Title: SJD-PAC: Accelerating Speculative Jacobi Decoding via Proactive Drafting and Adaptive Continuation

Abstract: Speculative Jacobi Decoding (SJD) provides a method to speed up autoregressive text-to-image synthesis that does not rely on a draft model. Nevertheless, the high entropy inherent in visual generation often results in low acceptance rates for draft tokens within complex areas, establishing a bottleneck that drastically restricts overall throughput. To address this limitation, we propose SJD-PAC, an improved version of the SJD framework. Our approach incorporates a proactive drafting strategy designed to enhance local acceptance rates in these difficult, high-entropy regions. Additionally, we present an adaptive continuation mechanism that maintains sequence validation even after an initial rejection, thereby eliminating the necessity for complete resampling. By combining these two optimizations, we significantly extend the average acceptance length per step, which enhances inference speed while strictly maintaining the target distribution. Evaluations on standard text-to-image benchmarks reveal that SJD-PAC delivers a $3.8\times$ acceleration with no loss in image quality. The code is accessible at https://github.com/KangJialiang/SJD-PAC.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC