The Principles of Diffusion Models

Posted by Anon84 11/9/2025

The Principles of Diffusion Models(arxiv.org)

239 points | 31 comments

smokel 11/9/2025|

If you're more into videos, be sure to check out Stefano Ermon's CS236 Deep Generative Models [1]. All lectures are available on YouTube [2].

[1] https://deepgenerativemodels.github.io/

[2] https://m.youtube.com/playlist?list=PLoROMvodv4rPOWA-omMM6ST...

storus 11/10/2025|

I wish Stanford kept offering CS236 but they haven't run it for two years already :(

dvrp 11/9/2025||

hn question: how is this not a dupe of my days old submission (https://news.ycombinator.com/item?id=45743810) ?

borski 11/9/2025|

It is, but dupes are allowed in some cases:

“Are reposts ok?

If a story has not had significant attention in the last year or so, a small number of reposts is ok. Otherwise we bury reposts as duplicates.”

https://news.ycombinator.com/newsfaq.html

Also, from the guidelines: “Please don't post on HN to ask or tell us something. Send it to hn@ycombinator.com.”

dlcarrier 11/10/2025|||

I presume that email address is for when you want to ask something of Hacker New, not to ask something about Hacker News.

For example they probably didn't want posts like "Hey Hacker News, why don't you call for the revival of emacs and the elimination of all vi users?" and would rather you email them so they can ignore it, but they also don't want email messages asking "How do I italicize text in a Hacker News comments, seriously I can't remember and I would have done so earlier in this comment if I could?" and would rather you ask the community who could answer it without bothering anyone working at Y Combinator.

fragmede 11/10/2025||

Are you saying this based on experience or are you projecting? In my experience (tho not asking how to italicize text using * characters) Dang and tomhow are happy to answer all sorts of questions. Sometimes they do get bogged down by the reality of running a site of this site manually, as it were, but I can't remember a question that didn't eventually get answered. I'll even tell them I vouched for this bunch of dead comments, was that the right thing to do? And one of them will write back saying mostly, but just fyi comment xyz was more flamebaity than idea, but thank you for asking and working on calibrating your vouch-o-meter.

stathibus 11/10/2025|||

in other words - "it is lol, also go pound sand"

bondarchuk 11/10/2025|||

What's the problem? Someone submitted it for people to read but it didn't catch on, now it's resubmitted and people can read it after all. Everyone happy. Don't be so attached to imaginary internet points.

borski 11/10/2025|||

That's not what I said, but okay.

JustFinishedBSG 11/9/2025||

CTRL-F: "Fokker-Planck"

> 97 matches

Ok I'll read it :)

joaquincabezas 11/9/2025|

why am I only getting 26 matches? where's the threshold then? :D

tim333 11/9/2025||

It's all about the en dashes and Fokker-Planck vs Fokker–Planck.

dlcarrier 11/10/2025|||

PDF files often break up sentences in ways that the find utility can't follow, so even if they ask have the same dash, it might not find them all. At least those names are uncommon enough you could search for just one.

joaquincabezas 11/10/2025|||

AI is definitely related to dashes!!

gdmaher 11/11/2025||

Cool (but long) text. I wanted an overview so I used claude to make a chapter-by-chapter summary, sharing in case anyone else finds it useful

https://github.com/gmaher/diffusion_principles_summary

bondarchuk 11/10/2025||

Is there something equivalent in scope and comprehensiveness for transformers?

scatedbymath 11/10/2025||

i m scared by the maths

BrokenCogs 11/10/2025|

Are you sure you're not scated?

mlmonkey 11/9/2025||

470 pages?!?!?!? FML! :-D

leptons 11/9/2025|

Reading this reinforces that a lot of what makes up current "AI" is brute forcing and not actually intelligent or thoughtful. Although I suppose our meat-minds could also be brute-forcing everything throughout our entire lives, and consciousness is like a chat prompt sitting on top of the machinery of the mind. But artificial intelligence will always be just as soulless and unfulfilling as artificial flavors.

dhampi 11/9/2025||

Guessing you’re a physicist based on the name. You don’t think automatically doing RG flow in reverse has beauty to it?

There’s a lot of “force” in statistics, but that force relies on pretty deep structures and choices.

Bromeo 11/11/2025|||

Are you familiar with the "Bitter Lesson" by recent Turing Award winner Rich Sutton? http://www.incompleteideas.net/IncIdeas/BitterLesson.html

tim333 11/9/2025|||

Always is a long time. It may get better.

theptip 11/10/2025||

Intelligence is the manifold that these brute-force algorithms learn.

Of course we don’t brute-force this in our lifetime. Evolution encoded the coarse structure of the manifold over billions of years. And then encoded a hyper-compressed meta-learning algorithm into primates across millions of years.

uecker 11/10/2025||

Learning a manifold is not intelligence as it lacks the reasoning part.

esafak 11/10/2025||

Leaning the manifold is understanding. Reasoning, which takes place on the manifold, is applying that understanding.

uecker 11/10/2025||

I am not sure what you definition of "understanding" is that you apply here.

esafak 11/10/2025||

I mean understanding physics and the universe of natural possibilities; what can happen. Then comes why.

uecker 11/10/2025||

Fitting a manifold to a bunch of samples does not allow you to understand what can happen in the universe. For example, if you train a regular diffusion model on correct sudokus, it will produce sudokus with errors because it does not understand the rules.

esafak 11/11/2025||

You raise a good point for the diffusion case, which trains only on positive examples, but generally speaking negative examples will warp the manifold appropriately.

uecker 11/11/2025||

Sure, but you show a few correct examples to a human it will quickly pick up the correct rules. And this is understanding.