正在加载视频...
视频加载失败
When reading diffusion papers, my most common reaction is "mkay the math works out but WHY WOULD A SANE PERSON CHOOSE TO DO THIS". This included the (foundational) DDPM formulas. The saving grace is that you can visualize it as 2D particle motion and get a solid intuition.
20,604 次观看 • 8 个月前 •via X (Twitter)
17 条评论

I’ve only read one early diffusion paper, but using a statistical physics perspective made sense to me: you’re teaching a model how to reverse high entropy states (noise) to low entropy states (e.g cats). Is it useful to keep thinking about it like this or does it get much worse?

Things clicked for me when I tried to make the connection with physics, I have a video on this:

Are diffusion models making a comeback? Seeing lots of diffusion content lately but i'm out of the loop since nano-banana era autoregressive image gen

I don't think there's public information on how nano-banana really works. My assumption is that, even if there is some sort of auto-regressive planning involved, the actual pixels are still generated through diffusion/flow matching. IDK, either way, I'm doing this because I want to understand the fundamentals rather than whatever is sota right now.

papers don't really explain intuition - 3rd party and blogs and videos are really the best way to truly understand concepts.

In DDPM the particles got shrink-rayed and pulled towards the center to preserve variance and the math keeps them housekeeping, My imagination going wild sorry for wabbling !🙂

@MufassaM you're not imagining it, that's what the animation shows as well if you watch till the end :)

I watched it! :) Are you using Langevin Dynamics in your 2D sandbox? If I’m not mistaken,

@MufassaM This particular animation shows forward diffusion, Langevin is an inference/sampling process IIUC. But the common denominator between them is that they're stochastic and there's an element of Brownian motion in how the particles move.

Thanks for the clarification! It’s cool to see how that shared Brownian motion is the 'common denominator' for both sides of the process

This is amazing, thank you

what is it exactly that is becoming more intuitive after seeing this animation?

Looks like a pilot wave lol

Still p5.js ?

Visualizing DDPM as particles is genius

what site is this?

still working on it, might make it live this week


