Flow With What You Know
Blogpost Url: https://d2jud02ci9yv69.cloudfront.net/2025-04-28-flow-with-what-you-know-38/blog/flow-with-what-you-know/
Abstract: We provide an accessible introduction to flow-matching and rectified flow models, which are increasingly at the forefront of generative AI applications. Typical descriptions of them are often laden with extensive probability-math equations, which can form barriers to the dissemination and understanding of these models. Fortunately, before they were couched in probabilities, the mechanisms underlying these models were grounded in basic physics, which provides an alternative and highly accessible (yet functionally equivalent) representation of the processes involved.
Conflict Of Interest: None that I'm aware of. But in the interest of full disclosure / TMI, perhaps there are 4 relevant points:
1. Some of Esser et al's authors and I overlapped while we worked at Stability a couple years ago, but were on different teams, never worked together, and have all been at other companies for over a year. My referring to their "success" is pretty objective: they won Best Paper at ICML 2024 and closed $31 million in Series Seed a week later. I do wish them well, but we're not pals.
2. Tanishq Abraham (whose "tweet" is cited) and I know each other a bit from years ago doing the Fast.AI course, and were both at Stability -- again, different teams, no collaboration. He's a well-known "ML person of interest" with 62K followers on X.com and does not need my citation.
3. I mention Katherine Crowson's k-diffusion GitHub package. She & I interacted a bit in 2021 on the Eleuther Discord server, overlapped at Stability and interacted a bit re. modifying her package for audio, but have had no interaction since early 2022. I also used her package and cited it in a preprint (https://arxiv.org/abs/2407.01499). But her package has over 2000 stars, and again her accomplishments are widely recognized in generative text-to-image and AI art circles (e.g. for VQGAN-CLIP, LAION-5B,..) and are independent of my acknowledgement.
4. Towards the end, I cite a recent paper on Hamiltonian flows by researchers at NVIDIA, and my startup (Hypesrtate AI) is part of the NVIDIA Inception program. However, I don't know those researchers at all; I just saw their paper/poster on the list for NeurIPS 2024.
Submission Number: 19
Loading