Different words for different parts of the frame

The problem it solves

Two people, one in a blue coat and one in a red scarf. The model swaps them. You emphasise, reorder, add weights. It still swaps them, and the previous block explained why: attribute binding degrades, especially across chunk boundaries and especially between similar subjects.

Regional prompting removes the ambiguity by removing the need to bind. Instead of one prompt describing two people, you give the left region one prompt and the right region another. There is nothing to confuse, because the descriptions never meet.

How it works

Implementations differ (ComfyUI has conditioning-area and mask nodes, A1111 and Forge have regional prompter extensions) but the mechanism is the same everywhere. The image is divided into regions, either by simple splits or by painted masks. Each region gets its own conditioning. During sampling, each region’s prediction is computed with its own text embedding and composited according to the masks.

A simple setup:

base prompt (whole image, low weight)
  a documentary photograph, two people in a market, hard sunlight

region left (0.0-0.5 width)
  an older man in a blue wool coat, grey beard

region right (0.5-1.0 width)
  a young woman in a red silk scarf, laughing

The coat stays blue. The scarf stays red. No emphasis syntax, no seed lottery.

What it is good for

Two or more subjects with distinct attributes. The canonical case.

Sky and ground treated separately. A dramatic sky over a specific landscape, each described properly, which one prompt handles poorly because the words compete.

A specific object in a specific place. “Red kettle on the left of the counter” as a region rather than a hope.

Keeping a character’s description off the background. If your prompt says “woman in a green sari” and the background acquires a green tint, that is the description bleeding into the whole frame. Regions confine it.

Mixing styles deliberately. A photographic subject against an illustrated background, which is otherwise very hard.

The settings that decide whether it works

Base prompt weight. There is usually a global prompt applied across the whole image, controlling overall style and light. Keep it descriptive of the scene and let regions describe the contents. If the base is too strong it overrides regions; too weak and the image loses coherence. Around 0.2 to 0.4 of the total weight is a reasonable start.

Mask feathering. Hard region boundaries produce a visible seam: a straight vertical line where the style changes. Feather the masks by 30 to 60 pixels and the transition disappears.

Overlap. Regions that touch but do not overlap can leave a strip that belongs to neither. A small overlap is safer than a gap.

What it does not fix

It does not create objects. A region saying “a red kettle” on an empty counter produces a red kettle only if the composition already supports one there. Regional conditioning biases what appears in an area; it does not place objects. If the base composition has a window there, you will get a reddish window.

It does not handle interaction. Two people shaking hands cannot be split into two regions, because the hands are in both. Anything where subjects touch or overlap needs a different approach: generate them separately and composite, or accept a repair pass.

It struggles with irregular shapes. Rectangular splits are reliable. A complex painted mask around a figure that has not been generated yet is guessing at a shape that does not exist, and the result is often worse than not masking at all.

When to use something else

Regional prompting is fiddly to set up, and there is usually a simpler route. Before reaching for it, consider two alternatives.

Inpainting. Generate the scene, then mask the coat and regenerate it blue. Slower per image, far more reliable, and it works on any interface. For one image, this is usually the better answer.

Generate separately and composite. Two clean subjects on plain backgrounds, cut out and assembled over a generated environment. More work, complete control, and the standard approach for anything a client will scrutinise.

Regional prompting earns its complexity on sets: when you need twenty images with the same two-subject structure and cannot afford twenty inpainting sessions. For a single hero image, set it up only if you already know the tool; otherwise inpaint.