The Rise of Single-Step Generative Models
Автор: Jia-Bin Huang
Загружено: 2025-06-14
Просмотров: 10071
Diffusion and flow-matching models are key techniques for the current generative AI boom. However, their fundamental limitation is that they require multiple steps to generate samples. MeanFlow is a recent one-step end-to-end generative model that achieves compelling results without pretraining, distillation, or curriculum training.
In this video, we will discuss the fundamental concepts of Flow Matching and Mean Flow.
00:00 Introduction
00:24 Flow Matching
01:18 Conditional Flow Matching
02:46 Iterative sampling
04:06 MeanFlow
09:09 Results
Reference:
Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, and Kaiming He
Mean Flows for One-step Generative Modeling, arXiv 2025
https://arxiv.org/pdf/2505.13447
Video made with Manim: https://www.manim.community/
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
![How AI Taught Itself to See [DINOv3]](https://ricktube.ru/thumbnail/oGTasd3cliM/mqdefault.jpg)
![Как внимание стало настолько эффективным [GQA/MLA/DSA]](https://ricktube.ru/thumbnail/Y-o545eYjXM/mqdefault.jpg)
![[GCV @ CVPR25] Kaiming He - Towards End-to-End Generative Modeling](https://ricktube.ru/thumbnail/4VwXBrMoC0E/mqdefault.jpg)
![This Simple Optimizer Is Revolutionizing How We Train AI [Muon]](https://ricktube.ru/thumbnail/bO5nvE289ec/mqdefault.jpg)
![Что ошибочно пишут в книгах об ИИ [Двойной спуск]](https://ricktube.ru/thumbnail/z64a7USuGX0/mqdefault.jpg)
![How LLMs Learn to Reason [GRPO]](https://ricktube.ru/thumbnail/mg-iU-WxiNs/mqdefault.jpg)