이미지 생성 인공지능의 가능성과 한계 - 한국 초상화 기반 생성사진 제작 사례 -
Possibilities and limitations of image generative artificial intelligence - a case of generative photography based on korean portraits -
0
조회
0
다운로드
3
인용
0
참고문헌
키워드
초록
This study examines the performance and limitations of image-generative artificial intelligence through actual generative photo production. The goal is to identify the possibilities and problems of generative artificial intelligence by generating images that are visually no different from photographs based on machine learning of Korean portraits. The images produced using generative artificial intelligence based on Generative Adversarial Networks not only maintained similarities with portraits, but also showed very realistic visual characteristics. However, although the generated image looks like a photograph, the Chosun's hat expressed in the portrait were not synthesized properly. In other words, generative AI has revealed its limitations in transforming and distorting elements that are not in the learning data. In addition, when the GAN value is set to 100, in short, in the case of fully automatic generation, all portraits showed convergence to the stereotype. It means that the characteristics of individual characters disappear. After all, the direction of generative artificial intelligence has been decided, and the ultimate result is the world of robots. To combat this contradiction and grim prospect, active human intervention is required.
Abstract
This study examines the performance and limitations of image-generative artificial intelligence through actual generative photo production. The goal is to identify the possibilities and problems of generative artificial intelligence by generating images that are visually no different from photographs based on machine learning of Korean portraits. The images produced using generative artificial intelligence based on Generative Adversarial Networks not only maintained similarities with portraits, but also showed very realistic visual characteristics. However, although the generated image looks like a photograph, the Chosun's hat expressed in the portrait were not synthesized properly. In other words, generative AI has revealed its limitations in transforming and distorting elements that are not in the learning data. In addition, when the GAN value is set to 100, in short, in the case of fully automatic generation, all portraits showed convergence to the stereotype. It means that the characteristics of individual characters disappear. After all, the direction of generative artificial intelligence has been decided, and the ultimate result is the world of robots. To combat this contradiction and grim prospect, active human intervention is required.
참고문헌 (18)
- 권행가, "고종 황제의 초상 : 근대 시각매체의 유입과 어진의 변용 과정", 박사, 한국근대미술사, pp.2006
- 이기주, "시각예술에서 이미지란 무엇인가" (2003), 눈빛
- 이태호, "사람을 사랑한 시대의 예술, 조선 후기 초상화" (2016), 마로니에북스
- 이태호, "조선 후기 회화의 사실정신" (1996), 학고재
- 조선미, "한국초상화 연구" (1983), 열화당
- 조선미, "석지 채용신" (2001), 삶과 꿈
- 조선미, "한국의 초상화 : 형과 영의 예술" (2009), 돌베개
- 조선미, "어진, 왕의 초상화" (2019), 한국학중앙연구원출판부
- 최인진, "한국사진사 1631-1945" (1999), 눈빛
- Clément Rosset, "L’Objet singulier" (1979), Minuit
- Paul Ricoeur, "Soi-même comme un autre" (1990), Seuil
- Edouard Pommier, "Théories du Portrait: de la Renaissance aux Lumières" (1998), Gallimard
- Susan Sontag, "On Photography" (1973), Picador
- Tero Karra, "A Style-Based Generator Architecture for Generative Adversarial Networks", ~ (2019)
- Brock, Andrew, "Large Scale GAN Training for Hige Fidelity Natural Image Synthesis", ~ (2019)
- Goodfellow, Ian, "Generative Adversarial Nets", ~ (2014)
- Karras, Tero, "Progressive Growing of GANs for improved quality, stability, and variation", ~ (2018)
- Ledig, Christian, "Photo-Ralistic single image Super-Resolution using a Generative Adversarial Network", ~ (2017)