Hey, Look, I'm an Artist Now © ML Image Generator
summary
Describes how modern text-to-image systems pair CLIP (trained on ~400 million image–text pairs) with generators such as GLIDE, VQ‑VAE+GAN hybrids and latent diffusion models (Stable Diffusion, DALL·E, Midjourney) to create and iteratively refine images (examples: axolotls, inpainting). Demos run on Hugging Face GPU notebooks using float16, showing image quality improves with more diffusion steps and guidance scale around 7–9; some results appear realistically photorealistic (possible for fake profile photos), and the talk highlights copyright ambiguities and racial bias in generated images.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:48Generative model theory
- 12:50GAN and CLIP explained
- 23:22Image generator demos
- 38:54Societal impact and biases
- 43:57Conclusion and Q&A
related talks
43mEncounters of the Third Kind With the Multimodal Generative AI Revolution
27mGenerative AI: What's Next?
53mGenerative AI: Overview and use cases
45mGenerative AI for Image and Video Production
30mHow to train a large diffusion model
22mGenerative AI & its Transformation of Business, Research, and Society
22mEfficient Pretraining of Text-to-Image Models
32mHow generative AI is Changing the way people interact with computers