This workshop, presented by the Metacreation Lab for Creative AI at Simon Fraser University, provides an introduction to generative AI and GANs, the neural networks used in Autolume [02:10].
Key topics covered:
Introduction to Generative AI & GANs: The workshop provides an overview of generative AI and details how GANs are used in their “Autolume” tool [02:33].
Autolume Demos & Training: Participants learn about real-time generation with Autolume and the process of training custom models using their own data [02:39]. The workshop covers what’s needed for training and offers advanced tips and tricks [02:47].
Small Data & Model Crafting: Philip Pasquier, a professor at Simon Fraser University, introduces concepts of “small data” and “model crafting” as a vision for ethical AI that is accessible to non-coders [04:08].
Ethical AI Considerations ⚖️: The discussion touches on the ethical implications of AI, particularly concerning data usage and copyright in creative domains. The Metacreation Lab emphasizes participatory design, involving stakeholders in tool development [06:48]. They mention ongoing lawsuits against AI companies for alleged data theft [20:37] and discuss solutions like using clean, proprietary, or public domain datasets, as well as models that offer retribution and attribution [22:40].
Autolume’s Design Principles 💻: Autolume is presented as a free, open-source tool that runs locally, requires no coding, and features built-in latent space navigation [31:25]. This allows artists and creators to control the generation process directly and in real-time, even with limited hardware (though a GPU is recommended for optimal performance) [30:05].
Video Training đź“ą: The workshop also covers using videos as a dataset for training models, discussing recommended video codecs and troubleshooting common issues with file formats [03:05:28]. Users can potentially find salient dimensions of change in the latent space, such as time, allowing for the generation of coherent temporal videos or controlling specific features like facial movements [03:03:08].