AI Now Designs Like a Human Creative Director

A new image generation model is making decisions about visual design that feel surprisingly human. Recraft V4, the latest iteration from the Recraft team, introduces what they call 'design taste'—the ability to make intentional choices about composition, lighting, and color rather than producing generic images. This means the model can create visuals that look art-directed even from simple text prompts, bridging the gap between automated generation and professional design work. The approach represents a shift from models tuned for broad preference to those developed with specific aesthetic principles in mind.

Recraft V4 comes in four distinct versions: two raster and two vector models, all sharing the same core design taste and prompt accuracy but differing in output format, resolution, and speed. The raster models produce traditional pixel-based images, while the vector models generate actual SVG files with editable paths and structured layers. This dual capability allows users to choose between high-resolution photographic outputs and scalable vector graphics depending on their needs, with all versions maintaining consistent visual decision-making.

The model was developed in collaboration with designers and tuned specifically around professional design aesthetics and expectations. Unlike most image generation models that aim for general appeal, V4 focuses on art-directed composition, making choices about how elements are arranged, how colors relate to each other, and where the eye moves through the frame. The model accepts prompts up to 10,000 characters and follows them closely while maintaining overall image coherence, demonstrating strong prompt accuracy whether the input is brief or detailed.

Several examples from the paper demonstrate V4's capabilities across challenging design scenarios. For complex typography layouts, the model successfully integrated text as a first-class element of composition, creating cinematic editorial posters with specific typographic hierarchy where text interacted naturally with the image rather than appearing stamped on top. In product shot rendering, V4 handled precise material physics including brushed steel finishes, matte aluminum surfaces, and micro-reflections from engraved monograms while maintaining playful commercial energy.

The model also excelled at extreme detail rendering, as shown in a macro photograph of snake scales where it resolved each overlapping scale individually with iridescent color shifts from neon green to violet under dramatic side lighting. For vector graphics, V4 produced consistent icon sets with uniform stroke weights and cohesive styles across multiple distinct shapes, outputting real SVG files that open directly in design tools like Figma or Illustrator. The illustration capabilities extended to creating images with artistic character, such as a modern vector illustration of a levitating man with distorted proportions and a vibrant color palette that felt hand-drawn rather than generated.

What makes Recraft V4 particularly significant is its native vector output capability. The SVG models produce actual editable vector files with clean paths and structured layers—not traced rasters or bitmaps wrapped in SVG containers. This means designers can open the output directly in professional tools and edit paths, recolor elements, or scale to any size without conversion steps. The paper demonstrates this with a set of six icons that form a single editable file, useful for brand assets, icon sets, illustrations, and logos that need to scale cleanly or be modified after generation.

Extend across multiple design domains. For posters, packaging, and editorial layouts, V4's integrated text rendering treats typography as a structural part of the image that bridges visual elements and responds to spatial context. For commercial applications, the model's material rendering and lighting control enable realistic product shots with specific aesthetic requirements. The vector output capability opens new possibilities for design workflows where assets need to be editable and scalable, potentially reducing the time designers spend on manual vectorization or asset creation.

While the paper demonstrates impressive capabilities, it also reveals certain limitations through the examples provided. The model appears optimized for specific design aesthetics—modern, commercial, cinematic, and minimalist styles—which may limit its applicability for other visual traditions or historical art styles. The vector output, while groundbreaking, currently focuses on relatively simple geometric shapes and iconography rather than complex illustrations with numerous detailed elements. Additionally, the paper doesn't address how the model handles prompts requesting styles outside its trained aesthetic range or how it manages conflicting design instructions within complex prompts.