I was tracing Jascha Sohl-Dickstein’s original paper [1503.03585] Deep Unsupervised Learning using Nonequilibrium Thermodynamics. The intuition seems to come from physics and statistical mechanics which is fantastic. I am just wondering in life diffusion seems quickly leads to chaos, as there are some theories including butterfly effect which seems pretty true, so the eventual/inference “timeline” of distribution should have high entropy and not Gaussian, so why diffusion is Gaussian? Please share your thoughts!
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Difference between origin EDM and diffusers EDM | 1 | 668 | December 4, 2023 | |
| A few questions about how (vanilla) diffusion works | 1 | 901 | September 25, 2022 | |
| Why the output of the UNet is noise? | 0 | 291 | November 14, 2023 | |
| Why is the loss of Diffusion model calculated between "RANDOM noise" and "model predicted noise"? Not between "Actual added noise" and "model predicted noise"? | 12 | 5577 | November 27, 2023 | |
| Why sigma_data = 0.5 in EDM? | 1 | 1131 | January 12, 2024 |