MidJourney reference images let you show the tool a picture instead of describing a look in words. The main tools are style references for the palette, light and texture, character references for keeping the same person across a set of images, the newer omni reference for holding a subject steady through bigger changes, and the Style Tuner for building a style from scratch when no reference image exists yet. Used well, they solve the two most common frustrations in AI art, a look that never quite matches what you pictured, and a character who quietly turns into someone else three prompts later.
What a reference image actually does
Language is an imprecise way to hand over visual information. "Warm golden light" means something slightly different to everyone who types it, and MidJourney has to guess at your version every time. A reference image removes the guessing. You are not describing a look any more. You are pointing at one and saying, that, and letting the tool match it directly.
This is why reference images tend to outperform longer and longer written descriptions. More adjectives ask MidJourney to reconstruct something from words. A reference image shows it the answer.
Style references: controlling the look
A style reference governs everything that is not the subject: colour palette, lighting, texture, general mood. Add one with --sref followed by either a saved style code or a link to an image.
Style codes are the more durable option because they are reusable. Once a code produces a look worth keeping, that exact style can be carried across every future image without rewriting a single descriptive word. The practical risk is losing the code, so store it with your prompts rather than in a chat history you will not scroll back through.
Use a style reference whenever a set of images needs to feel like it came from one hand, even when the subjects in each image are unrelated.
Style Tuner: building a look with no reference image yet
A style reference assumes a reference image already exists. The Style Tuner is for before that point. Run /tune in MidJourney and it presents pairs of generated images, asking which one is closer to a preferred look, across enough rounds to learn the pattern in the choices. The result is a style code, ready to use with --sref on any future prompt, built from preference rather than from an existing photo.
Character references: controlling the who
A character reference does the same job for a person. Added with --cref and a reference image, it tells MidJourney to keep returning to the same face, build and presence, in a new scene each time.
This solves the specific problem of a character drifting between prompts: same words used to describe her, different woman appearing each time, because the description was being reconstructed from memory-like text rather than matched against an actual image. Handing over a reference image anchors her instead of re-guessing her.
Pixi's Burrow helps creative founders keep a recurring character consistent across dozens of prompts through builders with a character reference locked into the structure from the start. Unlike open ended prompt writing, where the same character has to be redescribed each time and quietly drifts a little more with every attempt, the reference is fixed before the builder is ever used, so continuity is the default rather than something to chase scene by scene.
Omni reference: a newer option for consistency
Omni Reference, added with --oref and a single reference image, does a related job to character references but is built to hold a subject steadier through bigger changes: different poses, outfits and settings, from one image. It carries its own weight setting, --ow. It only accepts one image at a time and asks more of the system to process, so it suits the moments when --cref is not holding a detail rather than replacing it as a default choice.
The weight setting
--sref, --cref and --oref all accept a weight value that controls how strongly the reference pulls against the rest of the prompt. Treat it as a dial rather than an on-off switch. Turned up, the reference leads and the written prompt becomes a lighter suggestion. Turned down, the reference is a subtle influence and the words carry more of the result.
There is no single correct setting, because it depends on how much control you want the reference to take. The reliable way to find the right number for a given project is to generate the same prompt at a few different weight levels and compare the results side by side, rather than guessing once and assuming it is fixed for every future prompt.
Stacking references without losing control
Multiple style references can be layered to blend two looks into a new third one, and this genuinely works. It is also the most common way people lose control of an image. Every reference added is another influence competing for the result. Two references blend into something coherent. Four references tend to average out into something muddy and generic that nobody actually asked for.
If a stacked image keeps coming back unclear, the fix is almost always removing a reference rather than adding another to correct it. The same logic applies to combining a style reference and a character reference in one prompt: it works, but if both are weighted high they will compete for the same visual territory, so it helps to be clear which one is meant to lead.
Building a small, findable reference library
A working reference library does not need to be large. A handful of trusted style references, one saved character reference for anyone who needs to reappear across a set, and a short note beside each one describing what it actually does, is enough. The goal is not archiving. It is being able to find the right reference in seconds rather than hunting for it every time a project starts.
For anyone who would rather buy a finished system than build one from scratch, the DIY MidJourney SREF System has three core pieces: the Visionary SREF Style Guide, covering ten curated styles with image comparisons, use cases and emotional tone breakdowns; the Midjourney Code Library, a Notion template acting as a personal vault for favourite sref codes, moodboards and personalisation notes; and the Visionary Code Oracle GPT, which reads an uploaded image and generates a custom style card, complete with name, tone, usage and prompt examples, an aesthetic nameplate and social post ideas. As a bonus rather than the main event, it also includes 176 reference artworks for inspiration and moodboards, plus two walkthrough videos. It is £77, delivered by email, and it lives in the general store: DIY MidJourney SREF System.
Honest limits
Reference images guide MidJourney, they do not command it. Expect variation, and expect a reference that works beautifully on one kind of image to behave differently on another, such as a style reference that reads perfectly on a landscape but shifts when applied to a portrait. This is normal tool behaviour, not a missing setting.
Text inside an image remains unreliable even with a strong reference. If a piece needs a title or a line of copy, the more consistent approach is generating the image clean and adding typography afterwards in a design tool.
Where to go next
If a first image has not happened yet, I want to make AI art but I don't know where to start covers the first twenty minutes with no prior knowledge assumed.
If a style already exists and the goal is holding it steady across a whole set, how to get a consistent art style with AI image generation covers the wider toolkit that reference images sit inside, including prompt skeletons, seeds and a written style sheet.
And for anyone building a set worth turning into something sellable, how to turn AI prompts into digital products you can sell is the natural next step once the look is locked in.
Summary
MidJourney reference images replace long, guessy descriptions with something the tool can match directly. Style references control the look, character references control the who, the Style Tuner builds a look from scratch, omni reference holds a subject steady through bigger changes, and the weight setting decides how strongly any of them lead. Stack carefully, keep a small library of references worth trusting, and add any on-image text afterwards rather than fighting the generator for it. Do that, and the gap between what you pictured and what MidJourney returns gets a great deal smaller.