One reference image can define an entire art direction

Recraft V4 Styles uses up to ten reference images to generate consistent visual series without training, in raster or vector formats.

Producing a convincing first image does not guarantee that the next ones will belong to the same visual world. A color may shift, a texture may disappear, a photograph may suddenly become too realistic, or an illustration may lose its graphic treatment. Recraft is addressing this problem with V4 Styles, a new family of models designed to carry an art direction from one generation to the next.

The process begins with one to ten images. They can come from a brand guide, an earlier campaign, a photoshoot, an illustration, a screenshot, or another generator. No fine-tuning is required. The references are analyzed when the style is created, and that style can then be reused across multiple prompts.

The documentation distinguishes between using a single reference and a broader set. One image keeps the output closely aligned with its most visible characteristics. Several similar references define the direction more precisely, while a more varied collection expands the range of compositions and subjects the model can produce without leaving the same visual language.

V4 Styles does more than reproduce a general color palette. Recraft says the model also follows rendering techniques, textures, lighting, and certain compositional principles. This interpretation works across photography, character design, flat illustration, and vector artwork.

Two fidelity settings control the distance from the references. Precise, selected by default, attempts to reproduce the palette, textures, and visual treatment as closely as possible. Flexible gives the model more freedom and mainly preserves the overall mood. This second behavior is closer to the style system previously available with Recraft V3.

This distinction helps separate stylistic consistency from strict content reproduction. V4 Styles is designed to maintain an art direction across different subjects. The documentation does not describe it as a system for preserving the exact identity of a character, face, or product from every angle.

The feature comes in four models: V4 Styles and V4 Styles Pro for raster images, plus V4 Styles Vector and V4 Styles Pro Vector for SVG files. The standard versions cover routine generations, while the Pro variants produce higher-resolution images. In Recraft Studio, the system is included in the combined Recraft V4.1 model selector. All four models are also accessible through the API.

V4 Styles is not Recraft’s first reference-based style feature. The platform already supported custom styles created from images, with options to add a description, define a palette, or mix multiple sources. The new element is a family of models specifically trained to maintain that direction across a series, with separate Precise and Flexible modes.

In the API, references can be attached directly to each request or saved as a reusable style. When they are submitted with a generation request, Recraft creates a private style and returns its identifier in the response. Future calls can reuse that identifier without uploading the files again. PNG, JPG, and WebP files are supported, with a limit of ten images, 64 MB in total, and 10 MB per file.

Recraft is supporting the release with a benchmark focused exclusively on style fidelity. Precise mode was compared with GPT Image 2 High, Nano Banana Pro, Seedream 5.0 Pro, Flux 2 Max, Nano Banana 2, Midjourney, and Krea 2 Large across 159 real-world styles. According to the company’s results, V4 Styles won 91.6% of its head-to-head matchups, rounded to 92% in the announcement.

Each comparison was judged blindly by five people recruited through an external crowdsourcing platform. They did not know which model had produced each image. Recraft specifies that it designed the testing protocol itself, and the complete methodology was not published with the results, though it is available upon request. The score measures how closely a generation follows a reference style, not overall image quality, prompt understanding, or content accuracy.

The published examples compare V4 Styles directly with Nano Banana and Midjourney. Recraft highlights better retention of color palettes, textures, and visual treatment across several generations. These comparisons were still selected and described by the company. They illustrate the intended behavior but do not replace an independent evaluation conducted with different references, prompts, and subjects.

API pricing starts at $0.035 per raster image with V4 Styles and increases to $0.10 with the Pro version. Vector generations cost $0.05 and $0.12 respectively. Creating a separate reusable style costs $0.005. V4 Styles is available in Recraft Studio, the web and mobile apps, the live demo, and the API.