Where mix-ups come from

Diffusion models connect words to image areas through **cross-attention**. At every step, every position of the
image "looks at" the prompt's words and takes in some of each. If "dog" and "cat" pull on the same positions, the
model draws a mixture: a cat-dog, a dog's head on a person, or colours that land on the wrong thing ("purple"
meant for the teapot ends up on the apron).

You can see this attention directly. The act-steer extension for Forge records, for every word, how much of
the attention each image area gave it, and shows it as a **word map** above the image:

- **all words**: every area is tinted in the colour of the word that dominates it. The coloured chips under the
image are the prompt's words in their colours; grey words are filler words.
- **one word** (hover or click a chip): a spotlight. The word's area lights up in its colour and the rest of the
image is darkened.


*Spotlight on the word "dog". Without the LoRA its attention sits on a head attached to a person; with the LoRA
it covers a dog of its own.*

From these maps the extension computes how much two words **overlap** (do they claim the same area?). Its
*Optimize* button already uses that to adjust word strengths for one prompt at a time. The idea of the LoRA:
**learn that fix once, inside the model, so it applies to every prompt.**

The idea

The LoRA changes only the two parts of cross-attention that decide *where* a word goes:

- the **query** projection (what an image position is looking for), and
- the **key** projection (how a word presents itself).

It does **not** touch the **value** and **output** projections, which decide *what* a word paints. So "red"
still means the same red and "watercolor" the same style; only the layout of which word acts where can change.
Training never sees pixels either: it only looks at attention maps, so it can't pick up a new style from its
training images.