Style consistency is the promise every image generator makes and few keep past the second image. Recraft answers it with a reference image saved as a reusable style, not a cleverer prompt. The vendor’s creation guide puts the split in one line: “The reference sets the style, the prompt does everything else.”
What Recraft documents
Recraft V4 Styles shipped in August 2026 for exactly this job: “Style it once, every image matches.” It holds “texture, composition, and rendering, not just the general vibe.” A reference set runs from one image to ten with no training, and references can come from anywhere: “other generators, brand libraries, or an asset pulled from a client’s old campaign.”
The four V4 Styles models always require a style, so a request with neither a style_id nor references is rejected. Other V4 and V4.1 models treat style as optional, at “the same price as generating without a style.” The V4.1 breakdown covers which return vectors.
Two style systems, two outcomes
Recraft ships two generations of styling, and they are not interchangeable.
| Mechanism | Field | Documented for |
|---|---|---|
| Curated name | style | V2 and V3 models |
| Custom style | style_id | V2, V3, V4, V4.1, V4 Styles |
| Reference images | style_references / style_reference_urls | V2, V3, V4, V4.1, V4 Styles |
Only the last two hold a look across a batch: a curated name is a category, and two images inside one category can still look like distant cousins. Studio names its controls exactly as Style, Create style, Add reference material, Adjust weights, Type, Palette, Test it, Save style.
Precise or Flexible
A style stores a match value, and it decides how literally the next generation reads it. V4 and V4.1 have two modes; V2 and V3 accept only regular.
What Precise holds
Precise is the default. The model page lists what it keeps: “rendering technique, color, composition, and lighting.”
When Flexible wins
Flexible “matches the general vibe, giving the model slightly more liberty.” Pick it when the mood is the deliverable. A generation request can override the stored value with style_match, so one style can serve a strict product set and a loose illustration set.
A reference set that agrees
One reference is the safe default: “One strong image gives you a clean signal.” The model fixes on that image and reproduces it across the batch.
More references change the result in one direction: “Every reference has to come from the same visual world.” Mismatched references “compete, and the output pays for it.”
File quality sets a ceiling before any prompt is read: “A low-resolution reference caps your results before the prompt gets a say.” API limits are PNG, JPG or WEBP, ten images, 64 MB total, under 10 MB each.
Colour is the one promise Recraft declines to make: “Reference-to-output color consistency isn’t guaranteed.”
Worked examples: two saved sets
Both runs follow one shape: the reference locks the rendering logic once, the style is saved as its own object, and every prompt after that names only a subject.
Example one: an icon set
Save a style from one flat vector icon you own, then generate each object with a short prompt, because the style already carries line weight, corner radius and fill.
fox head, front view, flat vector icon
leaf, front view, flat vector icon
mug, front view, flat vector icon
paper plane, front view, flat vector icon
light bulb, front view, flat vector iconThe nouns change and the drawing language does not, which is why an icon grid is the fastest place to watch a style hold or fail.
Example two: a product set
Here the reference is a product photograph, so lighting, background and camera angle come from the image. Change the noun and keep the clause:
a ceramic mug on a linen cloth, three-quarter angle, soft morning light
a leather wallet on a linen cloth, three-quarter angle, soft morning lightRepeating the scene clause is the point. Run the long version through the API and let each request carry the same style_id instead of the same file.
Where a set still drifts
Repetition is the commonest complaint, and the cause is an underwritten prompt. “A person” returns the same pose every time; “a person walking forward with their hair flowing in the wind, carrying a small bag” does not. Adding a fourth reference to shake up a repetitive result is how a blend happens.
In vector styles, line weight, corner radius and fill logic visibly break a set, and all three come from the reference. The SVG export walkthrough covers the file end.
Reusing one style across a team
A saved style stops depending on its sources: it becomes “its own object, separate from the source.” It can be shared to a team account, and its identifier copied from the Styles panel three-dot menu. A style records the model it was created for and must be used with a matching model.
The API exposes the same two routes. Send the returned style_id on every generation, or attach references and let the server create a private style and return the resolved style_id. The two are mutually exclusive; the Studio walkthrough covers the hand-driven version.
Frequently asked questions
How many reference images should I start with? One. A strong reference gives a clean signal, and more only helps when the references already agree.
Is Precise better than Flexible? No. They differ in how closely the style is followed, not in quality. Precise holds rendering, colour, composition and lighting; Flexible keeps the mood.
Can I loosen the match without editing the style? Yes. Sending style_match overrides the stored value for that request alone. A value the resolved model does not support is rejected.
Why did the colours shift when the style matched? Because reference-to-output colour consistency is not guaranteed. Treat the palette as approximate and check the output.
Do the V4 Styles models work without a reference? No. Those four always require a style, as a style_id or as references. Other V4 and V4.1 models treat style as optional.
Start from the reference, not the prompt. The style carries the look, the prompt carries the subject, and a set drifts when one does the other’s job.