Carnegie Mellon University
Browse
PhD_dissertation_manuel_final (3)_R.pdf (204.87 MB)

Parametric-Based Models for Artistic Representations

Download (204.87 MB)
thesis
posted on 2024-03-08, 21:24 authored by Manuel Rodriguez Ladron de Guevara

  Driven by the popular adoption of AI for artistic purposes, this research examines its  technical and ethical implications, and presents new approaches for parametric algo rithms that generate different type of artistic representations, based on optimization  methods and learning models. In the midst of the success of machine learning algo rithms in image generation and other creative tasks of pixel-based diffusion and large  language models, my dissertation is framed within the field of parametric representa tions, which is closer to human art than pixel-based methods. An array of computa tional methods including procedural, optimization, and machine learning, are analyzed  and proposed with the intention of disseminating the generation of artwork by computers.  

To understand and situate the current AI models used for art, we provide a com prehensive review and implementation of artistic computational methods, ranging from  classical procedural hand-made, rule-based algorithms, to the most advanced AI meth ods. After analyzing the technical possibilities of existing methods, we present new  algorithms that focus on particular gaps in the literature.  

Particularly, this research studies three main problems that exist in the literature:  stylization, controllability, and identity preservation. Stylization, the process of apply ing a particular artistic style to an image, is a common subject within generative algo rithms, and key to artistic success. Controllability enables intentional painting processes  and thus, it is a steppingstone for further stylization. Identity preservation is the ability  for a learning model to preserve key content features from the original image through  the artistic representation process.  

 

Finding the right combination of stroke primitives for a particular artistic style and, by extension, for a painting strategy that leads to certain styles is not fully resolved under parametric framework, as there exists a trade-off between reconstructions of the input image and a controllable stylistic variation.  Existing works that address style are still limited in style variations and controllability, and principally use different stroke modes and textures to output styles.  State-of-the-art algorithms normally output strokes in an uncontrollable manner without a planned strategy that might help stylization. However, human artists employ painting techniques such as “blocking in”, grouping by semantics or colors, “background-foreground” or “color-then-contours” that help them convey artistic styles

Throughout the different algorithms presented, I demonstrate that we can achieve new ways to find stylization, controllability and identity preservation.  We disentangle such a complex landscape of artistic styles and strategies, and leverage some artistic vision under some perception of art. We finally tap into computational creativity, whether algorithms can be creative, and discuss future steps in the field of machine learning and art


History

Date

2024-01-27

Degree Type

  • Dissertation

Department

  • Architecture

Degree Name

  • Doctor of Philosophy (PhD)

Advisor(s)

Ramesh Krishnamurti Daragh Byrne Jun-Yan Zhu

Usage metrics

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC