One character across every frame
Build a reference image first, then pick the right reference feature for your Midjourney version.
Start with an anchor image
A story short falls apart when the hero's face or jacket changes between shots. Before you write any scene prompts, generate one clean reference of your character: full body, plain background, even light, and an outfit with a few bold, simple features. Creators who do this well pick the result that shows the whole costume, because later angles have more to copy from.
Use the reference feature for your version
Midjourney has changed this feature several times, so check which model you are on. On V6 the tool is --cref with --cw: low weights keep mainly the face, the default of 100 also carries over hair and clothes. On V7 it is Omni Reference (--oref plus --ow, 1 to 1000, default 100). It takes one image, costs about twice the GPU time and does not run in Fast or Draft mode. On V8.1 and V8.2 the Edit model replaces both. You attach up to four reference images and can write instruction-style prompts such as "show this character from behind".
Let the text do its part
A reference does not replace the prompt. Describe the scene, action and lighting clearly, and repeat the physical traits you care about, especially when you lower the weight. If you want a different art style than the reference, name that style at the start and end of the prompt or add a style reference. High stylize or experimental values compete with the reference, so raise the weight to match, but on V7 stay under about 400 or results get erratic.
Know the limits
Small details such as freckles, tattoos and logos on clothing often drift. Design your character so they don't depend on them. Two people in one frame is the hardest case: the docs suggest putting both characters in a single reference image and describing each one in the prompt. Keep scene text short so the model has room to hold on to the face.