70.4K
DADiffusion models create new data by learning how to undo a gradual process where noise is added step by step to an image or signal. This idea is inspired by **Brownian motion**, where particles move randomly as noise slowly increases until the original data becomes almost pure noise.
During training, the model studies how to reverse this process. A neural network learns the directions that move a noisy sample back toward clearer, more structured data. In other words, it predicts how the noise should be removed at each stage.
By following these learned directions through many small steps, the model can transform random noise into realistic outputs, which is how diffusion models generate things like images, videos, and other synthetic data.
Credits; Welch Las
Follow @datascience.swat for more daily videos like this
Shared under fair use for commentary and inspiration. No copyright infringement intended.
If you are the copyright holder and would prefer this removed, please DM me. I will take it down respectfully.
©️ All rights remain with the original creator (s)
@datascience.swat
