The fastest way to settle a lookMuse Image · references and combining visuals
Show it. Then say what to take.
A reference answers in one step what adjectives only approach. But an image carries a dozen properties at once style, palette, light, framing, subject and if you do not say which one you meant, the tool picks. Naming the job is what turns a reference from a suggestion into an instruction.
Schematic, not output the argument is the tick and the cross, not the pictures.
What actually comes across
Seven attributes, and how well each one travels
A reference influences the character of what you get rather than pinning down its details, so the properties that are about feel come across strongly and the ones that are about identity do not. The ordering is ours, from how these tools behave in use.
Style and medium
Watercolour, flat vector, film photography, block print. The single strongest transfer, and the reason most people reach for a reference in the first place.
Travels well
Palette
The colour relationships come across reliably, which is what makes references the practical way to keep a set on-brand without naming hex values you cannot specify anyway.
Travels well
Lighting
Direction, hardness and warmth carry over. Worth naming explicitly, because light is the attribute most often taken by accident when you meant something else.
Travels well
Composition and framing
Where things sit, how much space is left, the crop. Useful when you need a series to sit the same way a run of headers that all leave room on the left, say.
Mostly
Texture and finish
Grain, paper tooth, gloss. Comes across as an impression rather than a match close enough for mood, not close enough to pass as the same stock.
Mostly
Subject identity
The one people expect and do not get. A reference showing your product, your dog or your face yields something in the same family a relative, not the individual. For the actual thing, use the actual photograph.
Barely
Exact details, text and counts
A reference does not fix what approximation breaks. Lettering stays unreliable, counts stay approximate, and fine specifics stay generic no matter what you supply.
No
Which is why you name the attribute
Supplying a photograph of a sunlit kitchen and saying “like this” leaves five plausible readings on the table. You might get the warm light you meant, or a kitchen, or the pale wood, or that particular wide crop.
One clause removes the ambiguity: “use this image for its lighting only, and ignore the subject and the setting.” Saying what to ignore does as much work as saying what to take the same principle that governs wording an edit.
Combining visuals
One job per reference, and no more than three
Two references without roles will compete. Each one is offering style, palette, light and framing simultaneously, so handing over a pair and hoping produces a result that splits the difference in ways you cannot predict and therefore cannot correct.
Give each picture a job and say it out loud: this one for its colour, that one for its layout. Then a disappointing result tells you which reference to change.
Conflicts resolve badly. If both references have strong and different lighting and you have not said which to use, you rarely get either you get an average that belongs to neither image.
Two is comfortable; three is the ceiling. Past three, attribution collapses: you can no longer tell which input caused which part of the output, and once you cannot attribute you cannot steer. At that point you are back to rolling the dice, which is what references were meant to stop.
Schematic the point is that a named role makes a poor result diagnosable.
Your subject, someone else’s look
The most useful pair there is. A photograph of the real thing supplies the subject, and a second image supplies the treatment so you get your actual product or place rendered in a look you chose.
your photo → the subjectstyle image → medium and palette
Use the first image for the subject and the second for its style and palette only, ignoring what is depicted in the second.
Watch for: the style reference quietly bringing its lighting too. Name the light explicitly if it matters.
Palette from one, arrangement from another
For work that has to sit in a layout: one reference fixes the colours, the other fixes where things go and how much empty space is left. Neither contributes subject matter.
colour image → palettelayout image → composition
Take the palette from the first image and the framing and spacing from the second; the subject is described in my prompt.
Watch for: lighting conflicts two strongly lit references with nothing said about light average into something flat.
Your own earlier result as the anchor
Once a generation lands, that image becomes the most valuable reference you own, because it already contains the exact combination you were chasing. This is how a set gets made the anchor method.
your best result → everything but the subject
Match the style, palette and light of this image exactly, keeping the same treatment, but the subject is now a different one.
Watch for: chaining. Always reference the original anchor, never the previous image in the run.
Real object, invented setting
A product photograph plus a scene reference: the object stays the one you photographed, and the environment around it comes from somewhere else. Usually better done as an edit but worth knowing as a reference pattern when you want the scene rebuilt rather than swapped.
product photo → the objectscene image → setting and light
Keep the object from the first image as it is and place it in a setting like the second, matching its light.
Watch for: the object drifting away from the real one. Check it against the original photograph product images.
Consistency across a set
Anchor to one image, not to the last one
Twelve images that look like they belong together is the job references exist for. There is one mistake that quietly ruins it, and it is the intuitive way to work.
Do this
Every image references the same anchor
anchor→02·03·04·05·06
Choose one image as the reference for the whole set and point every subsequent generation at that same file, changing only the subject clause. Item six is as close to the anchor as item two.
Not this
Each image references the previous one
01→02→03→04→05→06
Every step adds a small deviation, and the deviations compound. By item six you have drifted somewhere the first image would not recognise the same accumulation that degrades repeated edits.
Keep the wording constant too: identical style and light clauses across the set, with only the subject changing. The reference carries the look; the repeated clauses stop the prompt from fighting it.
Keep the anchor file, because you cannot recreate it
The anchor is usually a lucky generation, and generation is not reproducible the prompt that produced it will not produce it again. If you lose that file, the set ends when you lose it.
Save the anchor somewhere deliberate before you start the run, which on a Mac means a folder rather than the chat why the desktop matters for this.
Whose pictures
Where the references should come from
Technique aside, a reference is someone's photograph or artwork. Most of the time that someone is you, and there is nothing to think about.
Reference source
Where it stands
Your own photographs and artwork
Straightforward. The best source, and usually the most useful, because it already looks like your work.
Images you have licensed
Fine, subject to what the licence actually permits some stock licences address AI use directly, so it is worth reading rather than assuming.
Genuinely public-domain work
Fine. Note that a museum's or studio's modern photograph or restoration of an old work may carry its own rights.
A living artist’s distinctive style
The case to think hardest about. A style is often the part of a practice that took longest to build, and reproducing it from their work is different from being influenced by it whatever any tool permits.
A photograph of someone you do not know
Avoid. You get a family resemblance rather than a likeness anyway, so it fails technically as well as raising the obvious problem.
Brand assets and product shots you do not own
Avoid. The output resembles the brand without being it, which is the worst of both outcomes.
What you may do with the resulting images is a separate question again, governed by Meta’s terms rather than by where the reference came from what Meta says it uses. None of the above is legal advice; it is the shape of the question rather than an answer to it.
A picture you supply alongside your prompt so the tool can see what you mean instead of inferring it from adjectives. It is the fastest way to settle a look, because showing is precise about exactly the things language is vague about.
Does a reference copy the image I give it?
No. It takes properties from it style, palette, light, arrangement rather than reproducing it. A reference showing a specific person or product will not clone that person or product; you get something in the same family what travels and what does not.
Why did it take the wrong thing from my reference?
Because an image carries many attributes at once and nothing in the picture says which one you meant. Name it: say you want the palette, or the lighting, or the framing, and say what to ignore.
Can I use more than one reference?
Yes, and the rule that makes it work is one job per reference. Say what each image is for this one for its colour, that one for its layout because two references given without roles will compete and the result splits the difference unpredictably four patterns that work.
How many references is too many?
Two is comfortable, three is usually the practical ceiling. Past that, attribution collapses: you can no longer tell which image caused which part of the result, which means you can no longer steer it.
How do I keep a whole set looking consistent?
Pick one image as the anchor and reference that same file for every item in the set, changing only the subject. Do not chain each new image off the previous one, or small differences compound down the line the anchor method.
Can I use a photo of myself as a reference?
You can supply it, but expect a family resemblance rather than a likeness, and expect it to vary between generations. For anything where the person needs to be recognisably themselves, use the photograph itself.
Is it alright to use another artist’s work as a style reference?
Using a living artist’s work to reproduce their distinctive style is the case to think hardest about, whatever the tool permits it takes the part of their practice that is hardest won. Your own work, licensed images and genuinely public-domain material raise none of the same problems the fuller picture.
Does a reference fix text and counts?
No. Rendered text and exact counts stay unreliable no matter what you supply, because a reference influences the character of the output rather than pinning down its details.
Is a reference better than a longer prompt?
For anything visual you can already see, yes. The rule of thumb is that if your third attempt is a longer version of your first, more words are not the answer a picture is. Why prompts grow when they are not working.
One job per picture · anchor the set
The picture is the easy part. Naming its job is the skill.
Supply the image, say which property you want taken and which to ignore, and give every reference in a combination a role of its own.
Independent guide not affiliated with Meta. How references are supplied differs by surface and rollout wave; confirm what your app offers in Meta’s own documentation.