3 ms·
Overall this post is a pretty fair take, but I'm not sure intuitive the "latent space" explanation is. I work with diffusion models full time and really appreci
by samsartor 3y ago
Overall this post is a pretty fair take, but I'm not sure intuitive the "latent space" explanation is. I work with diffusion models full time and really appreciate how easy they are to explain at an introductory level (assuming you don't get distracted by the mathematical derivations). But I don't feel like I've seen that many good explanations online. I made my own attempt in a presentation last year (https://youtu.be/c-eIa8QuB24?t=86 https://youtu.be/c-eIa8QuB24?t=86) but haven't had time to clean it up. I keep hoping I'll see a better series from a real communicator pop up in my feed but I'm not sure this is it.
- catgary 3y agoHonestly, as a mathematician-turned-ml-researcher, Song and Ermon’s mathematical derivations based on SDEs are what sold me on diffusion/score-based generative models. This (https://youtu.be/wMmqCMwuM2Q?si=fvujznGDHjKH2yUi https://youtu.be/wMmqCMwuM2Q?si=fvujznGDHjKH2yUi) is probably the best ML seminar talk I’ve ever seen.
- samsartor 3y agoOh definitely. I think that 2021 paper was what sold me as well! But I don't think it is very approachable to non-technical audiences, and there are a number of other was to interpret diffusion models which are pretty intuitive. That is also why I like the alpha-(de)blending paper from this year.