Loading video...
Video Failed to Load
I which Sander Dieleman beautifully illustrates why diffusion models work so well with images 🐳 Our visual world is spatially coherent, and large scale structures dominate. 〜 This means low frequency spectral components in images tend to be stronger. 🌫️ This means when you noise up an image, high... show more
16,429 views • 3 months ago •via X (Twitter)
2 Comments

Carlos Pinheiro3 months ago
@sedielem Indeed, very good explanation. 👍🏻

Alex Stern2 months ago
@sedielem Do you think this phenomenon of a fine to coarse data corruption schedule is necessary condition for a diffusion model to work well. Because afai understand, this condition isn't necessarily met with noise schedules used for DLM for example.
