You are one upload away from the look you want, or one paragraph away — and the tools happily accept either. Here is the short answer: a reference image gives you fidelity you could never write down, and prompt words give you control you can edit. If you can already see the exact look and only need to reproduce it, load the reference. If the look has to bend to fit a brief, vary across a set, or survive an art director changing their mind, write it. And once you have run both a few times, you stop choosing: you describe the parts you can name and reference the part you cannot.
That is the trade in one line — fidelity against steerability. The rest of this piece is how it plays out in real work, including the cases where the choice honestly does not matter.
What a reference image actually gives you
A reference is a bundle. Whatever the tool extracts — colour relationships, contrast curve, texture, edge quality, composition, lighting direction — arrives together, whether or not you wanted all of it.
That bundling is the whole strength. It carries the parts of a style that have no vocabulary: the specific dustiness of one film stock, the way a particular illustrator lands a shadow, a palette you could name three colours of and still miss. If the look is already fixed and shared — a client sent it, a brand deck defines it, an existing set has to be matched — the reference is the closest thing to a lossless handoff.
It is also the honest choice for consistency. Words drift between generations because your phrasing drifts; an image does not change between runs. When a set has to hold together across a dozen frames, the same reference in every call removes one whole source of variation. Our guide to locking down a consistent image set covers the rest of that problem.
The costs are real, though, and they are the mirror image of the benefits:
- You get everything, including what you did not want. Reference a photograph for its light and you may inherit its composition, its subject pose, and the ghost of its background.
- You cannot ask for half of it in plain terms. Most tools give you one blunt dial — a strength or weight — and turning it down weakens every attribute at once, not the one that is overpowering.
- It resists being modified. "The same, but colder and with more negative space" is a sentence. Getting there from an image means finding a different image.
- Rights and provenance follow the file. A reference you do not own, or a living artist's work used to reproduce their exact look commercially, is a problem no output quality solves. Use your own images, licensed ones, or work far enough from a single artist's signature that the result is genuinely yours.
What prompt words give you that an image cannot
Words are the opposite bundle: nothing arrives that you did not ask for, and nothing arrives as precisely as you imagined it.
What you gain is a set of dials you can actually turn. Each clause is independent, so you can change the light without touching the palette, drop the shallow depth of field while keeping the grain, or swap "muted sage and rust" for "cold slate and bone" and hold everything else steady. A described style is also portable — it survives a tool change, a teammate, a brief document, next quarter — where a reference lives in a file someone has to still have.
Words are also the only option in a few common situations: the look exists only in your head, the client wants three directions rather than one, or you are working somewhere that will not take an image input at all.
The costs:
- Vocabulary is the bottleneck. The look you cannot name comes out generic, because vague words land at the average of everything the model has seen described that way. That is a solvable skill problem — describing a visual style you can see but cannot name breaks it into concrete attributes — but it is a skill, and it takes practice.
- Long prompts fight themselves. Past a point, adding clauses starts diluting the ones already there, and the model quietly picks favourites.
- Reproducibility depends on discipline. The same words produce related-but-different images, and a phrase you improvised on Tuesday will not be there on Friday unless you saved it.
Which one for what you are making
The decision is easier when you stop asking which is better and ask what has to stay fixed.
One-off social image, look is loose. Words. Faster to iterate, nothing to manage, and precision is not the point.
Matching an existing set, brand, or client look. Reference. This is the case references were made for — the target already exists as pixels, and describing it is a lossy detour.
A batch of product shots that must look identical. Reference for the style, words for the per-item differences. The reference holds the light and finish steady; the words say what is on the table.
Exploring directions for a new brand or campaign. Words. You need to generate variety on purpose, and you want to change one variable at a time and see what moved.
A look you can see clearly but cannot name. Reference, then reverse-engineer it. Generate from the image, then try to write what you got. If your words can reproduce it, you have upgraded a file into an asset you can edit forever.
Anything where someone will ask for changes. Words, or at minimum words layered on a reference. "Make it warmer" has to have somewhere to land.
When the choice genuinely does not matter
Plenty of the time this is a false decision, and it is worth knowing when so you stop deliberating:
- The style is a well-known, well-labelled category — a clean studio product shot on white, a flat vector icon, a black-and-white documentary frame. The label is unambiguous enough that words land where the reference would.
- The subject dominates the image. A tight crop of one object carries very little style surface. Whatever you feed it, the result looks much the same.
- The output is a rough for a human. Comps, mood tiles, storyboard frames and layout tests only need to communicate an intention. Use whichever is at hand.
- You are going to redraw or heavily edit it anyway. If the generation is raw material rather than the deliverable, precision at this stage is wasted effort.
In all of those, the faster path wins, and the faster path is usually whichever you can start in ten seconds.
Using both: describe what you can name, reference what you cannot
The productive setup is not a compromise, it is a division of labour.
- Pick the reference for one job, not all of them. Decide before you upload whether it is carrying the light, the palette, the texture, or the composition — and say the others in words.
- Write the subject and the framing. These are what references most often contaminate, so state them explicitly to override what comes along for the ride.
- Name the attributes you want to keep steady across the set even though the reference implies them. Redundancy between the image and the words is not waste; it is insurance against the strength dial moving.
- Turn the reference strength down until it stops leaking. Start higher than you need, then reduce until the borrowed composition or subject disappears and only the look remains. That number is your working setting for the whole set.
- Say what must not come across. If your tool takes negative prompts, this is where they earn their keep — the reference's background, its subject, its text.
- Write down what worked. A reference file plus the prompt plus the strength setting is a recipe. Without all three you are starting over next week.
How to tell which one failed when the output is wrong
When a result misses, the fix depends on which input let you down, and the symptom tells you:
- The output looks like the reference's subject, not yours — the reference is dominating. Lower the strength, or crop the reference to just the surface you want.
- The output looks generic and nothing like your description — the words are doing no work. They are probably abstract compliments rather than concrete choices; replace "cinematic" and "aesthetic" with a specific light and a specific palette.
- Every image in the set is slightly different — you are relying on words for something that needs a fixed input. Move that attribute to the reference.
- You cannot make the small change you want — you are relying on a reference for something that needs to be a word. Move that attribute into the prompt.
- The look is right but feels like someone else's — you may be sitting too close to a single source. Fuse a second influence in, or write the style from your own attributes instead of borrowing it whole. That merge is the technique in our image style fusion guide.
Iteration here is not failure; it is the method. Two runs usually tell you which side of the line the problem is on, and after that it is one variable at a time.
FAQ
Does a reference image override the prompt?
Usually it dominates, but not completely — most tools weight them, and the balance sits on a strength control. If the words seem ignored, the strength is too high, not the prompt too weak.
Can I use more than one reference image?
Often yes, and it is the basis of style fusion — one for palette, one for texture, for example. Expect mush if the two disagree strongly, and expect to spend your iterations balancing them rather than refining either.
Is it legal to use someone else's image as a style reference?
Rights depend on the image, the platform, and where you are, and this is not legal advice. The safe practice is simple: use images you own or have licensed, avoid reproducing a living artist's signature look for commercial work, and never use a real person's likeness.
Why does the same prompt give me a different image every time?
Generation is stochastic by design. Words define a region of possibilities, not a point in it. A fixed seed where the tool offers one, plus a reference image, is what shrinks that region.
Should the reference be high resolution?
Clean matters more than large. A small, sharp, well-lit crop of exactly the surface you want beats a big image full of things you will then have to suppress.
How do I turn a reference into words I can reuse?
Generate from it, then describe the result attribute by attribute — medium, light, colour, framing, texture, detail — and test the description on a different subject. Whatever survives that transfer is the real style; the rest was the photograph.
Fidelity or steerability: pick the one your project actually needs, then use the other to cover its weakness. If you want that combination handled as a single recipe — a reference, a description, and a strength you saved once — see what FusionZap is building at FusionZap.
