FreqPose: Frequency-Aware Diffusion with Fractional Gabor Filters and Global Pose-Semantic Alignment

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Generation

Background:

• Pose-guided person image generation faces challenges in preserving high-frequency texture details and semantic identity during pose transitions.
• Existing methods often struggle with blurriness and loss of detail in appearance transfer and consistency during pose changes.

Purpose of the Study:

• To propose a novel diffusion-based generative framework that addresses texture detail loss and semantic inconsistency in pose-guided person image generation.
• To enhance image detail fidelity and ensure consistent semantic identity across varying poses.

Main Methods:

• Developed a multi-level fractional-order Gabor frequency-aware network for accurate extraction and reconstruction of high-frequency texture features.
• Implemented a global semantic-pose alignment module using cross-modal attention for mapping pose features to appearance semantics.
• Integrated these modules within a diffusion-based generative framework.

Main Results:

• The proposed framework effectively reconstructs high-frequency texture details like hair strands and fabric wrinkles.
• Achieved superior performance in maintaining structural integrity and natural textures under complex pose variations and large-angle rotations.
• Demonstrated state-of-the-art results on DeepFashion and Market1501 datasets, validated by SSIM, FID, and perceptual quality metrics.

Conclusions:

• The proposed diffusion-based framework significantly enhances texture fidelity and semantic consistency in pose-guided person image generation.
• The integration of frequency awareness and global semantic alignment effectively overcomes limitations of previous approaches.
• The method shows strong potential for realistic and consistent human image synthesis across diverse poses.