A Zero-Shot High-Dimensional Feature Fusion With STF-GAN for Cross-Domain Image Reconstruction

Abstract

High-dimensional imaging techniques are rapidly evolving in the signal processing field, with applications spanning a wide range of domains such as autonomous driving, industrial manufacturing, healthcare, remote sensing, and robotics. Generative Adversarial Networks (GANs) have shown remarkable success in image generation, yet challenges persist in high-dimensional cross-domain reconstruction tasks—particularly when target-domain samples are unavailable. Existing zero-shot methods often suffer from expression distortion, mode collapse, and feature degradation due to inadequate fusion of multi-scale representations and poor semantic consistency. Inspired by the need for high-dimensional image recovery, we propose the Share Token Tensorized Attention Fusion Generative Adversarial Network (STF-GAN) framework. This is a new framework for multi-scale feature fusion of features from pre-trained generators using the Share Token Tensorized Attention Fusion approach. Our main contributions are the tensorized attention mechanism that fuses high-dimensional features across spatial and semantic domains while maintaining structural consistency, fast learning guided by CLIP, and bridging the source-target domain gap in the absence of paired data. Experimental results on FFHQ and AFHQ datasets demonstrate that STF-GAN achieves superior reconstruction fidelity, with an average improvement of 12.3% in Inception Score and 8.7% in Structural Consistency compared to IPL, while effectively mitigating distortion artifacts. These experiments demonstrate the ability of STF-GAN to process high-dimensional features, showing strong potential for applications in industrial imaging systems, medical diagnostics, and other fields where cross-domain image reconstruction must be performed with limited data availability.