Title: GuidedBridge: Training-Free Enhancement of Bridge Models via Prior Guidance

Abstract:

While guidance techniques like classifier-free guidance (CFG) and auto-guidance (AG) have significantly propelled noise-to-data generation in diffusion models, bridge models offer a distinct advantage by employing a data-to-data generative process capable of leveraging clean, instructive priors. Building on the principle that quality disparities in denoising outputs can serve as effective guidance, we introduce Prior Guidance (PG), a novel method that requires no additional training. Our approach introduces a "weak prior"—a signal absent from the bridge model’s pre-training phase—which naturally impedes prior exploitation and results in suboptimal denoising. By contrasting this weak prior with the standard, seen prior, we employ a scaling factor to accentuate and strengthen the model's ability to utilize the prior information. Furthermore, we investigate the mechanics of prior exploitation within the bridge framework, leading to the development of Frequency-Modulated Prior Guidance (FMPG). This technique adapts the guidance scale across low- and high-frequency bands, aligning with the specific dynamics of bridge generation. For applications such as image in-painting, we propose a cascaded architecture, CFG-FMPG. This framework initially produces a noisy hidden representation using CFG, which is subsequently utilized as a generative prior by FMPG, thereby combining the strengths of both methods without sacrificing inference speed. Empirical results confirm that our PG strategies consistently boost the performance of pre-trained bridge models across a variety of image translation tasks.

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