GCache Optimizes Diffusion Models for 2.17x Speedup, Better Quality

Diffusion models have become dominant in visual generation, but their inference process incurs substantial computational overhead. While cache-based acceleration methods exist, they often rely on local similarity heuristics, which researchers have identified as being poorly aligned with the overall quality of the final generated output. This misalignment stems from the non-uniform way errors propagate and accumulate during the denoising trajectory. To overcome this, a new approach called Global-Impact Cache (GCache) has been proposed. GCache first provides a rigorous theoretical characterization of the error propagation upper bound. Recognizing that this bound can be overly conservative for complex diffusion models, it then reparameterizes the propagation exponent using a Bernstein form and reframes the cache policy search as a bilevel optimization problem. This framework allows GCache to identify an optimal reuse policy in the inner objective, while simultaneously aligning the error-weighting function with the actual generation quality loss in the outer objective. This reconciliation of theoretical rigor with empirical performance enables GCache to prioritize computation where it has the most impact on visual fidelity. Experiments show GCache consistently outperforms prior caching strategies, achieving a 2.17x speedup on a state-of-the-art video diffusion model (Wan2.1) while significantly improving generation quality.