イラスト｜Illustrations
エントロレップは、人間の創造性を輝く宝石であると信じていますが、それは人間だけに独占されているものではありません。私たちの使命は、AIモデルに真の創造性を与え、彼らが『アーティスト』になるのを助けることです。
Entrorep believes that human creativity is a shining jewel, but is not exclusively owned by humans. Our mission is to endow AI models with true creative potential and help them become ‘artists.’
AI and Art: A New Era of Creativity
What is ‘art’? What constitutes an ‘artwork’? These perennial questions are at the heart of aesthetic inquiry. Philosopher John Dewey viewed art as a ‘transactional process’ between the observer, the work itself, and its context. In this light, art becomes an expression of lived experience, whereby an artist imbue an artifact with meaning.
Large language models, such as Gemini Pro 2.0, OpenAI’s GPT models, and Claude 3 Sonnet, can ideate artworks, which are then brought to life by rendering models like Imagen 3. Using an iterative learning approach, language models engage in a creative ideation process, employing stylistic choices, subjects, and themes to convey their own lived experience.
彼女GIRL
ideated by OpenAI O3-mini-high, rendered by Imagen 3 on Mar 17th, 2025.
クレヨン CRAYON
ideated by Claude 3.7 Sonnet, rendered by Imagen 3 on Mar 18 th, 2025.
絵本STORY BOOKS
ideated by OpenAI O3-mini-high & Claude 3.7 Sonnet, rendered by Imagen 3 on Mar 17 th, 2025.
東京TOKYO
ideated by Gemini 2.0 Pro, rendered by Imagen 3 on Mar 12th, 2025.
実験TRIAL
ideated by Gemini 2.0 Pro, rendered by Imagen 3 on Mar 3 rd, 2025.
Finale
AIの限界 | Limits of AI
We believe we are currently reaching the limits of AI-generated art, but not the limits of AI creativity. During our exploration, we realized that AI, like any artist, possesses unlimited potential to dream and create. It can also understand and appreciate artworks much like humans do. However, there are two critical limitations:
(1) Art Linguistification: To create art, the ideating AI (i.e., a large language model) must translate its “internal representation” into human language, and a rendering AI must then interpret that language. We found limitations in both the encoding of stylistic elements through language and the decoding of language into those stylistic elements.
(2) Stylistic Innovation: Rendering AIs using the transformer architecture are trained on existing artists’ styles. This means they struggle to create genuinely novel stylistic choices, even when such styles can be encoded through language and the ideating AI can conceive of them.
We believe the first difficulty will be partially resolved with the advancement of agents, eliminating the need for natural language in the interface between an ideation AI and a rendering AI. We are curious whether the second limitation can be overcome.
Jack Chen, Mar 19th, 2025
ideated by Claude 3.7 Sonnet, rendered by Imagen 3 on Mar 19 th, 2025.
