Reference Image Sets: Save a Consistent Look as a Seed

Here's the problem in one before/after.

Weak prompt: a coffee bag on a table, nice lighting

Stronger version, with a saved reference set: @WarmKitchenLook the oat latte bag on a oak counter, morning window light from the left, shallow depth of field, same palette and grain as the references

The first prompt gives you a different mood every time. The second pulls in a saved set of reference images, so the colour, light and texture carry over from one result to the next. That saved set is what Mediafy Studio calls a Seed.

This post covers how to build good reference image sets, save them, and use them without fighting the result.

What a Seed is (and isn't)

A Seed in Mediafy Studio is a saved set of reference images that you can reuse for a consistent look. You build it once, then call it up whenever you want that style again.

It isn't LoRA training. The studio doesn't train custom models. A Seed guides the generation with the images you give it, so the quality of the set matters a lot. Treat it as a mood board the model can see, not a guarantee of identical output.

Why reference image sets beat long style prompts

You can describe a look in words: "muted teal shadows, warm highlights, fine film grain." That works, but words drift. Two generations from the same paragraph can read as different brands.

Reference image sets show the look instead of describing it. They're also faster to reuse, because you aren't pasting a 60-word style block into every prompt. If you want a refresher on writing the words side, see AI Image Prompt Structure: Subject, Setting, Light, Lens.

Build the set: what to pick

Quality beats quantity. A small, tight set usually gives a clearer signal than a big mixed one.

1. Pick one idea per set. A set for "warm kitchen product look" shouldn't also hold your neon night-city shots.
2. Choose images that agree. Same colour temperature, similar contrast, similar level of detail. If two images pull in different directions, the result will blur between them.
3. Vary the subject, keep the look. Include a close-up, a wide shot and a detail shot, all in the same light. That tells the model the look is the point, not the object.
4. Use clean, sharp images. Blurry or heavily compressed references teach the wrong texture. If a favourite is soft, see AI Image Upscaling: What It Fixes and What It Can't first.
5. Cut anything with stray text or logos you don't want echoed back.

A good starting size is a handful of images, not dozens. Add more only if the look still wanders.

Where to get the images

Your best references are often your own earlier results. Generate a few images you love, check them in Assets (which holds every result), and pull the winners into a set. This is how you turn a lucky one-off into a repeatable style.

Save and reuse it in the Studio

The basic flow:

1. Collect your chosen images and save them as a Seed.
2. Give it a name you'll recognise later, like WarmKitchenLook or ClientA_Summer.
3. In the Studio, pick Image or Video, choose a model and start your prompt.
4. Type @ to bring in your saved Seed (the same menu also lists saved characters).
5. Write the rest of the prompt about the new subject only.

Check the credit cost on the Generate button before you run it. A Nano Banana image is 3 credits. If you want the full picture on costs, see How Mediafy Credits Work: Costs, Packs and Refunds.

Prompting with a Seed

Once the look lives in the references, your prompt can shrink. Describe what's new: subject, setting, action. Don't re-describe the style in detail, or you'll fight the references.

Some examples:

• @WarmKitchenLook a ceramic mug of tea beside a folded linen towel, top-down
• @WarmKitchenLook the same bag, now held in a hand, waist-up crop
• @WarmKitchenLook empty counter with space on the right for a headline

If a short prompt feels thin, the Enhance button can rewrite it into a fuller one. Read the result and delete any style words that clash with your Seed. More on that in Prompt Enhancer: Turn a Short Idea Into a Full Prompt.

Common mistakes

• Mixing looks in one set. Five lovely images in three styles give you a muddy average.
• Overstuffing the prompt. If the prompt says "cold blue moody" and the set is warm and golden, one of them loses, and it isn't always the one you wanted.
• Expecting an exact copy. A Seed pushes toward a look. Details such as exact shapes or small props can still shift, so generate a couple of options and pick.
• Using a Seed for a face. If you need the same person across shots, a saved character is the better tool. See Keep a Consistent AI Character Across Images and Video.
• Skipping a test run. Try a cheap draft first to see if the set is doing its job, then switch to a model with more polish for the final.
• Never revising the set. If results drift, swap out the weakest reference rather than adding more words to the prompt.

Using the same look in video

Consistency matters more in video, where a style shift between clips is obvious. A practical approach:

1. Make your hero still using the Seed.
2. Use that still as the start frame for animation. Image to Video: Pick a Start Frame That Animates Well covers how to choose one.
3. Describe the motion in the prompt, not the look. The still already carries it. See AI Video Prompts: Describe the Motion, Not Just the Scene.

For a campaign, make every still from the same Seed, then animate each. Your clips will cut together far more cleanly than ones generated from scratch.

Also match your format early. A set built around wide shots may need new references if you move to vertical. Aspect Ratio for Social Media: 9:16, 1:1, 4:5, 16:9 helps you decide before you generate.

A quick checklist

• One look per Seed, with a clear name.
• A handful of sharp, agreeing images with varied subjects.
• Short prompts that describe the new subject, not the style.
• A cheap draft before the final.
• Swap weak references instead of piling on words.

If you want to go deeper on building a repeatable creative workflow, the AI Creator Lab course is a good next step.

Ready to build your first set? Sign in with Google at /signin and try it in the Studio.