Region-guided Latent Diffusion for Training-free Style Transfer

Training-free diffusion methods show promise in style transfer but often suffer from spatial distortion and style-content leakage due to the lack of explicit structure guidance and spatial control. To address these limitations, we propose a region-guided dual-path diffusion framework that incorporates multi-level structural and positional priors to improve content preservation and controllability. First, we introduce a structural and positional image manipulation (SPIM) module that generates control maps via ControlNet for unified resizing, structure alignment, and mask-based region control. To better balance regional stylization and texture fidelity, we design a latent feature enhancement and normalization module (LaFEN) that adaptively injects high-frequency style residuals into the content latent space. Furthermore, we propose a region-aware and step-adaptive QKV fusion strategy to enhance the flexibility and precision of feature mixing across the denoising process. Extensive experiments show that our method consistently achieves high-quality stylization and structural fidelity in the style transfer stage without any additional training.