Making Art with GANs : Week 4 Datasets


Datasets are most of the creativity in ML Art. Datasets are the hardest part in Machine Learning Art.Pre trained models are limited, as they are mostly looking for results and use low res images.

Premade datasets are biased
• all datasets are made by humans
• most training datasets are designed to be very broad, often avoid minority considerations
(celebrities, creative commons)
• almost all photograph-based
• often for business purposes
  (facial recognition, self-driving cars)
• often limited resolutions
• almost none are made by artists

NVIDIA makers of StyleGan , images usually are copyright abiding


Dataset Considerations


  • Domain, StyleGAN works best within a single domain of images (faces, trees, cars, cats).


    Other models may work better for multiple domains.

  • Diversity, There’s a “Goldilocks” range of diversity to your images. You don’t want all your images to be too similar (what’s the point) and you don’t want them to be too different (output will be “soupy”)


  • Alignment , FFHQ (the model StyleGAN is most known for) works particularly well because some of the facial features are aligned. This alignment makes the animations look really impressive.




Popular posts from this blog

Socially Engaged Art and Digital Practice week 9 - Project 2

Socially Engaged Art and Digital Practice : week 5 Materials Presentation

Socially Engaged Art and Digital Practice : week 3