Using Reference Images to Steer Style and Composition

A common myth: if you upload a reference image, the model copies it. Drop in a photo you love, type a short prompt, and you get that look back.

It doesn't work like that. A reference image is more like a pile of evidence. The model sees a face, a color palette, a pose, a camera angle, a mood and a background, and your prompt has to say which of those matter. Without that, it guesses. Sometimes it guesses well. Often you get the subject from one image, the lighting from nowhere, and a composition that matches neither.

This post covers how to use reference images ai tools accept, how to write the prompt that goes with them, and where the technique falls short.

What the model makers say about references

I read the documentation from a few model makers, and the pattern is the same across them.

Google's Gemini API image generation guide says its Nano Banana models can mix up to 14 reference images in one request, and its main example combines several photos into one group shot. The page lists different limits by model, and it lists style references only for one of them (Gemini 3 Pro Image, up to 3). So don't assume every model treats a style reference the same way. Its examples also show two habits worth copying: describe the composition you want, and when editing, say what must stay unchanged.

OpenAI's image generation guide describes using other images "as a reference" to generate new ones, with an example that combines four input images into one gift basket picture. It also notes that if you give multiple images and a mask, the mask applies to the first image. Image order matters in practice, not just in theory. The page also describes iterative edits across turns, so you refine rather than start over.

ByteDance's Seedream 4.0 page says you can upload multiple reference images and generate several outputs at once, and shows examples of turning photos into styles like watercolor and cyberpunk. It doesn't state a reference limit or any caveats, so treat it as a capability claim, not a guarantee.

None of these pages promise that a reference will be reproduced exactly. That's the honest baseline.

Decide the job of each reference

Before you upload anything, write one line per image: what is this for? Most references do one of four jobs.

• Subject: who or what appears (a product, a person, a pet).
• Style: the rendering look: film grain, watercolor, flat vector, glossy 3D.
• Composition: where things sit in the frame, camera height, crop, depth.
• Mood or palette: colors and lighting, separate from the subject.

The mistake people make is uploading one image and hoping it covers all four. A moody street photo carries a style, a layout, a palette and a person. The model can't know you only wanted the layout.

Pick images that are strong at one job. If you want the composition of a photo but not its colors, a rough sketch or a simple screenshot of the layout can work better than the polished original.

Write the prompt around the references

Refer to each image by its role or its position, and say what to take from it. A pattern that works:

1. Name each reference (image 1, image 2).
2. Say what to borrow from each.
3. Describe the new scene.
4. State what to ignore or keep.

For example:

Use image 1 for the bottle and its label exactly. Use image 2 only for the soft window light and muted green palette. Place the bottle on a pale oak table, centered, shot at eye level with the background slightly blurred. Do not copy the objects from image 2.

Another, for style:

Redraw my product photo (image 1) in the flat vector illustration style of image 2. Keep the layout and the product shape. Use only three colors from image 2.

Note the pattern: it says what to keep, what to take, and what to leave behind. The ordering matters most when you attach more than one image, so keep the same order in your prompt as in your upload.

If you're unsure how to phrase the rest of the scene, AI Image Prompt Structure: Subject, Setting, Light, Lens gives a frame to fill in. Mediafy Chat can also help draft the prompt, and the Enhance button in the Studio expands a short prompt into a fuller one. Check the result before you run it, because an expanded prompt can drift away from your reference's role.

Steering composition specifically

Composition is where references most often disappoint, because models treat layout as soft guidance. A few ways to make it stick:

• Say it in words as well. "Subject on the left third, camera low, empty space on the right for a headline" beats relying on the image alone.
• Use a simple layout reference. A blocky sketch or a grayscale mockup carries position without dragging in a style.
• Match the aspect ratio. If the reference is square and your target is 9:16, the model has to reinvent the layout. Set the ratio first. Aspect Ratio for Social Media: 9:16, 1:1, 4:5, 16:9 covers the common sizes.
• Change one thing per round. Google's guide recommends refining through follow-up prompts. If the layout is right and the color is wrong, only ask for the color.

For product work, where the object has to look like the real thing, see AI Product Photography: Packshots, Lifestyle Shots, Props.

Steering style specifically

Style transfer is easier when you describe the style in words too. "In the style of image 2" is vague. "Soft gouache texture, visible brush edges, limited warm palette" gives the model something to hold. If the reference has a distinctive look the model doesn't recognize, your own description fills the gap.

Be careful with references that include recognizable people, brands or an artist's signature look. Using someone else's work as a style reference can raise copyright and likeness questions. Use your own images, or ones you have the right to use, particularly for commercial work.

Mistakes that waste credits

• Too many references. Each extra image dilutes the others. Start with one or two, and add more only when something is missing.
• Conflicting references. A bright, flat illustration and a dark, grainy photo will fight. Expect a muddy average.
• Low-quality references. Blurry, heavily compressed or watermarked images pass those flaws on. Clean up or upscale first; AI Image Upscaling: What It Fixes and What It Can't explains the limits.
• No instruction at all. An image with a one-word prompt gives the model free rein.
• Expecting exact copies. Faces, logos and small text can shift between runs. If text is the point, see How to Get Readable Text Inside AI Images.
• Changing everything per attempt. If you swap the reference, the prompt and the model in one go, you won't learn what helped.

Results also vary from run to run, even with identical inputs. Generate a few variations before deciding a reference "doesn't work."

Using references in Mediafy Studio

In the Studio you pick Image, choose a model (models such as Seedream 5 Pro, GPT Image 2 and Nano Banana 2 are offered), write your prompt, and check the credit cost on the Generate button before you run it. Cheaper models are a sensible place to test a reference idea before you spend on a final version. If you reuse the same set of references often, a Seed saves them as a set so you can type @ to bring it in; there's no custom model training involved. For repeat characters, Keep a Consistent AI Character Across Images and Video goes further.

Different models handle references differently, as the documentation above shows. Run the same reference and prompt through two models and compare. Nano Banana 2 Explained: What Google's Model Does Well and GPT Image for Design Work: Layouts, Text and Edits cover two of them.

Once you have a still you like, you can use it as the starting point for video: Image to Video: Pick a Start Frame That Animates Well.

A quick workflow to try

1. Choose one reference and write down its single job.
2. Set the aspect ratio of your final image.
3. Write a prompt that names the reference, what to take, and what to leave.
4. Generate a few variations.
5. Pick the closest, then change one thing at a time.

If you'd like to try it, you can sign in with a Google account at /signin.

Sources

• Google AI for Developers, "Image generation with Gemini (aka Nano Banana)": https://ai.google.dev/gemini-api/docs/image-generation
• OpenAI, "Image generation" (API guide): https://developers.openai.com/api/docs/guides/image-generation
• ByteDance Seed, "Seedream 4.0": https://seed.bytedance.com/en/seedream4_0