LitExplorer: Training-Free Diffusion Guidance with Adaptive Exploration-Filtering Framework

ICLR 2026 Conference Submission101 Authors

Keywords: Diffusion Model;Traning-free

Abstract: Diffusion models possess strong general generative capabilities, yet they remain insufficient when aligned with specific target objectives. Fine-tuning methods can enhance alignment but incur high training costs and face the risk of reward hacking. Consequently, training-free guidance mechanisms have emerged, which leverage external signals during inference to steer the generative distribution toward high-reward regions. However, existing training-free approaches encounter two key challenges: first, the guidance process tends to over-bias generation toward the target distribution, at the expense of excessively narrowing the pretrained model’s generative space; second, the guidance signals are mechanically imposed throughout inference, lacking mechanisms to identify and filter out ineffective or redundant signals. To mitigate these limitations, we propose \ourmethod{}. Regarding the first issue, we introduce exploratory guidance signals through \pos{} to prevent generation paths from prematurely converging to a single mode, while dynamically balancing the trade-off between exploration and stable generation based on denoising progress. This alleviates the excessive contraction of the generative space without deviating from the target distribution or the pretrained distribution. Regarding the second issue, to enable precise and efficient guidance, we incorporate an adjudication mechanism that evaluates the validity of guidance signals and adaptively eliminates ineffective or redundant ones. To demonstrate the generality of \ourmethod{}, we conduct extensive evaluations in both single-objective and multi-objective scenarios. Results show that \ourmethod{} achieves significant improvements over existing training-free baselines in terms of generative diversity, target alignment, and inference efficiency.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 101