Inspirational Adversarial Image Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep generative models, such as Generative Adversarial Networks (GANs), offer creative potential but lack user-friendly control for artists and designers.
• Existing tools for exploiting GANs are often tedious, hindering the integration of these models into creative workflows.

Purpose of the Study:

• To develop a simple strategy for controlling deep generative models to inspire creators with novel image generations.
• To provide users with control over the output of generative models based on their chosen datasets and preferences.

Main Methods:

• A straightforward optimization method was designed to identify optimal latent parameters for generating images closely matching user-provided inspirational images.
• The method involves iterative optimization steps within the model's latent space to recover desired parameters.
• Exploration techniques, including gradient descent and gradient-free optimizers, were tested, with gradient-free methods leveraging human preferences for optimization.

Main Results:

• The proposed optimization strategy effectively retrieves satisfactory images across diverse datasets, including faces, fashion, and textures.
• Gradient-free optimizers demonstrated robustness by utilizing human preferences, enabling applications like facial composite and fashion generation algorithms.
• The approach successfully generated novel content inspired by user input and preferences.

Conclusions:

• The developed method offers a practical and controllable approach to leveraging deep generative models for creative image generation.
• Iterative refinement based on user preferences enhances the robustness and applicability of generative models in artistic and design contexts.
• This strategy significantly simplifies the process of obtaining desired outputs from generative models, making them more accessible to creators.