diffusion text lab / 01
Small model. Entirely local.

An experiment in discrete diffusion

Language, out of noise.

Ten hidden tokens. Eighteen small steps.
Watch a tiny transformer turn uncertainty into a sentence — right here in your browser.

01 / Live inference On-device model
Token canvasCompleted example · seed 100
the wandering tower brings secret blossoms across the cedars gently
Resolved token Masked token
Bidirectional attention · 10 positions

Start from pure noise. Let the model find the words.

34,998Learned parameters
278Token vocabulary
300Training sentences
0Runtime requests
02

20 model samples. Open any sentence to see it emerge.

03

From masks to meaning

01   Learn what is missing

Trained offline on 300 short, template-generated sentences. Random words are replaced with [MASK]; the transformer learns to predict the originals.

02   Look in both directions

One attention block, 32 embedding dimensions. Each masked position considers the whole sentence, including words that have already emerged.

03   Reveal the confident words

At each of 18 steps, the model scores the remaining masks. The most confident positions resolve first. Ten tokens emerge; some steps only reassess.

A tiny model, with tiny-model quirks. These are unconditional samples from a narrow, poetic training set. Expect repetitions, unusual combinations, and occasional grammatical slips. No prompt, no server, no sentence lookup during live generation.