Definition
Plain language
A system that writes text the way image generators make pictures — refining a whole passage out of noise instead of typing it word by word.
As stated in the literature
A continuous latent diffusion language model that encodes tokens into vectors with a VAE, then runs an image-style Diffusion Transformer with flow matching over those latents; reaches GPT-2-level continuation quality from scratch, with the key ingredient being REPA-shaped latent geometry.
Why it matters: It shows that the image-generation recipe can produce competitive text from scratch, hinting at models that handle words and pictures with one architecture.
For example, instead of typing a paragraph word by word, it starts from noise and refines an entire passage at once into readable text.
Heard on the show
“The paper is called "TextLDM: Language Modeling with Continuous Latent Diffusion.”Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm