Skip to content

Mixture-of-Attention (MoA)

Split personalised image generation across two branches

Visit official site

What MoA is

The architecture is described as a way to split personalised image generation between a personalisation branch and a non-personalised prior branch, with a routing mechanism distributing pixels across them.

What you can do with it

  • Generate personalised images of a subject
  • Keep the prior of the original model
  • Decouple a subject from its context
  • Compose several subjects in one image

Example uses

  • Swap a face in an uploaded photo for another
  • Draw a character with a set pose and expression
  • Change the initial noise while keeping the background

Pros & cons

✓ What we like

  • Multiple subjects can be composed and interact
  • The prior of the original model is preserved

! What to watch out for

  • Generation work is split across two branches
  • Pixel distribution is managed layer by layer

FAQ

What branches are used?

A personalisation branch and a non-personalised prior branch.

What does the routing do?

It manages pixel distribution between branches at each layer.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review