Note: This piece of news got overshadowed pretty quickly; the open-sourced, LAION-powered Stable Diffusion was release mid-August. The WikiArt dataset (see mention below) has 40,000 images, whereas the LAION has somewhere between 400 million and 5 billion (sorry I just can't tell anymore because I see both numbers in a bunch of different places).
A model to generate artistic images based on text descriptions
Jun 2022, phys.org
Dynamic memory generative adversarial network (DM-GAN) can automatically generate unique artistic images based on text descriptions."Due to the lack of datasets with paired text description and artistic images, it is hard to directly train an algorithm which can create art based on text input," the researchers explained in their paper."To address this issue, we split our task into three steps."First use the DM-GAN model to generate a realistic image that represents a text description.Subsequently use neural network ResNet to classify the image produced by the DM-GAN into one of the genre categories outlined by the WikiArt dataset, which contains more than 40,000 artistic paintings produced by 195 artists. Then use deep style transfer, and done.
via University of Waterloo and New York University Courant Institute: Qinghe Tian, Jean-Claude Franchitti, Text to artistic image generation. arXiv:2205.02439v1 [cs.CV],
Image credit: Tian & Franchitti
Post Script:
Your brain is better at busting deepfakes than you are
Jul 2022, phys.org
When looking at participants' brain activity, the University of Sydney researchers found deepfakes could be identified 54% of the time. However, when participants were asked to verbally identify the deepfakes, they could only do this 37% of the time."That tells us the brain can spot the difference between deepfakes and authentic images."
(Interpretation: In the future you will need an EEG just to communicate with your own brain.)
via University of Sydney: Michoel L. Moshel et al, Are you for real? Decoding realistic AI-generated faces from neural activity, Vision Research (2022). DOI: 10.1016/j.visres.2022.108079
No comments:
Post a Comment