Adaptive Complex Region Perception for Pose-Guided Image Generation

Pose-guided image generation aims to generate realistic images based on a reference appearance and target pose. However, existing methods often suffer from structural distortion and texture degradation in visually complex regions, such as intricate clothing textures, text, sentences, skin tattoos, and pose-interactive regions. We argue that these shortcomings stem from two fundamental limitations. First, existing methods lack precise perception of complex regions. Second, feature allocation using attention mechanisms remains too coarse-grained to effectively model adaptively extracted complex regions. To address this issue, we propose a novel framework called “ “Adaptive Complex Region Perception for Pose-Guided Image Generation” (ACRP-Gen), which improves overall image quality by adaptively locating “complex regions” in the target image. ACRP-Gen consists of two key components: the Adaptive Complex Region Perception Module (ACRP) and the Region-Aware Pose-Conditioned Generator (RPCG). First, ACRP adaptively identifies and locates visually complex regions crucial to generation quality through high-level visual understanding of the input image. Based on prior information about these regions, we design a Region-Aware Pose Conditional Generator (RPCG). This module achieves a unification of local details and global structure by feature modulation of the perceived complex regions. Furthermore, during training, we incorporate the consistency between the generated pose and its corresponding pose in the reference image into the loss function to improve generation quality. To validate the model’s effectiveness, we first conducted extensive experiments on the Deep Fashion dataset, and then further verified its robustness and generalization ability by evaluating it on the UBC dataset. Experimental results show that ACRP-Gen outperforms existing methods, especially in challenging scenarios with diverse materials and complex scenes.