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

Aravinthen 的头像
Aravinthen8 个月前

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?

Julia Turc 的头像
Julia Turc8 个月前

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

bat 的头像
bat8 个月前

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

Julia Turc 的头像
Julia Turc8 个月前

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.

Joshua Yang 的头像
Joshua Yang8 个月前

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

Murali Chandran 的头像
Murali Chandran8 个月前

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 !🙂

Julia Turc 的头像
Julia Turc8 个月前

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

Murali Chandran 的头像
Murali Chandran8 个月前

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

Julia Turc 的头像
Julia Turc8 个月前

@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.

Murali Chandran 的头像
Murali Chandran8 个月前

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

the normal form: Matios the Frobenius 的头像
the normal form: Matios the Frobenius8 个月前

This is amazing, thank you

bk 的头像
bk8 个月前

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

Toast X 的头像
Toast X8 个月前

Looks like a pilot wave lol

moskstraumen 的头像
moskstraumen8 个月前

Still p5.js ?

Kyriakos 的头像
Kyriakos8 个月前

Visualizing DDPM as particles is genius

Joshua Yang 的头像
Joshua Yang8 个月前

what site is this?

Julia Turc 的头像
Julia Turc8 个月前

still working on it, might make it live this week

相关视频

Brad Jacobs on how to identify A players: "I do a mental exercise where I picture the person coming into my office and saying, 'Brad, I quit.' And then I try to feel and visualize what would be my reaction if that person came in to me and quit. If my reaction to that is, 'Yes! I don't want to smile, so I don't want to act like I'm happy about this. Nobody likes firing people. No problem at all, we'll replace them'—that's a C player. That's someone you should get the courage to get off the team right away. On the second category, if my reaction to it is, 'You know, it kind of sucks. I would've preferred that person stayed, but it's not the end of the world. We'll hire a headhunter. We'll get someone as good, maybe someone even better, and things will work out'—that's a B player. But if when I visualize that person quitting, my reaction to that is pure terror and absolute panic, and like somebody took a baseball bat and just whacked me in the stomach and then punched me in the face. I'm going, 'Oh my God! Like I'm never going to find someone as good as her. No way. I'm never going to have someone as talented as that person. I'm never going to have someone who brings to the table their particular superpower.' And I can't even hear what they're saying anymore because I'm just having this internal panic dialogue going on—that's what you call an A player. So I want all A players around me. I want people whose relationship with me I value so much that if it was terminated, I would be lost."

David Senra

59,239 次观看 • 9 个月前

Brad Jacobs on identifying A players: "I do a mental exercise where I picture the person coming into my office and saying, 'Brad, I quit.' And then I try to feel and visualize what would be my reaction if that person came in to me and quit. If my reaction to that is, 'Yes! I don't want to smile, so I don't want to act like I'm happy about this. Nobody likes firing people. No problem at all, we'll replace them'—that's a C player. That's someone you should get the courage to get off the team right away. On the second category, if my reaction to it is, 'You know, it kind of sucks. I would've preferred that person stayed, but it's not the end of the world. We'll hire a headhunter. We'll get someone as good, maybe someone even better, and things will work out'—that's a B player. But if when I visualize that person quitting, my reaction to that is pure terror and absolute panic, and like somebody took a baseball bat and just whacked me in the stomach and then punched me in the face. I'm going, 'Oh my God! Like I'm never going to find someone as good as her. No way. I'm never going to have someone as talented as that person. I'm never going to have someone who brings to the table their particular superpower.' And I can't even hear what they're saying anymore because I'm just having this internal panic dialogue going on—that's what you call an A player. So I want all A players around me. I want people whose relationship with me I value so much that if it was terminated, I would be lost."

David Senra

691,258 次观看 • 1 年前