Search papers, labs, and topics across Lattice.
This paper introduces Global-Impact Cache (GCache), a novel caching strategy designed to optimize cache reuse policies for diffusion models, addressing the inefficiencies caused by local similarity heuristics. By establishing a theoretical framework for error propagation and reformulating cache policy search as a bilevel optimization problem, GCache aligns computation prioritization with visual fidelity outcomes. Experimental results show that GCache achieves a 2.17x speedup on the Wan2.1 video diffusion model while significantly improving generation quality, as evidenced by a reduction in LPIPS from 0.1095 to 0.0316.
GCache achieves a remarkable 2.17x speedup in video diffusion while enhancing visual quality, challenging the effectiveness of traditional caching heuristics.
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.
