Diffusion model

Speaking of discovery, let's talk about the technical side.

How does Emot explain diffusion models versus the standard transformers?

He contrasts autoregressive transformers with diffusion models by explaining that diffusion works by destroying data with noise, then learning to reconstruct it.

He sees that process as tied to compression, least action, and even how nature itself seems to minimize effort while preserving structure.

And stable diffusion itself was meant to democratize this technology, right? Exactly.

Fresh for everyone.

DALI2 actually generates the image using a process called diffusion, which has been described by the company as starting with a bag of dots and then filling in a pattern with greater and greater detail.

Diffusion is the hottest new method of AI generation and can be a whole other topic for another day.

But in order to save time, let's move on to something that I find a lot more interesting.

The mimicking of human preference.

Today, we're diving into a paper from the Hugging Face Daily Paper list of August 18, 2026.

It has garnered 26 upvotes and is titled An Empirical Study of Training Pixel Space Text-to-Image Diffusion Models.

The first two authors are Deng Yangjiang and Ruoyidu, with the corresponding authors being Peng Gao and Harry Yang.

They come from a collaboration between Alibaba Token Hub of Alibaba Group and the Hong Kong University of Science and Technology.

So, Ashley, let's dive into the introduction.

What's the main focus of this paper?

The primary focus is on pixel space diffusion models for generative tasks, specifically text-to-image synthesis.

This is an area that has seen significant interest but hasn't been as thoroughly explored as its latent space counterpart.