Trajectory-Consistent Calibration for Cache-Accelerated Diffusion Models

2026-05-24

AI Summary

Caching-based acceleration of diffusion models introduces bias when reusing intermediate representations, degrading generation quality. This work proposes Trajectory-Consistent Calibration (TCC), the first method to incorporate trajectory consistency into cache calibration design. TCC calibrates cached representations through an offline iterative procedure without requiring additional training, simultaneously mitigating both direct bias and trajectory deviations induced by calibration. The approach effectively alleviates cumulative errors caused by cache reuse, significantly improving generation quality on PixArt-alpha and DiT-XL/2. When combined with FORA, TCC reduces the FID of PixArt-alpha from 29.83 to 27.35, slightly outperforming the full-computation baseline.

Abstract

Diffusion Transformers require repeated denoiser evaluations during iterative sampling, making inference computationally expensive. Cache-based acceleration reduces this cost by reusing intermediate representations across denoising steps, but can introduce representation deviations and degrade generation quality. In this paper, we analyze these deviations and show that effective calibration should consider both the direct mismatch caused by reuse and the subsequent trajectory shift induced by earlier corrections. To address this challenge, we propose Trajectory-Consistent Calibration (TCC), a training-free method that calibrates cached representations toward their full-computation counterparts. Specifically, rather than estimating all calibration priors from a single uncorrected cache trajectory, TCC uses an offline iterative procedure so that each prior accounts for the trajectory shift induced by preceding calibrations. Experiments on PixArt-alpha and DiT-XL/2 show that TCC consistently improves FID across representative cache-based acceleration methods while preserving their underlying reuse policies. Notably, in a representative PixArt-alpha cache-acceleration setting based on FORA, TCC reduces FID from 29.83 to 27.35, slightly surpassing the full-computation baseline.

Problem

• diffusion models
• cache acceleration
• trajectory shift
• representation deviation
• calibration

Innovation

• Trajectory-Consistent Calibration
• Cache-Accelerated Diffusion
• Representation Calibration
• Diffusion Transformers
• Iterative Sampling