📝 SummarXiv

Abstract

Diffusion models have shown strong performance in conditional generation by progressively denoising Gaussian samples toward a target data distribution. This denoising process can be interpreted as a form of hill climbing in a learned representation space, where the model iteratively refines a sample toward regions of higher probability. However, this learned climbing often converges to local optima with plausible but suboptimal generations due to latent space complexity and suboptimal initialization. While prior efforts often strengthen guidance signals or introduce fixed exploration strategies to address this, they exhibit limited capacity to escape steep local maxima. In contrast, we propose Controlled Random Zigzag Sampling (Ctrl-Z Sampling), a novel sampling strategy that adaptively detects and escapes such traps through controlled exploration. In each diffusion step, we first identify potential local maxima using a reward model. Upon such detection, we inject noise and revert to a previous, noisier state to escape the current plateau. The reward model then evaluates candidate trajectories, accepting only those that offer improvement, otherwise scheming progressively deeper explorations when nearby alternatives fail. This controlled zigzag process allows dynamic alternation between forward refinement and backward exploration, enhancing both alignment and visual quality in the generated outputs. The proposed method is model-agnostic and also compatible with existing diffusion frameworks. Experimental results show that Ctrl-Z Sampling substantially improves generation quality while requiring only about 7.72 times the NFEs of the original.