Subscribe to Google Cloud Tech → goo.gle/GoogleCloudTech

Thanks for the crisp explanation. Very helpful

good job on nice summary

hey so I have a question. in the diffusion model we are ultimately reversing the process of adding noise. so wont it retrieve the same previous image ?

@nanelikahya9949

9 ай бұрын

Yeah how does it get a novel image by reversing the noise addiction?

@jakobullmann7586

4 ай бұрын

From a linear algebraic point of view: It should be clear intuitively that “smearing out” is irreversible, but you can also argue mathematically, e.g. with the fact that the graph Laplacian of any undirected graph is a singular matrix (the graph Laplacian is the generator of a random walk). You train these models to undo tiny amounts of noise. The solution is not unique from a linear algebraic point of view, but out of all possible solutions, only few will look realistic, hence the model is forced to learn the real-world distribution of images. But if you start from an extremely noisy image, there are many possible originals that might have produced this noisy version, too much information is lost.

the awesome stuff! thanks so much for sharing it!

will be a next video ?

Alguien conoce de algún video similar en castellano?

❤

why is the whole video script showing on the screen in the first 2 seconds? 😂

Can you translate in hindi or add subtitles....please....?

Will be a good idea to offer image resize service integreted with your CDN instead of this nonsense...

Honestly it was quite hard to understand