AGFF: Attention-Gated Feature Fusion for Multi-Pose Virtual Try-On

Image-based multi-pose virtual try-on tasks aim to synthesize a person wearing a garment in a desired posture. Current methods face three main challenges: i) natural warping of clothing images, ii) preserving identity information after pose transfer, and iii) reducing model complexity. While recent methods have made improvements, they struggle to trade off complexity and performance, limiting model generalization. To address these issues, we propose AGFF, a new image-based multi-pose virtual try-on network with <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</b>ttention-<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</b>ated <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</b>eature <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</b>usion (AGFF), which efficiently tries on garments in arbitrary poses with low complexity. First, we introduce a bi-directional feature-matching approach with feature warping to capture geometric matching information between the garment and human posture for complex posture alignment. Second, we propose an attention-gated feature fusion approach to preserve more identity information by suppressing irrelevant person features and enhancing salient ones. Additionally, our model integrates seamlessly into small-scale encoder-decoder architectures, further reducing complexity. Extensive experiments on popular benchmarks show that our method outperforms state-of-the-art approaches both qualitatively and quantitatively.