In order to improve the visual quality and style controllability of intangible cultural heritage art images, this paper proposes a generative adversarial implementation method for style transfer and image enhancement. A dataset consisting of 5240 images, including 4160 training images and 1080 testing images, is constructed from embroidery, paper cutouts, New Year paintings, and lacquer patterns. After denoising, brightness correction and label coding, the texture structure and color semantics are encoded and input into the CycleGAN framework under the constraints of PatchGAN, cycle consistency and structure-color joint constraints. The model was trained using Adam for 180 epochs with a batch size of 4 and an initial learning rate of 0.0002. Experimental results show that the PSNR value of the proposed method is 24.76 dB, the SSIM value is 0.842, and the style consistency evaluation value is 0.913. At the same time, the visual clarity and edge integrity of different heritage categories are improved. The framework provides an efficient computational solution for digital rendering, restoration, and cross-style generation of heritage images.
