Recraft: Recraft V4 Styles

Recraft V4 Styles is an image generation model from Recraft, a provider known for its design-oriented image tooling. It sits in the image category, and its defining trait is style consistency: rather than treating a prompt as a one-off request, it is built to reproduce the rendering of a reference image you supply. The model was released by the provider on 26 August 2026. It is not a general chat or text model, so it will not fit workflows that expect written answers; its entire output is visual.

On the input side, the model accepts text and images, and it produces images only. It has a context window of 65536 tokens and a maximum output of 58982 tokens. Every request requires at least one style reference image, and the endpoint accepts up to ten input references, which lets you steer the visual style with more than a single sample. Output shape is controlled through an aspect ratio setting that covers square, portrait, landscape and widescreen formats: 1:1, 2:1, 1:2, 3:2, 2:3, 4:3, 3:4, 5:4, 4:5, 16:9, 9:16 and auto. You can also request several images in one call, from one to six.

The practical value follows directly from those capabilities. If you are building a tool that must keep a house illustration style across many generated assets, such as marketing visuals, app artwork, editorial images or a product catalogue, feeding reference images and asking for a new subject in the same rendering is the core use case. Because references are accepted as images rather than described in words, this suits teams that already hold a defined visual language and want to extend it without redrawing style prompts each time. The aspect ratio list makes it straightforward to mint the same asset set for square feeds, vertical stories and wide banners from one workflow, and requesting several images per call helps when you want variations to choose from.

You call the model through api14's OpenAI-compatible API using the model id "recraft/recraft-v4-styles". Because the endpoint follows the OpenAI request shape, you can point existing clients at api14 and swap the model id without rewriting your integration. The model exposes no general supported parameters of its own; instead, generation is tuned through image parameters. Aspect ratio takes one of the listed enum values, n accepts an integer from 1 to 6, and input references accept between 1 and 10 images. Note that at least one reference is required, so a text-only request will not run.

It is worth checking whether a plain text-to-image model fits better when you have no reference material to work from, since this model depends on references for its style behaviour. It also returns images rather than text, so anything needing reasoning or a written response belongs on a text model. Within image work, though, it is the right pick when consistency with an existing look matters more than one-off novelty.