AI models like OpenAI's DALL-E 2 can now generate photorealistic and culturally specific images directly from natural language prompts.
GANs use a two-player system where a generator creates fake images and a discriminator judges them against real ones, forcing both to improve.
The training process is a min-max game governed by a value function, where the discriminator maximizes accuracy and the generator minimizes it by creating better fakes.
Mode collapse, where the generator produces limited variety, can be fixed by introducing a similarity check that penalizes a lack of diversity in outputs.
Address training deadlocks with a two-timescale update rule and solve the vanishing gradient problem by replacing sigmoid activation functions with ReLU.
Progressive GANs achieve high-resolution results by starting with a low-resolution image and gradually fading in new layers to increase detail during training.
GauGAN allows users to control image generation by providing a segmentation map for layout and a style image to set the overall mood and color palette.
The Q&A session explores the societal impact of AI-generated images, including deepfake detection, AI safety, AI-powered editing, and legal ownership.
46:42Thomas Endres & Martin Förtsch & Jonas Mayer
30:02Martin Förtsch & Thomas Endres
56:55Ekaterina Sirazitdinova
29:27David Estevez
29:54Fabian Pottbäcker, Thomas Endres & Martin Foertsch
29:26Alexander Birke & Silke Eggert
25:17Scott Hanselman
28:30Cheuk Ho
.png?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
DTS Systeme GmbH
Deutsches Medizinrechenzentrum GmbH
Conrad Electronic SE
RE-INvent Retail GmbH