Intermediate11 min

References: Style, Character, and Image Prompts

Pure text prompting hits a ceiling. To control exactly what you get, you feed the model images as well as words. In 2026 every serious tool supports reference images for style, for a specific subject, and for composition. This lesson covers all three and when to use each.

Step 1: Style references copy the look, not the content

A style reference (Midjourney --sref, or a style image slot in Flux based tools) tells the model to borrow the palette, texture, and mood of a reference while generating your own subject. Use it to keep a whole campaign visually consistent.

midjourney prompt bar
$a city bus stop in the rain --sref https://example.com/my-style.jpg --sw 80
--sref the style source image
--sw 80 style weight, how strongly to apply it (0-1000)
$

Step 2: Character and subject references

To keep the same face or product across many images, use a character reference (Midjourney --cref) or an identity feature in tools like Flux and Runway. You supply a clear, well lit reference of the subject and the model carries its identity into new scenes.

character-ref.txt
the same young woman from the reference,
now sitting in a cafe reading a book,
warm afternoon light
--cref https://example.com/her-face.png --cw 100

# --cw 100 = lock face strongly; lower it to let clothing/style vary

Step 3: Composition control with image prompts

Sometimes you want the exact layout of a reference, where the horizon sits, where the subject stands. That is a structural or depth reference, often called ControlNet in the open world. You give the model the skeleton of an image and let it repaint the surface.

Reference slots in a Flux web app
[ Style ref ] influence: 60%
[ Character ref ] influence: 90%
[ Structure ref ] influence: 45% (pose / layout)
Prompt: a portrait in a sunlit kitchen
[ Generate ]
Modern apps expose three distinct slots so you can mix a style ref, a character ref, and a structure ref in one generation.
Do not crank every weight to max
If style, character, and structure are all at full strength they fight each other and the image collapses. Start each around 50 to 70 percent and raise only the one that matters most.

Result

You can now hold one variable steady (a face, a palette, a layout) while changing the rest, which is the core skill behind every consistent series and brand look.

Hands-on tasks