Exploring the Artistic Potential of Neural Style Transfer
Fallah, Nazanin (2024)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2024052415600
https://urn.fi/URN:NBN:fi:amk-2024052415600
Tiivistelmä
This thesis explores the creative potential of Neural Style Transfer (NST), an advanced deep-learning method for producing new visuals. Combining material from one image with the creative approach of another well-known painting or artwork.
PopekD commissioned a study to evaluate the efficiency of NST in connecting user-uploaded photographs with different creative styles. The study examined the possible advantages and disadvantages of employing pre-trained models for style transfer.
The thesis provides a theoretical overview of NST and its methods, focusing on CNNs like VGG19. Pre-trained models divide images into content and style components, combined through iterative optimization to produce stylistic results. We developed a user-centric online application using advanced methods like React.js, PyTorch, and pre-trained VGG19 models for creative experimentation. This service allows users to effortlessly upload photos, select selected creative styles and generate unique results. Surveys were conducted to get input on the deployed solution's effectiveness and creative potential.
The study demonstrated that NST can accurately mimic the visual aesthetics of great artworks using user-supplied pictures, democratizing artistic innovation. The study emphasizes issues with employing pre-trained models, including biases in training data, limited user control over stylistic components, and high complexity that restricts customization. The thesis suggests developing clear, adaptable, and user-centric models to enhance NST as a powerful tool for digital art discovery.
PopekD commissioned a study to evaluate the efficiency of NST in connecting user-uploaded photographs with different creative styles. The study examined the possible advantages and disadvantages of employing pre-trained models for style transfer.
The thesis provides a theoretical overview of NST and its methods, focusing on CNNs like VGG19. Pre-trained models divide images into content and style components, combined through iterative optimization to produce stylistic results. We developed a user-centric online application using advanced methods like React.js, PyTorch, and pre-trained VGG19 models for creative experimentation. This service allows users to effortlessly upload photos, select selected creative styles and generate unique results. Surveys were conducted to get input on the deployed solution's effectiveness and creative potential.
The study demonstrated that NST can accurately mimic the visual aesthetics of great artworks using user-supplied pictures, democratizing artistic innovation. The study emphasizes issues with employing pre-trained models, including biases in training data, limited user control over stylistic components, and high complexity that restricts customization. The thesis suggests developing clear, adaptable, and user-centric models to enhance NST as a powerful tool for digital art discovery.