MidJourney character art consistency is a filing problem before it is a prompting problem. Nearly everyone arrives at it the same way: you made a woman you loved, you saved the words that made her, and when you ran those words again she came back as somebody else. Same jaw? No. Same hair, roughly. Same person, not at all.

The words were never her. Language is a loose way to describe a face, and MidJourney reads your description fresh every single time and makes a new guess from it. "Deep brown skin, silver locs, round glasses" fits a large number of real people. The one you loved was one draw from that pool.

So stop saving prompts. Save characters.

What actually drifts, and why

A prompt has blanks in it. Every adjective you did not write is a decision the model makes on your behalf, and it makes a slightly different one each run. Nose width. Cheek shape. How old she reads. Whether the jumper is ribbed.

That drift is not random enough to look broken and not stable enough to look like one person, which is the worst of both. Your set ends up looking like a family rather than a woman.

You cannot close those gaps with more adjectives. Past about thirty words the description starts fighting itself, and you get an image that obeys the sentence and loses the face.

The character file

Four things, kept together in one place with her name on it. Miss one and she slides.

The reference image. The single generation where she looked most like herself, upscaled and saved to a folder, not left sitting in a scroll of results you will never find again.

What breaks without it: everything. Nothing else works without one picture to point at.

The character reference parameter. The image link you hand MidJourney so it carries her face and build into the next scene. On current models that is Omni Reference, --oref with the link and --ow for pull strength. On older models --cref and --cw do the job and still work exactly as they always did. MidJourney adds tools rather than taking them away, so pick the one that matches the model you are generating on, and write down which you used.

What breaks without it: she becomes a description again, and descriptions drift.

The style code. A saved --sref code that fixes the palette, the light and the weight of the line. Written next to her image, not remembered.

What breaks without it: the right woman in the wrong painting. Correct face, different world, and the set still looks incoherent.

The written description. One fixed paragraph. Complexion, build, age range, hair, the clothes she usually wears, the one or two objects she carries. Pasted in word for word every time, never retyped from memory.

What breaks without it: slow erosion. You paraphrase in month two and she is a different woman by month four.

Using her in a new scene

The prompt becomes two halves. The scene, which changes, and her, which does not.

Write the scene the way you would write any prompt: where she is, what she is doing, the weather, the time of day. Then attach the reference link, the sref code and the fixed paragraph, unchanged.

Reference weight is the dial worth learning. Too low and she is a suggestion; too high and every image is the same pose and the same crop, because the model is copying the original picture instead of the person in it. Run one scene at three settings, put them side by side, and pick the number you keep. Then use that number.

When a picture comes out wrong, change the scene. She was already right.

One character, not a cast

The urge is to build a whole world at once. Six people, a family, a shop full of regulars.

One is better to start with, and it is not a patience lecture. With one character you can see what your weight setting does, whether your sref is fighting your reference, and which parts of the description are doing real work. With six you get six variables and no way to read any of them.

She also earns her keep faster than a cast would. The same face on Monday's pin, Wednesday's header and next month's story is how people start recognising your work before they read your name.

Pixi's Burrow helps creative founders hold MidJourney character art consistency through builders that lock the character and leave the scene slots open. Unlike a pack of finished prompts you paste and lose, the character stays in your hands and only the world around her changes.

The style half of this is the part that gets under-planned. Committing to one sref code instead of hunting a new one every week is what stopped the woman on my own blog headers drifting, and the DIY MidJourney SREF System at £77 in the EPIC store exists because that is a decision worth making once rather than every Tuesday. If you want the wider mechanics of holding a look steady across a whole set, how to get a consistent art style with AI image generation covers that ground, and I want to make AI art but I don't know where to start is the place to begin if you have not made a first image yet.

Build her this week

Give her a name before you generate anything, because a name stops the fiddling. Write her paragraph. Generate her against a plain wall in good light, so you are judging her and not the room. Upscale the best one and save it.

Then open one note, put her name at the top, and paste in all four pieces.

Which character have you already made and lost?