SYNTHESIZING STYLE-SIMILAR RESIDENTIAL FACADE FROM SEMANTIC LABELING ACCORDING TO THE USER-PROVIDED EXAMPLE

Abstract

Example-guided facade synthesis aims to synthesize realistic facade images from semantic labels drawn by architects and example images of user preferences. The automated synthesis approach allows for the efficient generation of facade solutions that will facilitate effective communication between stakeholders and creative inspiration for architects. This study proposes a conditional generation adversarial network with style consistency to solve the problem of example-guided image synthesis. Specifically, the synthesis model is divided into two stages: first, the domain of the semantic label map is transferred to the domain of the realistic image using the pix2pixHD framework to ensure that the synthesized facade in the intermediate stage can be semantically consistent with the designed facade; Second, we use the Deep Photo Style Transfer (DPST) framework to faithfully move the implied features of the realistic facade image synthesized in the previous step to the domain of the provided example to ensure consistency of style. In summary, the proposed method can constrain the synthesis of new residential facades from the semantic labels and example styles. The synthesized residential facades can be consistent with the example styles provided by the client while matching the semantic labels of the facade created by the designer, producing satisfyingly realistic transitions in various cases.