What does image consistency actually mean in Midjourney?
Image consistency means the same subject and the same visual treatment survive across separate generations. In Midjourney, that is not something your text prompt controls. It is controlled by reference parameters that pin a face, a palette, or a whole aesthetic to every new image you make.
If your outputs feel inconsistent, you are almost certainly writing longer prompts to fix a problem that prompts cannot fix. The prompt describes intent. The reference parameters carry identity.
Three parameters do the work. Omni Reference (--oref) carries the subject. Style Reference (--sref) carries the look. Personalization (--p) carries your accumulated taste. Everything else in this article is about how to combine them without wasting credits.
Why do your images drift when the prompt is identical?
An identical prompt does not produce an identical image because each generation starts from different noise. Midjourney re-invents the face, the fabric, the light angle and the colour grade every run. Nothing in plain text is specific enough to describe a face the model must reproduce.
This is why "same character, different pose" fails in practice. You can write "a 32-year-old Hong Kong woman with shoulder-length black hair, round glasses, navy blazer" and get twelve different women who all technically match that description.
Text describes categories. References describe instances. That distinction is the whole game.
Adding more adjectives makes drift worse, not better, because longer prompts dilute the weight of every individual term. The fix is to shorten the prompt and attach a reference image instead.
There is a second, quieter source of drift: version and setting changes between sessions. If you generated your first batch three weeks ago on a different model version, with a different stylize value, or before you built a Personalization profile, the model you are prompting today is not the model that made your original image.
So before you blame the prompt, check three things: the model version, the stylize value, and whether a Personalization profile is active. Any one of them changing will shift your output even with byte-identical prompt text.
How does Omni Reference lock a character across images?
Omni Reference pins the content of a reference image, most usefully a person, object or product, into new generations. You pass a publicly accessible image URL with --oref, and control how strongly it is enforced with --ow.
According to Midjourney's Omni Reference documentation, --ow accepts any value from 1 to 1,000, and defaults to 100. The documentation also notes that unless you are running a very high stylize value, weights above 400 tend to produce unpredictable results.
In practice the useful band is narrow. Around --ow 100 you get a recognisable person who still adapts to new poses and lighting. Push toward --ow 300 and you get a much closer likeness but far less flexibility, because the model starts reproducing the original pose and framing along with the face.
Two operational details matter. The reference URL must be publicly reachable, so a private Google Drive link will fail. And dragging an Omni Reference image into the Imagine bar automatically runs the prompt on V7 rather than an older version.
How does Style Reference lock the look rather than the subject?
Style Reference does the opposite job to Omni Reference. Instead of carrying who or what is in the frame, --sref carries the visual treatment: palette, tonal range, grain, lighting mood and compositional habits. It leaves the subject entirely to your text prompt.
The combination is where consistency actually comes from. Midjourney's own guidance is to use Omni Reference alongside Style Reference to build consistent sequences, and you can pass the same image as both --oref and --sref when you want to lock the subject and the look together.
Personalization adds a third, slower layer. A Personalization profile, applied with --p, is built from your own ranking choices over time and biases future generations toward images you have already favoured. Midjourney states that a V7 global Personalization profile is compatible with V8.2.
Treat --p as your house style and --sref as your campaign style. The profile is a long-term bias; the style reference is a per-project instruction you can swap in a single line.
What does a repeatable image workflow look like, step by step?
A repeatable workflow means you generate the reference once and reuse it, rather than re-rolling the dice each time you need a new image. Four steps, roughly twenty minutes for the first run, under two minutes for every image after that.
Step 1 — Generate the anchor. Write a short prompt for your subject, generate a batch, and pick the single strongest result. Upscale it and copy its public image URL. This is now your canonical reference and you should not regenerate it.
Step 2 — Generate the style plate. Separately, produce or choose one image that carries the look you want: the palette, the light, the finish. This becomes your --sref URL for the whole project.
Step 3 — Build the template line. Keep the text prompt short and change only the scene. Everything after the double dashes stays frozen for the whole project.
Step 4 — Vary one thing at a time. Change the scene words, or change --ow, but never both in the same run, or you will not know which change caused the difference.
Here is the template to copy. Replace the two URLs with your own and change only the scene description between runs.
Try this prompt:
[scene description, 8 to 15 words, e.g. "seated at a cafe table reading a printed report, morning light"] --oref [YOUR_ANCHOR_IMAGE_URL] --ow 120 --sref [YOUR_STYLE_IMAGE_URL] --ar 16:9 --hd
Parameter reference, for quick recall:
--- --oref [URL] — Omni Reference. Pins the subject. URL must be publicly accessible.
--- --ow [1 to 1000] — Omni weight. Default 100. Above 400 becomes unpredictable per Midjourney's docs.
--- --sref [URL] — Style Reference. Pins palette, light and finish, not the subject.
--- --p — Personalization profile. Long-term bias from your own ranking history. V7 global profile works with V8.2.
--- --ar — Aspect ratio. --no — exclude an element. --c — chaos, adds variety. --hd and --sd — resolution control.
Where does this technique break down?
Reference locking is reliable for faces, products and palettes, and unreliable for text, logos and exact geometry. Knowing the failure modes before you commit a campaign to this workflow saves the most time.
It costs more. Midjourney's documentation states that using Omni Reference costs twice the GPU time of a regular V7 image. On a Basic plan with limited fast hours, a 40-image campaign at --oref burns through your allowance roughly twice as fast as you budgeted for.
It does not cover every tool. Omni Reference is not compatible with features that still run on V6.1, including inpainting and outpainting. So the "generate with a reference, then patch it with inpainting" workflow does not hold together; you have to choose one.
High weights collapse variety. Beyond roughly --ow 400 you are no longer generating new images of your character, you are generating near-copies of your reference photo with slightly different colours.
Faces are approximate, not identical. For a fictional brand spokesperson this is fine. For a named real person, or for anything a client will compare side by side against a photograph, expect visible variance and budget for editing.
It will not hold a logo or exact text. Reference locking works on visual character, not on vector precision. If your image needs a real wordmark, a specific typeface or an exact product label, generate the image clean and composite the logo in Canva, Figma or Photoshop afterwards. Trying to force it through --oref wastes credits and still gives you a warped approximation.
One more practical caveat: version behaviour changes. Omni Reference behaviour was documented against V7, and V8.2 is the current default producing native 2K output. Verify parameter behaviour on the version you are actually running before you scale a batch.
The honest summary is that reference locking raises your hit rate rather than guaranteeing it. In practice you still discard results, so any workflow you build on top of this should assume a human review step rather than direct publication.
How can you test this in the next twenty minutes?
Run a controlled three-image test rather than a full campaign. The goal is to find your own working --ow value, because the right number depends on your reference image and how much pose flexibility your project needs.
Generate one anchor image of a subject and copy its URL. Then run the same short scene prompt three times, changing only the weight: once at --ow 60, once at --ow 120, once at --ow 300.
Put the three results side by side and ask one question: at which weight does the subject stay recognisable while the scene still changes? That number is your project default, and you will reuse it for every image in the set.
Write it down somewhere you will find it again. The reason most people never get consistent output is not that they lack the technique, it is that they rediscover their working settings from scratch every project.
If your wider problem is that AI output quality swings run to run, the same principle applies to text. We covered the reliability side of prompting in what changed about step-by-step prompting in 2026, and the cost side of tool subscriptions in what US$20 a month actually buys.
The takeaway
Consistency is not a prompting skill. It is a reference discipline: generate your anchor once, freeze it, and change one variable at a time.
That is the difference between using an AI image tool and running an AI image workflow. The first produces good single images. The second produces a set you can actually ship.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Turn a Technique Into a Workflow
Knowing the parameters is step one. Building a production pipeline your whole team can run is step two.
UD has spent 28 years helping Hong Kong businesses put technology to work, and we'll walk you through every step, from tool selection to workflow design to deployment.
