What Is a Style Reference? How AI Uses Visual Anchors
A style reference tells an AI model how output should look, not what it shows. Learn the types, use cases, limits, and how to make one work across tools.
A style reference is the fastest way to tell an AI model how you want something to look without writing a paragraph of adjectives. You hand the model an image, or a structured description, and it borrows the palette, lighting, texture, and mood while leaving the subject up to your prompt. Every serious image and video tool now supports some version of it. Most people still use it badly, because they treat a style reference as a magic picture instead of what it actually is: a constraint.
This guide covers what a style reference is, how it differs from a content reference, the types you can use, where it works, where it breaks, and how to make one that survives beyond a single tool.
What is a style reference?
Definition: A style reference is an input — usually an image, sometimes a structured text specification — that guides an AI model's aesthetic choices, such as color, lighting, texture, composition, and mood, without dictating the subject. It separates how something looks from what it depicts, so one visual language can apply to many different outputs.
You will also see it called an aesthetic reference, a look reference, or a mood reference. In film and design, the older term is a reference frame or a mood board — the same idea, aimed at a human crew instead of a model.
The core job is transfer. A style reference encodes a set of visual decisions, and the model applies those decisions to new content. A 1970s Kodachrome still can turn a prompt for "a cyclist on a coastal road" into a warm, grainy, slightly faded frame — without putting the original photo's subject into the image.
Key properties of a good style reference
- Isolated — it expresses a style, not a busy scene full of competing signals
- Consistent — every element in it points in the same aesthetic direction
- Reusable — it works across many subjects, not only the one it was made for
- Explicit — the traits you care about (palette, grain, light direction) are visible or written down
How a style reference works in AI generation
Image models turn a reference image into an embedding — a numeric summary of its visual features. With a style reference, the tool weights the parts of that embedding tied to surface qualities (color distribution, contrast, brushwork, grain) and down-weights the parts tied to objects and layout. Your text prompt supplies the subject; the reference supplies the look.
Most tools expose a strength or weight control. Low weight gives a hint of the style. High weight starts leaking the reference's content into your image. The name and mechanism differ by tool:
| Tool | How you give it a style reference |
|---|---|
| ChatGPT / GPT image models | Attach an image and say "match the style of this image, not its content" |
| Gemini | Attach one or more reference images alongside the prompt |
| FLUX | Redux or IP-Adapter image conditioning with a strength setting |
| Stable Diffusion / ComfyUI | IP-Adapter style models, or a style LoRA |
| Midjourney | The --sref parameter, with --sw for weight (full guide) |
| Claude and other text models | Paste a written style spec; there is no image-weighting control |
Every one of these interprets the same image a little differently, which matters once you work in more than one tool.
Style reference vs content reference vs prompt
A style reference is one of three ways to steer a model. Mixing them up is the most common reason outputs drift.
| Input | What it controls | What it ignores | Best for |
|---|---|---|---|
| Style reference | Palette, lighting, texture, mood, rendering | Subject, pose, identity | Applying one look across many subjects |
| Content / character reference | Subject, identity, layout, pose | Rendering style | Keeping the same person, product, or object |
| Text style prompt | Whatever you describe, loosely | Anything you forget to mention | Quick one-off direction |
| Structured style spec | Every named attribute, explicitly | Nothing it lists | Consistency across sessions, people, and tools |
A useful analogy: a content reference says paint this; a style reference says paint it like this. A text prompt describing style says paint it roughly like this, from memory — which is why prompt-only style drifts from one generation to the next.
Types of style references
Style is not one thing. A reference can carry several dimensions, and the best results come from knowing which one you actually want.
Color and palette references
Borrow a color scheme — a limited palette, a specific grade, a duotone. Useful for brand work, where exact hues matter more than rendering technique.
Lighting references
Carry light direction, quality, and temperature: hard noon sun, soft window light, neon rim light. Lighting often does more for mood than any other attribute.
Texture and medium references
Film grain, watercolor bleed, risograph misregistration, 3D clay. These define the medium the image appears to be made in.
Composition references
Framing, negative space, camera height, lens feel. Many tools transfer composition weakly from a style image, so this dimension is usually better written down than implied.
Mood and atmosphere references
The hardest to pin down and the easiest to lose — melancholy, playful, clinical. Mood is the sum of the other dimensions, which is why a single image often captures it better than words.
Voice references (for text)
Style references are not only visual. A few paragraphs of your writing given to ChatGPT or Claude work as a style reference for tone and rhythm. See how to define an AI brand voice for the text side.
Multi-reference workflows
Many tools accept several references at once. Use one reference per dimension — one for palette, one for lighting — rather than several images that each carry everything. Two to three focused references beat five mixed ones.
Common style reference use cases
- Brand content — keep social posts, ads, and blog imagery in one visual language across months of production.
- Film and video pre-production — lock a grade and lighting look across storyboards and AI-generated shots.
- Advertising campaigns — produce variants for many placements without the campaign look fragmenting.
- Character and illustration series — pair a style reference with a character reference so a cast stays on-model and on-style.
- Art direction — hand a team, or an agent, a single source of truth. More on this in AI art direction.
This is exactly the problem StyleRef solves — build your style spec in 60 seconds →
The limits of image-only style references
An image reference is powerful and vague at the same time. Three problems show up quickly:
- It doesn't travel. A reference tuned in one tool behaves differently in the next, and a tool-specific code or LoRA doesn't carry over at all. Each tool interprets the same image differently.
- It is implicit. The model decides which traits in the image matter. You can't tell it "keep the grain, ignore the teal."
- It carries content. Push the weight too high and objects from the reference appear in your output.
The fix is to make the style explicit. Describe each dimension — palette with hex values, lighting, texture, composition, mood, and what to avoid — in a structured specification. That text works in every tool that accepts a prompt, and you can still attach the image where the tool supports it. We cover the per-tool approach in consistent style in ChatGPT and consistent style in FLUX.
How to create a style reference that works everywhere
Direct answer: Pick one or two images that express the look you want, then write down each style dimension explicitly — palette, lighting, texture, composition, mood, and exclusions. Use the images where a tool accepts them and the written spec everywhere else, with the same wording every time.
Step by step:
- Collect 3–6 candidate images that share the look, then cut to the one or two most consistent.
- Name the dimensions you want transferred. Skip the ones you don't care about.
- Write each one concretely — "warm tungsten key from camera left, deep shadows" rather than "moody lighting."
- List exclusions — "no lens flare, no pure black, no neon."
- Test on three unrelated subjects. If the style holds across a portrait, a product, and a landscape, the reference is working.
- Store it in one place so you paste the same version every time.
StyleRef automates steps 2 to 4: upload an image, extract a structured style, edit it, and export it as a portable spec or a STYLE.md file. Browse the gallery to see finished examples across 20 disciplines.
Frequently asked questions
How does an AI model use a style reference image?
The model encodes the image into features and weights the ones tied to surface appearance — color, contrast, texture, lighting — over those tied to objects. Your text prompt then supplies the subject, and the model renders it with the borrowed look. A weight parameter controls how strongly the reference applies.
What makes a good style reference image?
A good reference has one clear, consistent aesthetic and a simple subject that won't leak into outputs. Avoid collages, busy scenes, and images mixing several styles. If you can't describe the style in a sentence, the model will struggle to isolate it.
Can you use a film still as a style reference?
Yes. Film stills make strong references for grade, lighting, and framing. Pick a frame without recognizable faces or logos, so the model borrows the look rather than the content, and respect the rights of the source for commercial work.
How many style references should you use at once?
One to three. Each additional reference dilutes the others, and conflicting references average into something generic. Assign each reference a single job, such as palette or lighting, when a tool accepts several.
How do style references relate to LoRA models?
A LoRA is a small fine-tune trained on many images, so it bakes a style into the model's weights. A style reference applies a style at generation time with no training. LoRAs are stronger and more consistent within one model; style references are faster, cheaper, and easier to change.
Can style references be used in video generation?
Yes. Most AI video tools accept a reference image or a styled first frame, and a written style spec in the prompt helps keep the look stable from shot to shot. Consistency across separate clips is harder than within one, so an explicit spec matters more for video.
What is the difference between a style reference and a prompt describing a style?
An image reference shows the style; a prompt describes it. Images capture nuance but are interpreted differently by each tool. Loose prompts drift. A structured written spec sits between the two — explicit, reusable, and portable across every AI tool.
Where should you store style references?
Keep them in one versioned place rather than scattered across chat histories and Discord threads. A StyleRef stores the source images and the extracted spec together, so you and your team paste the same style every time.