Семинар HDI Lab: Diffusion Probabilistic Models explosion (part 2)
28 ноября 2022 г. в 18:00 состоится очередной семинар Международной лаборатории стохастических алгоритмов и анализа многомерных данных. С обзорным докладом "Diffusion Probabilistic Models explosion" выступит стажер-исследователь лаборатории Антон Зубехин.
Diffusion probabilistic models are state-of-the-art method for image generation. More than that, diffusion models also perform good working with data of different modalities, such as speech and text. Text-to-image diffusions are the most well-known types of such models with Dalle-2, Imagen and Stable Diffusion being famous even outside of ML community. Recent advances in machine learning show even wider applicability of diffusion models - they can be used for video generation, 3D reconstructions, meta-learning and even direct optimization of other neural networks.
Outline of the second part: current state of things, overview of recent papers; stable (latent) diffusion.