一字真跡仿真跡

Generative Font Stylistic Imitation Engine

設計師為了尋找有個性的字往往要花上不少的時間在碑文、拓印本及各名家法帖上，但不一定會找到古人曾經寫過的字及被數位化的字。例如「二南堂」三個公司標題字就是由不同的來源所組成。因此本研究以人工智慧的擴散運算為基礎，發展出可以從少數字元範例生成其他字，再從生成的圖形字轉成unicode向量字，更進一步快速完成一套完整的字體。本研究採用二種方式進行，第一種方法為擴散模型，透過迭代加噪及去噪來生成字體。第二種方法是以pix2pix為基礎的條件生成對抗網路的方法。結論是第一種方法，擴散加上加噪及去噪的方式更能保留字的筆劃細節，達到經由一字真跡後可由擴散加噪及去噪法仿真跡的境界。

In order to find unique characters, designers often spend a lot of time looking through inscriptions, rubbings and calligraphy works of famous calligraphers, but they may not necessarily find characters that were once written by the ancients or digitized. Therefore, this study is based on the diffusion operation of artificial intelligence and develops a method that can generate complete characters from a small number of letters, and then convert the generated graphic characters into unicode vector characters, and further quickly complete a font set. This study adopts two methods. The first method is a diffusion model, which generates fonts through iterative add noising and denoising. The second method is a conditional generative adversarial network method based on pix2pix. The conclusion is that the first method, diffusion plus noise addition and denoising, can better preserve the stroke details of the characters, achieving the state where the original character can be simulated by diffusion plus noise addition and denoising.