A face template: Improving the face generation quality of multi-stage generative adversarial networks using coarse-grained facial priors

The current text-to-face synthesis models only utilize text descriptions for image synthesis, neglecting the prior information of basic facial features. This leads to insufficient learning of both coarse-grained facial features, such as face shape and positions of the basic organs, and fine-grained facial features, such as facial wrinkles and hair textures, by the generators in each stage. As a result, the quality of the generated faces is low. To address this issue, we propose a generic face template. It includes only common facial information, such as face contour, shapes and relative positions of the basic organs. Moreover, to embed the face template into the first-stage generator of three stages for assisting face generation, we design a Facial Coarse-grained Feature Excitation Module (FCFEM). FCFEM extracts the coarse-grained feature channel weights of the face template. And it performs channel recalibration on the intermediate feature maps of the first-stage generator. This helps to generate more precise and complete initial face images. Therefore, it can enhance the ability of the generators in the latter two stages to learn fine-grained features. Experiments on the Multi-Modal CelebA-HQ dataset demonstrate that the multi-stage models using our method generate face images with higher quality and realism compared to the original models. They also achieve higher semantic consistency between generated images and text descriptions.