In this paper, we propose an adaptive inference time scaling method that dynamically adjusts the computational load during the inference process to solve the inference time scaling problem of diffusion models. Unlike the existing methods that rely on a fixed noise removal schedule, we propose a new framework called Adaptive Bi-directional Cyclic Diffusion (ABCD), which improves the output through bidirectional diffusion cycles and adaptively controls the search depth and termination time. ABCD consists of three components: cyclic diffusion search, automatic exploration-exploitation balance, and adaptive thinking time. We experimentally demonstrate that it improves performance while maintaining computational efficiency on a variety of tasks.
