You found the look. Flat vectors with soft pastels. A hand-drawn feel with bold outlines. A style you saw on Dribbble and wanted across your whole brand.
Then you tried to reproduce it with an AI image generator. Prompt after prompt, close but never the same. You tweaked words, added adjectives, got something in the neighborhood, and moved on because you ran out of patience. By the end you had spent more time chasing the style than designing the thing yourself.
The fix is not a better description. It is a style you capture once and paste every time. Screenshot the style, have the AI turn it into a JSON template, then paste that template under every new image request. Same colors, same textures, same feel. Different subject, identical look. That is the entire method, and the rest of this guide is how to run it.
What You Need
One tool: Google Gemini, at gemini.google.com. It does both jobs here, reading the style out of your screenshot and generating the new images. The free tier is enough to run this whole workflow. You do not need a paid plan to start.
Everything downstream depends on one artifact: a JSON style template. It is a small block of text that describes your look attribute by attribute. Colors as hex codes. Line style. Shape language. Texture. Shadow behavior. Mood. Once you have it, every image is a thirty-second job.
The Method In Four Steps
Here is the shape before the detail.
| Step | What happens | Tool |
|---|---|---|
| 1 | Screenshot any illustration style you want | Your phone or computer |
| 2 | Upload it to Gemini and extract a JSON style template | Google Gemini |
| 3 | Describe a new subject and paste the template | Google Gemini |
| 4 | Get a style-matched image, then repeat forever | Google Gemini |
The one-time cost is Steps 1 and 2. After that, Steps 3 and 4 are all you ever touch.
Step 1. Screenshot The Style You Want
Find a style worth copying. A social graphic you admire. A landing-page illustration. A book cover. A Dribbble shot. An icon set. A character design. The only rule is that it clearly shows the visual language you are after.
Take a clean screenshot. Four things to check before you move on:
- The image is sharp, not blurry or heavily compressed.
- It shows the style plainly: colors, shapes, textures, line work.
- It is a single illustration, not a collage. Gemini reads one style at a time.
- It has enough detail for the model to see the full aesthetic, not a tiny thumbnail.
Some styles carry over cleanly. Some do not. Here is the difference.
| Style | What makes it read | Reference quality |
|---|---|---|
| Flat vector | Clean lines, solid colors, geometric shapes | Strong |
| Hand-drawn / textured | Visible brush strokes, paper grain, imperfections | Strong |
| 3D rendered | Soft shadows, glossy surfaces, depth | Strong |
| Isometric | 30-degree angles, no vanishing perspective | Strong |
| Minimalist line art | Single-weight strokes, lots of white space | Strong |
| Retro / vintage | Limited palette, halftone dots, aged texture | Strong |
| Generic stock illustration | Nothing distinctive to lock onto | Weak |
The rule underneath the table: the more distinctive the style, the better the result. A generic stock look gives you a generic template, and a generic template gives you generic output. Pick something with an opinion.
Step 2. Turn The Screenshot Into A Style Template
This is where the whole thing gets built. You are asking Gemini to look at your image and write down its style in machine-readable form.
Do this:
- Go to gemini.google.com and sign in with your Google account.
- Click the image upload icon, or drag your screenshot straight into the chat.
- Once it is uploaded, type the prompt below exactly.
- Press Enter and wait for the response.
Copy The Analysis Prompt
Copy this.
Make a JSON prompt for this exact style.
That is the whole prompt. Short on purpose. You are not describing the style yourself, which is the mistake that started this problem. You are making the model read the style off the pixels.
Gemini walks through every visual attribute in your screenshot. The color palette, down to hex codes. The shape language, rounded against angular, geometric against organic. Line weight. Texture and surface. Lighting direction and shadow behavior. Composition. Overall mood. Then it packs all of it into one structured block. Think of it as the style DNA of your illustration.
What You Get Back
A JSON block. The exact contents depend on your image, but the shape looks like this:
Copy this.
{
"style": "flat vector illustration",
"color_palette": {
"primary": "#4A90D9",
"secondary": "#F5A623",
"background": "#FFF8F0",
"accents": ["#E74C3C", "#2ECC71"]
},
"line_style": "no outlines, shape-defined edges",
"shapes": "rounded, soft geometric forms",
"texture": "subtle grain overlay, paper-like",
"shadows": "flat, offset, no gradients",
"proportions": "slightly exaggerated, friendly",
"mood": "warm, approachable, modern",
"rendering": "clean, minimal detail, bold shapes"
}
This is your template. Copy it and save it somewhere permanent. A Notion page, a Google Doc, a plain text file, wherever you keep reusable assets. You are going to paste it into every image you generate from now on, so treat it like a brand asset, because that is what it is.
Done. The style is captured. You never have to describe this look in words again.
Step 3. Generate A New Image In The Same Style
Now the payoff. You describe a subject, paste the template, and let the two combine.
One thing matters here more than anything else: start a new chat. Open a fresh Gemini conversation for each image. The old chat still holds the previous picture and the previous words, and Gemini will let that context bleed into the next result. A clean chat plus the template is the reliable combination.
Then:
- Describe what you want. The subject only, not the style.
- Paste your JSON template underneath the description.
- Press Enter.
Copy The Generation Prompt
Copy this.
Generate an illustration of a person sitting at a desk
working on a laptop with a coffee cup beside them.
Use this exact style:
[PASTE YOUR JSON TEMPLATE HERE]
Swap the first sentence for whatever you need. The pattern holds for anything:
Copy this.
Generate an illustration of a smartphone showing a
dashboard with charts and notifications.
Use this exact style:
[PASTE YOUR JSON TEMPLATE HERE]
Copy this.
Generate an illustration of a team of three people
meeting around a whiteboard covered in sticky notes.
Use this exact style:
[PASTE YOUR JSON TEMPLATE HERE]
You are always writing the same two parts. A plain-language subject on top. The unchanged template below it. The subject moves. The style never does.
Step 4. Repeat, And Get The Same Look Every Time
Every new image is a subject line plus a paste. Same colors, same shapes, same textures, same mood. New subject, identical aesthetic, on demand.
It works because the template leaves nothing to interpretation. When you type "make it look flat and modern with pastel colors," you hand the model a vague instruction, and it guesses differently each time. When you hand it #F5A623 and no outlines, shape-defined edges, flat offset shadows, no gradients, there is nothing left to guess. Every attribute is pinned.
This is the same discipline a real brand system runs on. A brand guideline does not say "use blue." It says "use #0066CC at 100% opacity on white." The template is a brand guideline for AI image generation, and it enforces itself.
The Mistake That Makes It Fail
Generating several images in one long chat.
It feels efficient. It quietly breaks the whole thing. As the conversation grows, Gemini starts blending your template with everything else it has seen in that thread, and the style drifts one image at a time until image five no longer matches image one. You will not always notice until you line them up.
The fix costs nothing. New chat, paste the template, generate, close it. New chat again for the next one. One image per conversation keeps the output clean, and the template is right there in your saved file, so pasting it takes two seconds.
Five Ways To Push It Further
Once the base workflow is running, these are the moves that make it worth building on.
Refine the template instead of restarting. If the first image is close but off, do not throw it out. Tell Gemini what to change: "the shadows should be softer," or "the colors are too saturated." Then ask it to update the JSON with those corrections and save the new version as your master. Your template gets more accurate the more you use it.
Keep one template per use case. Blog illustrations, social graphics, and icon sets often want slightly different looks. Build a separate template for each, name them clearly, and store them together. You are building a small library, not one file.
Bake subject rules into the template. If you always draw people, add a character section: "proportions: slightly exaggerated, friendly faces, simple features, no detailed expressions." If you always draw product shots, add lighting and perspective rules. The template can carry more than style.
Reuse the template across other tools. The JSON is not locked to Gemini. Paste it into Midjourney, DALL-E, Stable Diffusion, or any generator that reads text prompts. It is just a precise description of a look, and a precise description travels.
Build a template per client. If you make visuals for other businesses, store a style template beside each client's brand guidelines. When they need new graphics, paste their template, describe the subject, and every deliverable matches their brand without a fresh design pass. Consistent branding, generated in seconds, is a service you can charge for.
If you want the same idea applied to video instead of stills, the reference-image method for AI video runs on the same logic: lock the look once, let it drive every frame.
Honest Limits You Should Know
This is a consistency tool, not a magic wand. A few things are true and worth saying out loud.
The first template is rarely perfect. Expect one or two rounds of refinement before it nails your style, and treat that as normal, not failure.
Very intricate or highly rendered styles are harder to pin down than flat, graphic ones. The more subtle the style, the more manual correction you will do.
Gemini's free tier has daily generation limits. For a normal week of brand visuals it is plenty. If you are producing at volume, you will hit the ceiling and want a paid plan.
And it does not replace a designer for work that needs real craft or originality. It replaces the hours you were losing to inconsistency. That is the job it does, and it does it well.
If You Do One Thing This Week
Pick the one style you reach for most. Screenshot it, run the analysis prompt, and save the template. That single file is what turns "close enough" into "the same every time," and it takes about five minutes to make. Do not build the whole library yet. Build one template, generate three images from it, and see the match for yourself. If it does not hold up, you spent five minutes. If it does, you never fight for a consistent look again.