acceptodds
Under review as a conference paper at ICLR 2027

The Golden Path Hypothesis: Reusable Schedules in Diffusion Caching

Abstract

Diffusion caching accelerates generation by replacing transformer computation with cached or predicted features at selected denoising steps. We introduce the **Golden Path Hypothesis** (GPH): under fixed inference conditions, prompt-independent cache schedules can achieve final-output quality comparable to the best prompt-specific schedules across prompts. We investigate GPH across ten caching methods, four image and video models, and three cache ratios. Prompt-adaptive methods repeatedly select a small number of schedules, and reusing their most frequent schedules on new prompts closely matches the quality of prompt-specific choices. Exhaustive evaluation of 1.4 million schedules on four examples further identifies a prompt-independent schedule that remains competitive on unseen prompts. To explain this transfer, we analyze denoising trajectories and the accumulation of caching errors. Latent-state trajectories exhibit similar structure across datasets and seeds, while an exact error decomposition shows that accumulated effects of earlier errors predict final latent-state error better than local approximation errors. This motivates searching for complete schedules using final-output quality. With only a small set of examples, the resulting golden paths transfer across prompts and datasets and can be tuned to the desired quality objective, including reconstruction fidelity or perceptual similarity.

Primary area:
generative models
Keywords:
diffusion model, diffusion caching

Reviews

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

This paper studies the caching schedule for accelerating diffusion inference: which denoising steps run a full transformer evaluation, and which steps reuse or forecast features. It proposes the Golden Path Hypothesis: under fixed inference conditions, there exists a prompt-independent schedule whose final output quality matches that of the best prompt-specific schedule for each prompt.

Critical strengths

The research question is important, as a general schedule can provide inspiration for the community and avoid searching for each prompt individually. However, such a general schedule may be strongly model-dependent and unable to adapt to various general models.

Critical weaknesses
  1. Poor readability of the paper: The values in Table 5, Table 6, and Table 7 are stacked without proper labeling; the lack of a Pipeline diagram for the main method reduces readability.
  2. Lack of Novelty: Scoring the complete schedule based on the final output quality is not a novel idea: BudCache already evaluates the complete cache generation using final latent MSE, and the main difference of this paper from BudCache is merely changing the target to PSNR/LPIPS and adding more scoring samples.
Questions for authors
  1. Please explain the core innovations in relation to existing work.
  2. How well does it generalize across models?
Reviewer_g2MgOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether a fixed schedule can remain competitive across prompts when the model, sampler, noise grid, cache ratio, approximation policy, and evaluation metric are held fixed. It analyzes recurring schedules from adaptive methods, replays shared schedules on held-out prompts, performs an exhaustive high-cache-ratio search on four FLUX examples, and evaluates selected schedules on larger sets. It also studies error propagation and searches schedules using final-output metrics. The resulting evidence substantially informs the design of diffusion caching, although general near-optimality and the attribution of search gains need more careful qualification.

Critical strengths
  1. The paper tests an important assumption across a broad set of methods. The study covers ten caching methods and four backbones. Selecting recurring schedules on 544 prompts and replaying them on 1088 different prompts provides meaningful evidence that a schedule can transfer beyond the inputs used to choose it. This finding can simplify deployment and challenges the assumption that each prompt requires a substantially different schedule.

  2. The exhaustive search gives concrete evidence about schedule quality. Evaluating all 1,370,754 feasible schedules for four FLUX prompt–seed examples at cache ratio 0.82 allows direct comparison with the optimum in that setting. Evaluating selected schedules on 2381 new prompts with three seeds then tests transfer. Although the broader comparison uses a restricted candidate pool, combining exhaustive analysis with larger-scale evaluation is a substantial empirical contribution.

  3. The error analysis explains why local cache metrics can be misleading. The paper connects final distortion to both the injected residual error and its subsequent propagation. In Table 1, propagation weighting raises the reported correlations from approximately 0.53–0.68 to 0.89–0.90. The different schedules preferred by PSNR and LPIPS also show that the evaluation objective is part of the scheduling problem. These results provide insight beyond simply reporting that one schedule performs well.

Critical weaknesses
  1. The evidence for near-optimality is narrower than the evidence for reuse. The exhaustive comparison covers only four examples under one model and cache ratio; even there, the schedule selected by mean PSNR is 1.249 dB below the optimum on one example. On the 2381 new prompts, the reference is the best of 337 selected candidates, rather than the full feasible schedule space. The reported coverage within 0.5 or 1 dB therefore shows competitiveness within that pool. It does not establish equally small gaps to each new prompt’s true optimum. The conclusions should preserve this distinction.

  2. The search comparison does not isolate the value of the final-output objective. The proposed search combines multiple procedures, starting schedules, and validation selection, while the BudCache comparison uses different scoring data and a smaller search budget. Its gains can therefore reflect search effort and initialization as well as the objective. The PSNR-versus-LPIPS comparison within a procedure is informative, but a matched comparison using the same optimizer, initialization, scoring examples, and evaluation budget is needed to attribute the baseline improvement specifically to optimizing final outputs.

  3. One benchmark does not provide fully independent evidence of transfer. The final PartiPrompts evaluation includes the 544 prompts used to obtain some of the initialization schedules. This overlap is disclosed, and the separate DrawBench, GenEval-style, and DiffusionDB evaluations remain useful independent evidence. However, combining the overlapping and unseen PartiPrompts examples obscures how well the selected schedules transfer on that benchmark. Reporting the two subsets separately would clarify the size of this effect and the scope of the generalization claim.

Questions for authors
  1. How do the held-out coverage curves change when the initial four-example selection set changes, or when the candidate pool includes schedules optimized on independent prompts? Please distinguish the best evaluated candidate from the full-space optimum in the conclusions.

  2. Can you compare local-surrogate and final-output objectives with the same optimizer, starting schedules, scoring data, evaluation budget, and validation rule? Please include the total search cost when selecting among multiple procedures.

  3. Can you report final PartiPrompts results separately for the 544 prompts used to obtain initialization schedules and the remaining prompts? Do the gains persist on the fully held-out subset?

Reviewer_fnsmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper challenges the statement that diffusion caching schedules need to be prompt-specific. Based on extensive schedule studies (schedule-frequency analysis, fixed-schedule reuse and exhaustive search) and experiments (10 caching methods, 4 image and video models, 3 cache ratios) authors introduce the Golden Path Hypothesis (GPH), which propose that under fixed inference conditions prompt-independent cache schedules can be comparable to prompt-specific schedules in terms of final output quality. The authors also analyze diffusion denoising trajectories and caching-error propagation to motivate searching for reusable schedules using final-output quality objectives.

Critical strengths
  1. Authors study a very clear and practically important question

This paper addresses the practically relevant question if we really need prompt-specific schedules while losing inference simplicity. This matters because prompt‑specificity add decisions during inference, while reusable schedules provides fixed computational budget plus fixed sequence of full and cached steps. Shared schedules also simplify batched inference since you don't need to route samples with different caching designs. So, the knowledge when such schedules preserve output quality is rather valuable.

  1. Extensive and diverse experimental evaluation

The study covers 10 caching methods (SeaCache, DiCache, TaylorSeer, etc.), 4 image and video models (FLUX.1-dev, Qwen-Image, HunyuanVideo and Wan2.1) and 3 cache ratios (0.58, 0.74 and 0.82) across multiple datasets and noise seeds. It also provides wide range of schedule experiments: schedule-frequency analysis, replacement of adaptive schedules with fixed ones, exhaustive search and evaluation on unseen prompts. Such a wide range of experiments provides empirical evidence of usefulness of reusable schedules.

Critical weaknesses
  1. The formulation of GPH is broader than the experimental evidence, and the added value of schedule search over prior work is unclear

GPH claims shared schedules can achieve quality comparable to the best prompt-specific ones, but the held-out experiments only compare against the best among a small candidate set (e.g., 337 schedules in Figure 2). The exhaustive comparison covers the full constrained search space, but only on 4 calibration examples. While the authors acknowledge the held-out reference as the “best evaluated,” these results only show that shared schedules compete with tested alternatives on unseen prompts, leaving their closeness to true optima unresolved.

Under this more realistic interpretation, the distinction from prior work becomes less clear for the schedule-search component. Searching for prompt-independent schedules with a fixed compute budget is not new; for example, DPCache [1] and ReCache [2] already address this explicitly. However, the paper lacks direct empirical comparisons against these methods in key settings. In particular, ReCache is especially relevant because it already optimizes reusable schedules using final-output quality, but it is not evaluated. Similarly, DPCache is evaluated in other benchmarks but is omitted from the main schedule-search comparison in Table 2. Without direct comparisons against these approaches, the practical benefit of the proposed search methodology is difficult to assess.

  1. The paper is difficult to follow and lacks a clear line of argument

The paper presents a lot of different experiments, but they sometimes feel separate despite the connections outlined by the authors. Each paper's section describes its own setup and observations, and it is not always clear how the findings build on one another. For example, Section 3 studies schedule recurrence and reuse, Sections 4–5 move to trajectory geometry and error propagation, and Section 6 returns to searching for reusable schedules. The error-propagation analysis motivates final-output scoring, but the role of the extensive geometry analysis in explaining schedule transferability remains unclear. Different experimental protocols and details scattered across the appendices further complicate the reading. As a result, the paper contains many individual observations, but their connection to a coherent explanation of why and when shared schedules work is insufficiently developed.

[1] Cui, Bowen, et al. "Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache." arXiv preprint arXiv:2602.22654 (2026).

[2] Aliev, Mishan, et al. "ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE." arXiv preprint arXiv:2606.06060 (2026).

Questions for authors
  1. GPH

Could you revisit the formulation of GPH and explain more carefully which aspects are supported by the experiments and which remain conjectural?

  1. Comparison with BudCache [1]

In the reported setup BudCache uses 2 or 3 preparation prompts, while the proposed search uses 8 prompt–seed scoring examples and 50 additional validation prompts, with different search procedures and initializations. Could you compare latent MSE and final-output objectives under the same data, initialization, and search budgets to separate their effect from that of additional resources?

  1. Comparison with ReCache [2]

Could you add to the schedule-search results in Section 6 a comparison against ReCache? Since ReCache is agnostic to the underlying caching mechanism, it could be evaluated using the same approximation policy and inference budget, giving readers a concrete reference for interpreting the reported changes in output quality.

  1. Organization and clarity

Could you reorganize the presentation around a clear sequence of research questions, explaining how each experiment builds on the preceding findings? In particular, how do the trajectory-geometry observations explain schedule transferability, and how does the error-propagation analysis inform the search in Section 6? A compact overview of the calibration sets, candidate pools, and evaluation protocols would also make the experimental argument easier to follow.

[1] Lei, Mingkun, et al. "Budget-Constrained Step-Level Diffusion Caching." arXiv preprint arXiv:2606.13496 (2026).

[2] Aliev, Mishan, et al. "ReCache: Learning Budget-Aware Caching Schedules for Diffusion Models via REINFORCE." arXiv preprint arXiv:2606.06060 (2026).

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their responses to your reviews. Please take some time to carefully examine their rebuttals and engage in the discussion by sharing your thoughts, remaining concerns, or follow-up questions.

As the discussion deadline is approaching, I would appreciate your timely participation to allow sufficient time for further exchanges with the authors.

Thank you for your time and valuable contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.