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

2026-06-29

AI Summary

Existing graph-based caching methods struggle to efficiently accelerate diffusion model sampling under low NFE (number of function evaluations) regimes due to the breakdown of additive independence assumptions. This work proposes the first training-free cache scheduling framework, introducing optimal transport theory into diffusion cache scheduling for the first time. By modeling the smooth evolution of caching policies across varying inference budgets in policy space, the method leverages reference policies, anchor search, and quantile interpolation to achieve geometry-aware, efficient scheduling. This approach overcomes the independence limitations of conventional graph-based methods, achieving acceleration factors of 4.5×, 4.7×, and 3.66× on FLUX.1, Qwen-Image, and HunyuanVideo, respectively, while simultaneously improving generation fidelity and significantly outperforming state-of-the-art caching techniques.

Abstract

We propose OTCache, a training-free framework for accelerating diffusion sampling via caching schedule prediction. Existing graph-based caching methods reduce redundant computation by optimizing shortest-path objectives, but rely on an additive independence assumption, which often breaks down in the low NFE regime. To address this issue, OTCache models caching schedules across inference budgets as a smooth evolution in policy space, inspired by Optimal Transport (OT). The framework consists of three stages: (1) obtaining a high-fidelity \textbf{reference schedule} using a graph-based caching method under a conservative budget; (2) performing a lightweight anchor search under an extreme low-budget setting via Optuna optimization with an end-to-end perceptual objective; and (3) predicting schedules for target budgets via quantile interpolation between the reference and anchor policies using continuous warping representations. Experiments on FLUX.1 [dev], Qwen-Image, and HunyuanVideo show that OTCache achieves 4.5x, 4.7x, and 3.66x acceleration, respectively, while consistently improving generation fidelity over state-of-the-art caching baselines. This work provides a new perspective on accelerating diffusion models through Optimal-Transport-inspired schedule modeling. Code:https://github.com/UnicomAI/OTCache

Problem

• diffusion models
• caching
• sampling acceleration
• Optimal Transport
• NFE

Innovation

• Optimal Transport
• Caching Schedule
• Diffusion Sampling Acceleration
• Geometry-Aware Caching
• Quantile Interpolation