Early AI art was low-resolution, domain-specific, and uncontrollable, but modern models like Stable Diffusion can now generate high-quality images from any prompt.
Diffusion models are trained on billions of image-text pairs to learn concepts, then generate new images by reversing a "noising" process based on a user's prompt.
AI image editing works by masking a specific region and applying the diffusion model's denoising process only to that area, preserving the surrounding context.
By creating a new "word" in the AI's vocabulary, you can train it to generate images of a specific person, object, or artistic style using a technique like Textual Inversion.
Instead of relying solely on text prompts, you can control the AI's output composition by providing visual inputs like sketches, pose data, or depth maps.
The discussion covers practical topics like video generation and open-source models, as well as complex issues such as ethical risks, watermarking, and copyright ownership of AI-generated content.
28:30Cheuk Ho
25:17Scott Hanselman
30:30Kai Mueller & Sebastian Luxem
30:24Markus Walker
26:25Stephan Gillich
32:26Krzysztof Cieślak
29:54Fabian Pottbäcker, Thomas Endres & Martin Foertsch
30:38Timo Zander
.png?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
Conrad Electronic SE
RE-INvent Retail GmbH
Virtual Identity AG
RE-INvent Retail GmbH