Style transfer methods usually use pre-trained VGG or more complex models as encoders to achieve better effects. This leads to extremely slow processing of high-resolution images. To solve the problem, we propose an degree-controllable detail attention-enhanced lightweight fast style transfer (DcDaeLFST), which adopts a small, shallow, and compact architecture for efficient forward inference. Additionally, our exploit a global semantic invariance loss to preserve the semantic and structural information of content images, and a local detail attention-enhanced module to preserve the detail information of them, together with a style discriminator. Despite limited parameters, it can achieve overall better style matching performance. Most importantly, it is the first method that can control the degree of detail retention and style transfer based on subjective evaluation. In comparative experiments, our model is 17-250 times smaller and 0.26-6.5 times faster than other state-of-the-art models, with the fastest processing speed of 0.38s on 4K high-resolution images.
View on arXiv