I've been working with flow matching models for video generation for a while, and recently went back to my old notes from when I was first learning about them. I cleaned them up and turned them into this blog post.
Hopefully it’s useful for anyone exploring flow matching for generative modeling. Writing it certainly helped solidify my own understanding.
Yes, and other people I know don't have the same problem. Making me think that everyone who has this issue is part of some A/B testing [1] bucket for an internal Gmail spam filtering experiment.
These A/B tests are quite common for products like these to test new features, such as a new spam filtering algorithm. Or even test the efficiency of an existing spam filtering algorithm by degrading the experience of a subset of users.
Hopefully it’s useful for anyone exploring flow matching for generative modeling. Writing it certainly helped solidify my own understanding.