Controllable Image-Guided Generation Via Dynamic Gaussian Spectral Modulation

Abstract

Diffusion models have achieved impressive results in image generation, but existing approaches often struggle with fine-grained control over the synthesis process, limiting their adaptability across different tasks. To address this issue, we introduce a novel diffusion framework that integrates adaptive Gaussian filtering into the denoising process, allowing dynamic modulation of structural and textural information. Furthermore, we design a Bidirectional Optimization Framework, which consists of two progressive phases: (1) Noise-to-Structure Optimization, ensuring global structural consistency through controlled spectral modulation, and (2) Structure-to-Texture Optimization, enhancing fine-grained details via gradient-based refinement. The proposed approach operates without additional training, supporting various image translation tasks, including cross-domain transformations and image to image translation. Extensive experiments on multiple datasets, including FFHQ and AFHQ, demonstrate that the proposed method achieves significant improvements over existing approaches, delivering superior generative quality and broader applicability in real-world scenarios.

Keywords: Controllable Diffusion Models, Gaussian Filtering, Bidirectional Optimization Framework