To address the difficulties in constructing an infrared image dataset from measurements and the high cost of testing and production， this study proposes a generation countermeasure VTIGAN for unpaired training samples to achieve high-quality image conversion from visible-to-infrared in different scenarios. VTIGAN introduces a new generator inspired by the transformer module to learn the mapping relationship of image content， whereby image style conversion is realized by reorganizing the characteristics of the target style. PathGAN is used as a discriminator to strengthen the image detail information generation ability of the model. Finally， four loss functions， namely， resistance， multi-layer contrast， style similarity， and identity losses are combined to constrain the model training process. VTIGAN was compared with other mainstream algorithms in a wide range of experiments on visible infrared datasets. Three evaluation indicators， namely， peak signal-to-noise ratio （PSNR）， structural similarity （SSIM）， and Fréchet inception distance （FID）， were used for quantitative and subjective qualitative evaluation. The experimental results show that VTIGAN improves PSNR， SSIM， and FID by 3.1%， 2.8%， and 11.3%， respectively， compared with the suboptimal UGATIT algorithm， which effectively realizes image conversion from visible light to infrared under the condition of unpaired training samples， demonstrating stronger anti-interference ability for complex scenes. The generated infrared images have high definition， complete details， and strong realism.