Named Style References vs. Descriptive Attributes: Why "In the Style of X" Works in Some AI Tools and Backfires in Others

There are two ways to tell an AI image or video tool what look you want. One is to name it: a movement, a medium, an illustrator, a decade, a camera stock, and let the model unpack what that name implies. The other is to describe it: list the actual visual attributes, the line weight, the palette, the lighting, the grain, without ever naming a shorthand for them. Both approaches produce good results somewhere. The problem is that which one works depends entirely on which tool is reading the prompt, and the two approaches fail in opposite, non-obvious ways when used in the wrong place.

Two ways to point a model at a look

A named style reference compresses a huge amount of visual information into a few words, because it leans on the model having already learned what that name tends to look like. "Ukiyo-e woodblock print," "Bauhaus poster," "1970s film photography" — each of these carries an implicit bundle of composition, palette, and texture choices that would take a full paragraph to spell out attribute by attribute. A descriptive attribute prompt does the opposite: it names no shorthand at all, and instead states the palette, the line quality, the lighting, the texture, and the medium directly, as separate, literal instructions. The tradeoff is obvious on its face — names are fast and specific descriptions are thorough — but the deciding factor in practice is not speed, it is which of the two a given model is actually built to follow reliably.

Where a named style reference actually works

Midjourney is the clearest case of a tool built around named references working well. Its training leans heavily on art-historical and medium-based vocabulary, so terms like "art nouveau," "risograph print," or "tilt-shift photography" reliably shift composition and texture in the expected direction, and the dedicated --sref parameter extends the same idea to an uploaded reference image instead of a word, letting you point at a specific look with a picture rather than a label when no word fully captures it. Named movement and medium terms — styles that describe a historical period, a printing or photographic process, or a broad artistic tradition rather than one living person's personal body of work — are the safest and most portable category of named reference, because they describe a shared visual vocabulary rather than one individual's specific output, and every major image tool has at least some training exposure to them.

Where named references get filtered, blocked, or quietly ignored

Naming a specific living artist is the version of this shortcut that breaks down fastest, and it breaks down differently depending on the tool. DALL-E, generated through ChatGPT, is explicitly built to decline prompts asking for a living artist's distinctive personal style and to redirect toward a more generic art-historical description instead — the request does not error out, it just quietly returns something looser and less specific than what was asked for, which is easy to mistake for the model simply failing to understand the reference. Adobe Firefly takes a more structural version of the same position: it is trained on licensed stock imagery, openly licensed content, and public-domain work rather than a broad scrape of the open web, specifically so that it is not positioned to mimic an identifiable individual artist's style at all, which makes naming one largely a dead end rather than a filtered request. Midjourney itself has moved the same direction over successive versions, with some living-artist names now filtered or quietly down-weighted even though broad movement and medium terms still work exactly as before. The practical result is the same across all three: a prompt built entirely around one named individual's style is the least portable, least durable choice available, and it ages badly as policies tighten, even on the one tool where it currently still works.

Descriptive attributes: slower to write, far more portable

A fully descriptive prompt trades speed for reliability. Instead of "in the style of a 1960s psychedelic rock poster," it spells out the actual attributes that phrase implies: flowing hand-lettered-style shapes, a saturated palette of hot pink, orange, and purple, high-contrast flat color with no gradients, a slight halftone texture. This is more typing, and it requires actually knowing what the named style looks like well enough to describe it rather than just name-dropping it, which is the real cost of this approach — it shifts the thinking from "name a reference" to "actually look at the reference and write down what you see." What it buys back is portability: this list of attributes means roughly the same thing to Midjourney, Stable Diffusion, DALL-E, and Firefly alike, because none of those words depend on the model having memorized one specific named thing, they depend only on the model understanding ordinary visual vocabulary, which every current image model handles well. Stable Diffusion's open ecosystem reflects this split directly: rather than reliably nailing a named style from text alone, the common real-world route to a consistent specific look is a LoRA or fine-tune trained on that exact style, which is effectively someone else doing the descriptive-attribute work once, permanently, instead of you doing it per prompt.

Video tools tilt even further toward attributes

Named style references carry even less reliably into video generation than into image generation, because a video model is already spending most of its capacity on coherent motion and temporal consistency, with comparatively less of its training weight devoted to associating a named aesthetic with a moving result the way an image model associates one with a single frame. Runway, Kling, Sora, and Veo generally respond far more consistently to literal descriptive attributes — a specific film stock grain, a named color grade like teal-and-orange, a specific lens behavior like shallow depth of field with visible bokeh — than to a named directorial or movement reference, which tends to get loosely interpreted or partly dropped once motion is added to the equation. A video prompt built entirely around "in the style of a specific named film" is one of the least reliable things you can ask a video tool for; the same prompt rebuilt as a list of the concrete visual attributes that film is actually known for — grain, grade, framing, pacing — holds up dramatically better across every major video tool.

A hybrid approach that holds up across tools

The most reliable pattern in practice is not choosing one approach exclusively, it is using a named term as an anchor and then backing it up with two or three concrete attributes rather than trusting the name alone to carry the whole look. "Art deco, geometric gold linework, deep navy and black palette, symmetrical composition" uses a single safe, broad movement name as a starting point and then states the specific attributes that make the reference actually land, so that even on a tool where the named term is weighted lightly or ignored, the attribute list still does most of the real work. This hybrid structure is also the most future-proof option, since a named term that currently works but later gets filtered or down-weighted still leaves the attribute list doing its job on its own, where a prompt built entirely around the name has nothing left once the name itself stops functioning.

Common mistakes

Naming a specific living artist and assuming every tool will treat it the same way a broad movement term is treated — some will quietly substitute something generic, others will decline the request outright. Relying on a named style alone in a video prompt, where named aesthetic references carry much less reliably than in image prompts. Assuming a named reference that works well in Midjourney will transfer directly to Stable Diffusion, DALL-E, or Firefly, when each has a different relationship to named-style vocabulary built from a different training approach. And writing a long, fully descriptive prompt but still opening it with a named style the model is likely to ignore or substitute, which wastes the most valuable real estate in the prompt — the first few words — on a term doing less work than the attributes that follow it.

Where Promptima handles this

Whether a named style will actually hold up is not something most people can keep track of tool by tool, and it changes as policies and model versions change underneath it. Promptima's image and video prompt generators handle this translation directly: a rough style idea gets expanded into the literal attributes that the selected tool actually responds to, with a named reference used only where it is reliably supported, so the prompt doesn't quietly fail on a reference that got filtered somewhere between describing the idea and generating the result.