial redundancy is high, alpha-guided spatial selection provides an additional 9.32%–16.35% latency reduction over layer-only scheduling.

  • We added robustness analyses to missed alpha-support regions, full-update guard trigger frequencies and runtime costs, and scalability results on 7–15 layers and up to 1024×1024 resolution, including peak VRAM.

We sincerely invite the reviewers to refer to the revised figures, new reduced-step and alpha-support ablations, matched-budget scheduling comparison, and scalability/VRAM analyses for details. We believe these additions address the common concerns.

Last but not least, we thank the PCs, ACs, and reviewers again for their invaluable time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.