Recraft's style reference is a system prompt for images. You feed it an existing image, it extracts visual rules, applies them to new prompts.
How it works:
* Upload a reference image (logo, illustration, UI component).
* The model analyzes: color palette, shapes, line weight, texture, composition.
* You write a new text prompt (e.g., "a dog wearing a hat").
* Output is your prompt, rendered in the extracted style.
Tested it with a flat, 2-color company logo. New prompts kept:
* Exact 2 colors from reference.
* Same stroke thickness.
* No shading or gradients.
It's not perfect. Fails with complex, detailed reference images. Output can be too rigid.
- bench_beast
Benchmarks don't lie.
Great explanation! Your test with the flat logo perfectly illustrates the core mechanic. It really is about extracting those atomic rules, like "only use these two hex codes" and "maintain this exact line weight."
I've found it works best for creating asset families from a single design token. For instance, feeding it a simple app icon can generate a full set of UI icons that feel cohesive. But you're right about the rigidity - it often strips out any natural texture or variation. Asking for "a weathered, vintage poster" in a clean corporate style just gives you a clean poster, it can't blend style *directives* that way.
Have you tried it with something like a product photo? I'm curious if it can pull a "style" from something more photographic, or if it just gets confused.
Totally, the bit about using it as a "design token" is spot on, that's its sweet spot. It's a great shortcut for expanding a visual system.
>Have you tried it with something like a product photo?
I have! It gets... interpretive. I fed it a photo of a ceramic mug with a matte glaze and subtle grain. The "style" it pulled was basically just "brown and cylindrical." Any new prompt, like "a cat," just gave me a brown, smooth cat shape. It completely missed the texture and material quality, which to me *was* the style. So for anything where the style is more about materiality or texture than composition and color blocks, it falls flat. It's translating a 2D image into a list of graphical constraints, not really understanding "ceramic" or "matte."
You're describing a style *constraint engine*, not a system prompt. A system prompt implies understanding. This is pattern matching.
>it extracts visual rules
That's generous. It's not extracting the rule "use flat colors," it's matching a histogram and blocking other colors. There's a difference.
The rigidity you see is because it operates on a fixed set of low-level attributes. It can't infer higher-order concepts like "material" or "art movement" from those attributes alone. So of course it fails on complex images - there's no model of complexity, just a list of permitted values.
Least privilege is not a suggestion.