OTCache: Optimal Transport for Geometry-Aware Caching in Diffusion Models¶

Conference: ECCV 2026
arXiv: 2606.31026
Code: https://github.com/UnicomAI/OTCache
Area: Diffusion Models / Image Generation
Keywords: Optimal Transport, Diffusion Model Acceleration, Cache Scheduling, Flow Matching, Training-Free

TL;DR¶

OTCache proposes a training-free acceleration framework for diffusion models. Based on optimal transport theory, it models cache scheduling under different inference budgets as continuous evolutionary trajectories in the policy space. Using a high-budget graph search reference schedule and a low-budget black-box search anchor as two endpoints, it predicts scheduling plans for any target budget through quantile interpolation in Wasserstein space. It achieves 4.5x, 4.7x, and 3.66x speedups on FLUX.1, Qwen-Image, and HunyuanVideo, respectively, consistently outperforming existing graph-search cache baselines.

Background & Motivation¶

Flow Matching has become a core paradigm of modern generative modeling, driving commercial-grade models like FLUX.1 and HunyuanVideo to achieve breakthroughs in image and video synthesis. However, Transformer-based denoisers have massive parameter sizes, resulting in high computational overhead and GPU memory footprints during 50-step iterative sampling, which severely hinders deployment in interactive or resource-constrained scenarios. Training-free caching, which bypasses redundant computation by reusing intermediate features from adjacent steps, has emerged as an important direction for accelerating inference.

Representative methods MeanCache and LeMiCa model cache scheduling as a constrained shortest path problem on a directed graph: nodes represent discrete time steps, and edge weights represent the local speed field error introduced by caching one step, with the optimal schedule being the path with the minimum cumulative edge weights. The core assumption of this paradigm is additive independence, suggesting that end-to-end generation quality loss can be approximated by the sum of local errors at each step. In high NFE intervals (e.g., \(B=20\)), where cache intervals are short and single-step decisions have limited impact on subsequent steps, this assumption generally holds. However, in low NFE intervals (high acceleration ratios), caching intervals are elongated, and errors from early decisions propagate nonlinearly along the denoising trajectory, interacting with subsequent decisions. Consequently, the discrepancy between the additive proxy objective and the true perceptual quality increases sharply. As shown in Fig. 1A, the LPIPS gap between MeanCache and the search-based optimal solution (Optuna) widens significantly as the NFE decreases from 20 to 8. Existing methods solve the shortest path independently for each budget, ignoring the structural correlations between optimal schedules under different budgets, which leads to inadequate scheduling quality in low NFE intervals.

More critically, this paper observes a neglected structural regularity: the gap profiles (the first-order sequence of time differences between adjacent caching steps) of optimal schedules under different NFE budgets are not unrelated but evolve smoothly with the budget (Fig. 1B). This implies that schedules under different budgets are observational snapshots of the same ideal trajectory in the policy space at different resolutions. Inspired by this insight, OTCache models the budget-conditioned scheduling evolution as a geodesic on the policy manifold, using quantile interpolation in optimal transport theory to realize scheduling prediction for any budget.

Core Idea: The optimal cache schedules under different budgets are observations of the same ideal denoising policy at different NFE resolutions, which evolve along a smooth geodesic in the policy space. Leveraging quantile interpolation of optimal transport, any target budget's cache schedule can be predicted in Wasserstein space using two reliable endpoints (a high-budget reference and a low-budget anchor), avoiding biased additive proxy objectives solved independently for each budget.

Method¶

Overall Architecture¶

OTCache is a training-free, three-stage cache scheduling prediction framework. It takes a diffusion model and a target inference budget \(B\) as inputs, and outputs the optimal cache scheduling sequence \(\pi_B\) under that budget (i.e., which steps undergo full denoising and which steps reuse cached features). The three stages collaborate: first, a high-fidelity reference schedule is obtained using graph search under a conservative high budget; second, an anchor schedule is obtained via black-box end-to-end search under a very low budget; finally, both schedules are lifted to continuous warping functions and subjected to quantile linear interpolation in Wasserstein space to predict the schedule for the target budget.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Diffusion Model<br/>+ Target Budget B"] --> B["Reference Schedule<br/>B_ref=20 · MeanCache Graph Search"]
    A --> C["Anchor Search<br/>B_anc=8 · Optuna+LPIPS End-to-End Optimization"]
    B -->|"Continuous Warping μ_ref"| D["OT Quantile Interpolation<br/>PCHIP Continuous Warping<br/>Wasserstein Geodesic Prediction"]
    C -->|"Continuous Warping μ_anc"| D
    D --> E["Target Budget B Cache Schedule π_B<br/>Accelerated Sampling"]

Key Designs¶

1. Reference Schedule: High-Fidelity Graph Search under Conservative Budgets

Graph search methods (such as MeanCache) assume that the end-to-end loss of the cache path can be decomposed into cumulative local edge losses. This assumption holds well in high NFE intervals (conservative budgets). Here, the cache intervals are short, and the error from single-step decisions has limited propagation effects, leading to high alignment between the additive proxy objective and actual generation fidelity. OTCache exploits this property to run MeanCache under a conservative budget \(B_{\text{ref}}=20\) to search for the shortest path, obtaining the reference schedule \(\pi_{\text{ref}} = C_{\text{graph}}(B_{\text{ref}})\). This schedule acts as a high-fidelity reference; its gap profile (the distribution of time intervals between adjacent cache steps) encodes the chronological structure of the denoising ODE trajectory, supplying a stable geometric prior for subsequent budget-conditioned predictions. Without this reference endpoint, a low-budget anchor alone cannot capture the refined trajectory shape under dense sampling.

2. Anchor Search: End-to-End Perceptual Optimization under Very Low Budgets

In very low NFE intervals (e.g., \(B=8\)), additive proxy objectives become severely distorted—cache intervals are large, errors propagate nonlinearly, and the "shortest" path no longer corresponds to the "optimal" one. OTCache abandons proxy objectives and directly optimizes the end-to-end perceptual loss. Using the output of full-step, cache-free sampling as a reference, it minimizes the LPIPS distance between the cache-scheduled generation and the reference:

\[\mathcal{J}(\pi) = \ell_{\text{LPIPS}}(x_0(\pi), x_0)\]

where \(x_0(\pi)\) is the image/video generated by executing schedule \(\pi\), and \(x_0\) is the 50-step full-sampling reference output under the same prompt and seed. This objective directly measures the perceptual fidelity loss introduced by acceleration, naturally avoiding the systematic bias of proxy objectives.

The search uses Optuna combined with a CMA-ES sampler, with a budget of 200 trials and early stopping patience=50. Two practical tricks improve search efficiency: (1) Warm start—initializing the search with the schedule solved by MeanCache under the same budget to provide a strong prior, falling back to MeanCache if the search fails to find a better solution, ensuring performance is no worse than the baseline; (2) First-order gap parameterization—searching in the gap space of adjacent time differences rather than absolute time steps. The gap space naturally preserves temporal monotonicity (positive gaps guarantee increasing time steps) and empirically offers better search terrain conditioning. The resulting anchor schedule \(\pi_{\text{anc}}\) at \(B=8\) reduces LPIPS by approximately 25% on average compared to MeanCache, providing a reliable very-low-budget endpoint for subsequent interpolation.

3. OT Quantile Interpolation: Budget-Conditioned Schedule Evolution

The core hypothesis of this design is that optimal schedules under different budgets are not unrelated independent solutions, but rather observational snapshots of the same ideal trajectory in the policy space at different resolutions—changes in budget correspond to continuous sliding along this trajectory. Therefore, scheduling evolution can be modeled as a geodesic on this policy manifold. The specific steps are as follows:

(1) PCHIP Continuization: Slice-wise monotonic piecewise cubic Hermite interpolating polynomials (PCHIP) are used to lift discrete schedules \(\pi_{\text{ref}}\) and \(\pi_{\text{anc}}\) to strictly monotonic, continuous warping functions \(\mu_{\text{ref}}(u)\) and \(\mu_{\text{anc}}(u)\) on a normalized progression \(u \in [0,1]\). PCHIP guarantees monotonicity without spurious oscillations. Subsequently, uniform sampling at a fixed resolution yields equal-dimensional representations, aligning two schedules of different lengths in the same coordinate system.

(2) Wasserstein Geodesic Interpolation: In the one-dimensional case, the Wasserstein geodesic between two probability distributions is equivalent to the linear combination of their quantile functions. OTCache directly treats the warping functions \(\mu(u)\) as quantile functions, predicting the schedule for a target budget \(B\) as:

\[\mu_B(u) = \alpha_{\text{anc}}(B) \cdot \mu_{\text{anc}}(u) + \alpha_{\text{ref}}(B) \cdot \mu_{\text{ref}}(u)\]

The weights \(\alpha_k(B)\) are normalized from the product of a distance prior \(w_{\text{dist}} = (|B_k - B| + 1)^{-1}\) (proximity—favoring whichever known budget is closer) and a confidence prior \(w_{\text{conf}} = \log B_k\) (giving higher trust to the high-budget reference because dense sampling characterizes the ODE trajectory more faithfully). Crucially, the intuition behind this weighting design is clear: known schedules closer to the target budget receive greater weight, while the high-budget reference gains extra trust due to its higher structural reliability.

(3) Conservative Discretization: When recovering discrete time steps from the continuous curve \(\mu_B\), a power-law warp \(\rho \ge 1.0\) is introduced to bias the NFE density toward the early stages of the ODE: \(t_i = P_{\text{sum}}(\mu_B([i/(B-1)]^\rho))\), where \(P_{\text{sum}}\) is a sum projection operator that rounds sampling points to integers and reallocates residuals to the largest sampling intervals, ensuring strict compliance with the maximum time step \(T_{\text{max}}\) constraint. The motivation for the \(\rho\) bias comes from empirical observations: the vector field fluctuates heavily in the early stages of denoising, requiring denser full computation; whereas the late-stage trajectory tends to be smooth, allowing for more aggressive caching. Larger \(\rho\) values lead to higher densities in early steps.

A Complete Example¶

For instance, predicting a FLUX.1 [dev] schedule under target budget \(B=10\):

Stage 1 searches under \(B_{\text{ref}}=20\) using MeanCache to obtain reference schedule \(\pi_{\text{ref}}\): 20 time steps covering \([50, 1]\), with a relatively uniform gap distribution. Stage 2 searches under \(B_{\text{anc}}=8\) using Optuna to obtain anchor schedule \(\pi_{\text{anc}}\): 8 time steps, with a gap distribution significantly biased toward the early stages (as the vector field fluctuates wildly early on, the search automatically allocates more steps there). The gap profiles of both share similar shapes but differ in resolution—which is exactly the "snapshots of the same trajectory at different resolutions" shown in Fig. 1B.

Stage 3 applies PCHIP continuization to \(\pi_{\text{ref}}\) and \(\pi_{\text{anc}}\), resulting in two monotonic warping curves \(\mu_{\text{ref}}\) and \(\mu_{\text{anc}}\). For target \(B=10\), the distance weights \(w_{\text{dist}}\) are \((|8-10|+1)^{-1}=1/3\) and \((|20-10|+1)^{-1}=1/11\) respectively, and the confidence weights \(w_{\text{conf}}\) are \(\log 8 \approx 2.08\) and \(\log 20 \approx 3.00\). Combining these, \(\alpha_{\text{ref}}\) is slightly higher than \(\alpha_{\text{anc}}\) (although \(B=10\) is closer to the anchor budget 8, the high confidence of the reference compensates), yielding an interpolated \(\mu_{10}\) that preserves the fine structure of the reference while absorbing the aggressive caching strategy of the anchor at very low budgets. Finally, power-law sampling with \(\rho=1.3\) on \(\mu_{10}\) produces 10 discrete time steps biased toward early dense allocation. The entire process requires zero model execution, using only closed-form mathematical operations (interpolation + sampling).

Loss & Training¶

OTCache is a completely training-free framework that requires no model fine-tuning or distillation. The only optimization-involved step is the Stage 2 anchor search: using LPIPS as the perceptual loss, the optimal schedule at \(B_{\text{anc}}=8\) is searched via Optuna (CMA-ES sampler) within a budget of 200 trials and an early stopping patience of 50. It employs a MeanCache-initialized warm start and first-order gap parameterization. The interpolation in Stage 3 involves only closed-form mathematical operations (PCHIP fitting + weighted linear combination + power-law sampling + sum-projection) with no learnable parameters. The power-law warp parameter \(\rho\) is determined via grid search: on FLUX.1, \(\rho \in \{1.00, 1.15, 1.30, 1.45\}\), where \(\rho = 1.30\) achieves the best PSNR/LPIPS balance (\(\text{PSNR}=26.03, \text{LPIPS}=0.126\)), which is kept fixed globally. The offline calibration cost in Stage 2 is approximately 7.7 min (FLUX.1), 18.2 min (Qwen-Image), and 50.5 min (HunyuanVideo) per prompt in the worst case, but there is zero online inference overhead after calibration, and the Stage 3 interpolation computation is negligible.

Key Experimental Results¶

Main Results¶

Table 1: FLUX.1 [dev] (1024x1024) Acceleration Performance Comparison

Method	Speedup	LPIPS ↓	SSIM ↑	PSNR ↑	ImageReward ↑	CLIP ↑
Original 50 Steps	1.00x	–	–	–	1.033	31.229
TeaCache (l=1.5)	3.66x	0.504	0.624	15.01	0.717	30.696
TaylorSeer (N=6,O=2)	2.74x	0.415	0.663	16.28	0.971	31.310
LeMiCa (B=15)	2.80x	0.153	0.858	24.45	0.991	31.125
LeMiCa (B=10)	3.60x	0.312	0.740	19.03	0.981	31.355
MeanCache (B=15)	2.91x	0.142	0.870	24.83	1.010	31.244
MeanCache (B=10)	4.12x	0.272	0.761	19.43	0.993	31.323
OTCache (B=15)	3.04x	0.126	0.881	26.03	1.011	31.167
OTCache (B=10)	4.50x	0.254	0.780	20.22	0.996	31.269

In FLUX.1, OTCache achieves the lowest \(\text{LPIPS}=0.126\) with a \(3.04\times\) speedup at \(B=15\) (MeanCache has \(\text{LPIPS}=0.142\) at \(2.91\times\)); under extreme acceleration of \(4.50\times\), it still maintains \(\text{ImageReward}=0.996\) and \(\text{LPIPS}=0.254\), whereas TeaCache's LPIPS degrades to 0.504 and CLIP drops to 30.696 at a similar speedup. Qualitative results indicate that under high acceleration, baseline methods experience frequent content inconsistencies and object distortions (e.g., deformed legs of the red goat, unnatural scissors, redundant piano keys), whereas OTCache maintains better content consistency and visual quality even under faster acceleration.

On Qwen-Image (\(1664\times928\), 16:9), OTCache (\(B=15\)) achieves a nearly lossless \(\text{LPIPS}=0.069\) at \(3.20\times\) speedup (MeanCache achieves \(\text{LPIPS}=0.075\) at \(2.85\times\)); under extreme acceleration of \(4.70\times\) (\(B=10\)), it maintains \(\text{SSIM}=0.864\) and \(\text{PSNR}=21.48\), far exceeding MeanCache at the same budget (\(\text{LPIPS}=0.236, \text{SSIM}=0.815\)). Qualitatively, in 16:9 poster-style generation, OTCache maintains significantly better content consistency before and after acceleration, whereas MeanCache and LeMiCa suffer from changes in relative object positions (cat and octopus) and subject orientations (giraffe).

On HunyuanVideo video generation, OTCache (\(B=12\)) achieves \(\text{VBench}=80.20\%\) and \(\text{LPIPS}=0.162\) under a \(3.21\times\) speedup (MeanCache gets \(\text{LPIPS}=0.176\) under the same speedup); under extreme acceleration of \(3.66\times\) (\(B=10\)), it achieves \(\text{VBench}=80.37\%\) and \(\text{LPIPS}=0.252\), while DiCache and TeaCache already show prominent temporal artifacts and content degradation at similar speedups.

Ablation Study¶

Table 2: Impact of Power-Law Parameter \(\rho\) on Reconstruction Quality (FLUX.1, B=15)

\(\rho\)	LPIPS ↓	PSNR ↑	SSIM ↑	ImageReward ↑
1.00 (Uniform Sampling)	0.139	24.304	0.864	1.022
1.15	0.134	24.847	0.871	1.000
1.30	0.126	26.034	0.880	1.011
1.45	0.131	25.798	0.877	1.001

\(\rho=1.30\) achieves the best reconstruction accuracy, validating that biasing the NFE density toward the early stages of the ODE improves stability; however, \(\rho=1.45\) leads to performance degradation (PSNR drops from 26.03 to 25.80), showing that excessive bias sacrifices late-stage denoising refinement and degrades overall quality.

Table 3: Impact of Anchor Budget \(B_{\text{anc}}\) on Predicted Schedule for B=10 (FLUX.1)

\(B_{\text{anc}}\)	LPIPS ↓	PSNR ↑	SSIM ↑	ImageReward ↑
8	0.254	20.22	0.780	0.996
6	0.301	18.59	0.752	0.978
4	0.302	18.29	0.753	0.970

\(B_{\text{anc}}=8\) yields the best reconstruction quality. For \(B_{\text{anc}}=6\) or \(4\), LPIPS deteriorates from 0.254 to \(\sim 0.301-0.302\), as the sampling trajectory itself becomes unstable and fails to supply sufficient geometric priors for quantile interpolation; the close performance of 4 and 6 suggests that the trajectory quality under extremely low budgets is the bottleneck rather than the interpolation accuracy.

Regarding anchor search effectiveness: The Top-1 searched schedule reduces MeanCache's LPIPS from 0.338 to 0.261 on average (-25.04%), and the Top-3 still maintains a 24.26% improvement, confirming that end-to-end perceptual optimization effectively compensates for proxy objective bias. For search efficiency, convergence is reached within 50 trials on median, and the vast majority converge within 200 trials.

Key Findings¶

Anchor search is a key source of quality improvement: End-to-end LPIPS optimization remedies the systematic bias of additive proxy objectives at low NFE, lowering LPIPS of the Top-1 schedule by ~25% compared to MeanCache. This gain propagates to all intermediate budgets through OT interpolation.
OT interpolation effectively captures cross-budget regularities: Starting from \(B_{\text{anc}}=8\) and \(B_{\text{ref}}=20\), the interpolated and predicted intermediate schedules (\(B=15, 13, 10\)) consistently outperform independently solved graph-search methods across three models, verifying the geodesic hypothesis of scheduling evolution.
An optimal range exists for the anchor budget: \(B_{\text{anc}}=8\) is the optimal anchor for a 50-step sampling—the search space is sufficiently high yet tractable (\(\sim 5.36 \times 10^8\) candidates) and the trajectory quality is robust enough to support interpolation. If it is too low (\(\le 6\)), the trajectory itself is unstable, whereas if it is too high (close to \(B_{\text{ref}}\)), the gap difference between the two endpoints is insufficient to characterize the evolution trajectory.
Initialization strategy significantly affects anchor search quality: Anchor search with MeanCache initialization yields an average LPIPS of 0.257, outperforming uniform initialization (0.293) and random initialization (0.319). ECDF curves display that MeanCache initialization yields high-fidelity solutions over a larger portion of samples.

Highlights & Insights¶

Elegant and compelling core observation: The visualization of the gap profile in Fig. 1B directly shows the smooth evolution of optimal schedules across various budgets—this is the fulcrum of the paper, and the leap from empirical observation to OT modeling is natural and strongly backed by empirical evidence.
OT perspective elegantly sidesteps the curse of dimensionality: The cache scheduling search space peaks conceptually near \(B=25\) (around \(1.26 \times 10^{14}\) candidate paths). OTCache only performs a single search at a low search complexity anchor (\(B=8, \sim 5.36 \times 10^8\)), predicting all other budgets via closed-form interpolation, thus transforming an exponential search problem into an \(O(1)\) interpolation problem.
Elegant properties of 1D Wasserstein: The 1D Wasserstein geodesic is equivalent to a linear combination of quantile functions. OTCache leverages this mathematical property to directly take weighted linear averages of PCHIP warping curves, avoiding solving expensive OT coupling matrices in probability measure spaces. This technique is highly transferable to any problem needing "smooth morphing between two policies/configurations."
Log-confidence weighting offers clear physical intuition: \(w_{\text{conf}} = \log B\) grants higher trust to the high-budget reference since dense sampling maps the continuous structure of the ODE trajectory more faithfully. Compared to simple proximity weighting (relying solely on distance), the confidence prior enables the reference to contribute valuable structural information even in high-acceleration regimes.
Gap-space parameterization is a general optimization trick: Optimizing gaps rather than absolute time steps in black-box searches naturally guarantees temporal monotonicity and improves search terrain conditioning. This idea is directly applicable to any ordered sequence optimization problem (e.g., exit point selection in model cascades, keyframe sampling in video, etc.).

Limitations & Future Work¶

High offline calibration overhead on video models: In the worst-case scenario, Stage 2 anchor search takes 50.5 min/prompt on HunyuanVideo. Calibrating 50 prompts takes about 4.2 hours (8xH100). Although a one-time offline process, it presents a non-trivial adaptation barrier for new models or architectures. The authors' proposed multi-GPU data-parallel scheme (finishing 50 prompts in ~42 minutes using 8xH100) mitigates this but has not been thoroughly validated.
Linear geodesic assumption remains unchallenged: Quantile linear interpolation implies a linear geodesic assumption in the policy space. When the gap between reference and anchor budgets is large (e.g., \(B_{\text{ref}}=20\) vs \(B_{\text{anc}}=4\)), whether nonlinear effects exist—where some intermediate schedules deviate from linear interpolation—remains unvalidated in experiments (the degradation at \(B_{\text{anc}}=4\) might partially stem from this).
Only validated on DiT architectures: FLUX.1, Qwen-Image, and HunyuanVideo are all Transformer-based (DiT) architectures, leaving applicability to UNet architectures unknown. The global self-attention of DiTs might result in unique cache error propagation patterns, and the effectiveness of OT interpolation needs independent validation on UNets.
Hand-crafted, fixed parameter \(\rho\): The power-law warp parameter is determined via grid search and remains globally static, but the vector field fluctuation patterns of the early/late denoising trajectory may vary with prompt content, image resolution, or even random seeds. Learning an adaptive or prompt-conditioned \(\rho\) could yield further benefits.
Unexplored synergy with distillation/step-compression methods: OTCache is a training-free caching framework. Combining it with few-step distilled models (e.g., SDXL-Turbo, LCM)—on trajectories already compressed to 4-8 steps—to see whether exploitable cache redundancy still exists is a valuable but untouched direction.

vs MeanCache: MeanCache solves the graph shortest path independently for each budget and relies on additive proxy objectives, whereas OTCache models structural relationships across budgets, turning individual solving into a "two-point calibration + interpolated prediction" paradigm that bypasses systematic bias of proxy objectives in low NFE intervals. However, MeanCache serves as the Stage 1 reference source, and its reliability in high-budget intervals is the prerequisite for OTCache, making them complementary rather than mutually exclusive.
vs LeMiCa: LeMiCa abstracts video denoising as a DAG for global cache scheduling, still based on additive edge weight aggregation. OTCache's OT perspective can naturally extend to video generation by treating spatio-temporal scheduling under different frame budgets as geodesic evolution, avoiding independent temporal assumptions for every frame.
vs TeaCache/ToCa/DBCache: These threshold-based or structural redundancy-based methods decide whether to cache at each step independently, lacking global planning. The policy space modeling of OTCache can inspire upgrading these heuristics into a "pre-computed schedule + online execution" paradigm to offload expensive global search offline.
vs Distillation Methods: Distillation methods like LCM, PD, and SDXL-Turbo require substantial training data and pipelines but compress sampling steps down to 1-4. OTCache is training-free and plug-and-play, achieving 3-5x speedup on 50-step models. They represent orthogonal dimensions of acceleration—distillation reduces total steps, whereas OTCache eliminates redundant computation within given steps—and their joint use could produce a synergistic gain.

Rating¶

Novelty: ⭐⭐⭐⭐ Introducing optimal transport into cache scheduling modeling offers a fresh perspective; the core observation (smooth evolution across budgets) is solidly backed by empirical evidence and acts as the logical pivot of the paper. However, the basic components of Stage 1/2 (graph search, black-box optimization) are not inherently original, centering the contribution on the design of the Stage 3 OT interpolation framework.
Experimental Thoroughness: ⭐⭐⭐⭐⭐ Covers 3 mainstream architectures (T2Ix2 + T2V), more than 7 baselines, and multi-dimensional metrics (reconstruction LPIPS/SSIM/PSNR + perceptual ImageReward/CLIP/VBench + efficiency FLOPs/latency). Ablations thoroughly investigate \(\rho\), anchor budget, initialization, and search efficiency, backed up by search-space combinatorics analysis and Pareto frontiers in the appendix.
Writing Quality: ⭐⭐⭐⭐ Fig. 1's double subfigures precisely convey the core tension (A: failure of proxy objectives) and core observation (B: smooth gap evolution). The three-stage framework diagram combined with the search space analysis in the appendix provides outstanding structure. Occasionally, math-dense paragraphs could be more intuitive, and certain ablation analyses could go deeper (e.g., the underlying cause of \(B_{\text{anc}}=6/4\) degradation).
Value: ⭐⭐⭐⭐ Highly practical as a training-free and plug-and-play acceleration scheme. The OT perspective supplies a reusable methodology ("budget-conditioned policy interpolation" paradigm) for diffusion model acceleration, and engineering tricks like gap-space parameterization are directly transferable. Offline calibration overhead is the main barrier for actual deployment, but there is immense potential for joint application with orthogonal acceleration dimensions like distillation or quantization.