From Local Mismatch to Global Impact: Optimizing Cache Reuse Policy for Efficient Diffusion

2026-08-13

AI Summary

Existing caching strategies rely on local similarity heuristics, which struggle to balance acceleration efficiency and generation quality in diffusion models. This work proposes a novel bilevel optimization framework that explicitly links cache reuse policies with global generation quality for the first time. The inner loop searches for the optimal reuse strategy, while the outer loop aligns error weights with quality loss by leveraging an upper bound on error propagation derived via reparameterization in Bernstein form. Evaluated on the Wan2.1 video diffusion model, the method achieves a 2.17× speedup while significantly reducing the LPIPS metric from 0.1095 to 0.0316, substantially outperforming existing caching approaches.

Abstract

Diffusion models have achieved dominant performance in visual generation but suffer from substantial inference overhead. While cache-based acceleration has emerged as a promising solution, existing policies rely on local similarity heuristics, which we identify as being significantly misaligned with final generation quality. This discrepancy stems from the non-uniform propagation and accumulation of errors along the denoising trajectory. To address this, we propose Global-Impact Cache (GCache). We first establish a rigorous theoretical characterization of the error propagation upper bound. Recognizing that this bound can be overly conservative for complex, highly non-convex diffusion models, we further reparameterize the propagation exponent with a Bernstein form and reformulate cache policy search as a bilevel optimization problem. In detail, GCache identifies an optimal reuse policy in the inner objective while aligning the error-weighting function with generation quality loss in the outer objective. This framework effectively reconciles theoretical rigor with empirical performance, learning to prioritize computation where it most impacts visual fidelity. Extensive experiments demonstrate that GCache consistently outperforms prior caching strategies on both video and image generation. Notably, on the state-of-the-art Wan2.1 video diffusion model, GCache maintains a 2.17x speedup while significantly enhancing generation quality, reducing LPIPS from 0.1095 to 0.0316.

Problem

• diffusion models
• cache reuse policy
• error propagation
• generation quality
• inference acceleration

Innovation

• cache reuse policy
• error propagation
• bilevel optimization
• diffusion models
• generation quality