How masked diffusion generates DNA

A generative model learns to reconstruct biological sequences by iteratively unmasking tokens. This demo visualises the process on a simple image. The same principle applies to DNA, where each pixel becomes a nucleotide (A, C, G, T).

Fully masked, t = 1.0
1.00
A
C
G
T
Masked

Forward process: progressively masks tokens until fully corrupted.
Reverse process: a neural network predicts and unmasks tokens iteratively, generating new sequences that follow the learned grammar.

Interactive companion to poster, Rotation Project 2, 2026