What Recraft V4 Styles actually does<\/span><\/h2>\nThe model takes between one and ten reference images and turns them into a style you can reuse. That style becomes a thing you attach to a generation, the same way you would attach an aspect ratio.<\/p>\n
Two details make it different from the style controls you may already know.<\/p>\n
It holds more than the vibe.<\/strong> Most style matching captures palette and a general aesthetic. V4 Styles carries texture, composition and rendering through as well, which is why a batch stays consistent instead of loosening as it goes.<\/p>\nIt reads references from anywhere.<\/strong> Many models only work reliably with references generated inside that same model. V4 Styles does not care where an image came from. A frame from another generator, a page from a brand book, an asset from a campaign that shipped three years ago: if it exists as an image, the model can learn from it.<\/p>\nNo training step, no fine-tuning, no waiting. Attach the references and generate.<\/p>\n
Precise or Flexible: picking a mode<\/span><\/h2>\nStyle matching runs in one of two modes, set with the style match parameter.<\/p>\n
Precise<\/strong> is the default, and it is the one for briefs that say \u201cmatch this exactly.\u201d It follows the reference\u2019s texture, palette and rendering closely enough to hold across a full batch rather than a lucky first image. This is the mode the benchmark below was run on.<\/p>\nFlexible<\/strong> is the looser read. It matches the general feel of a reference rather than reproducing it, and it sits closer to the style experience Recraft V3 users already know, refined for V4. Reach for it when you want the mood of a reference and room to move.<\/p>\nA useful way to choose: if a client would notice the difference between your output and the reference, use Precise. If they would only notice that the two feel related, Flexible gives you more range.<\/p>\n
What to use as a reference<\/span><\/h2>\nThe reference set does more to shape your result than the prompt does, so it is worth thinking about.<\/p>\n
One reference works.<\/strong> You do not need a set. A single image is enough to build a style from.<\/p>\nSimilar references sharpen the match.<\/strong> Feed it five images that share a treatment and the model gets a tighter, more confident read of what that treatment is. This is what you want for brand work, where the answer should be narrow.<\/p>\nVaried references widen the range.<\/strong> Feed it five images that differ and the model works across a broader space. This is what you want for exploration, where the answer should have room in it.<\/p>\nTen is the ceiling. Most jobs do not need it.<\/p>\n
The four parts of a style-matched prompt<\/span><\/h2>\nWith a style attached, the prompt stops carrying the look and starts carrying everything else. Four things are worth stating explicitly.<\/p>\n
The subject<\/h3>\n
What is in the frame. Be concrete. \u201cA ceramic pour-over coffee dripper\u201d beats \u201ccoffee equipment,\u201d because the style is already handling how it will be rendered and the prompt only needs to handle what it is.<\/p>\n
The composition<\/h3>\n
Where it sits and how it is framed. Centred, three-quarter view, overhead, cropped at the edge. The style holds compositional character, but the prompt decides the actual shot.<\/p>\n
The context<\/h3>\n
What surrounds the subject. Background, surface, props, setting. This is where a batch gains variety while staying visually of a piece.<\/p>\n
What not to change<\/h3>\n
If something has to survive across every image, say so. Naming it costs you a few words and saves a regeneration.<\/p>\n
The benchmark, and what it measures<\/span><\/h2>\nRecraft ran a blind comparison across 159 real-world style references. In Precise mode, V4 Styles won 91.6% of its head-to-head matchups against seven leading image models, at the versions available on 14 August 2026.<\/p>\n
The methodology is worth stating plainly, because it is unusually transparent for a vendor benchmark. Recraft designed the test. The judging was not theirs: five independent evaluators per matchup, sourced through a third-party crowdsourcing platform, each blind to which model produced which image. Full methodology is available on request.<\/p>\n
Recraft is also clear about scope, and the caveat matters. The benchmark measures one thing: holding a style across a generation. It is not a measure of overall image quality, and the models it was tested against are strong models that are the reason style matching became a standard expectation in the first place. Read the number as \u201cthis is what it was built to win on,\u201d not as a general ranking.<\/p>\n