How to make a logo with AI: a prompt, an SVG and a trademark check

Before you start: what a logo file has to do

A logo gets used at the size of a browser tab and at the size of a shop sign, on white paper and on a dark web page. A picture that looks fine in a chat window can fail on all four. A logo prompt lists what a logo needs: flat shapes, few colors, a clear outline and no background.

An image model draws pixels. A raster file such as PNG stores a grid of colored dots, and the dots blur when you enlarge them. A vector file such as SVG stores shapes and colors as instructions, so it stays sharp at any size and you can edit its paths. OpenAI's image generation guide says to use SVG "for graphics that need editable paths, shapes, and colors, such as icons and illustrations".

You need an app that draws pictures from a text prompt. The prompt below is from OpenAI's image prompting guide, written for developers who call its GPT Image models. It is plain English, so you can paste it into other image apps too, and the results will differ. Settings such as size, quality and background are fields in OpenAI's API. In a chat app, say the same thing in words.

Write the logo prompt

OpenAI's guide has one logo prompt, for a bakery called Field & Flour. Swap in your name and your business and keep the rest.

The sentence before the prompt in the guide tells you what to ask for.

Create an original, non-infringing logo for a company called Field & Flour, a local bakery.
The logo should feel warm, simple, and timeless. Use clean, vector-like shapes, a strong silhouette, and balanced negative space.
Favor simplicity over detail so it reads clearly at small and large sizes. Flat design, minimal strokes, no gradients unless essential.
Fully transparent background. Deliver a single centered logo with generous padding, clean alpha edges, and no solid backdrop, scenery, checkerboard, or watermark.

• "original, non-infringing" asks the model not to copy a known mark. It is a request, and no prompt can search a trademark register.
• "clean, vector-like shapes, a strong silhouette, and balanced negative space" pushes toward bold forms that survive shrinking. Negative space is the empty area around and inside the shapes.
• "Favor simplicity over detail", with flat design, minimal strokes and no gradients, removes what a small print or a one-color stamp would lose first.
• "Fully transparent background" and the ban on a checkerboard matter because OpenAI's guide says a drawn checkerboard is not transparency.

1. Write a brand line: the name, what the business does and three feeling words. The sample uses warm, simple and timeless.
2. Put them into the prompt and ask for four versions. In OpenAI's complete example the API call sets n=4 and the script saves four PNG files.
3. Put the four on a white page and on a dark page. A mark that vanishes on one of them needs a fix, and the fix is a new color, not a background box.
4. Look at each candidate at browser-tab size. If you cannot tell what it shows, drop it.
5. Change one thing per follow-up, such as one color or the layout of the name, and repeat the details you want to keep. OpenAI's guide asks for one change at a time and the details to preserve.
6. Read the name letter by letter. Lettering is where image models slip, and the next lesson goes into it.

The steps after the prompt are our own routine. The prompt and the one-change rule are OpenAI's.

Turn the picture into a vector

The PNG from a chat app is a raster file. The table below sets out three routes to a logo file. The transparent PNG sits on any background but still blurs when enlarged. The other two routes give an SVG, which you can scale and recolor, and each gives up something different.

For developers, OpenAI's guide shows the call below, which writes four transparent PNG files. The model and prompt variables are defined earlier in its script, as the Field & Flour prompt above.

For everyone else, VTracer is an open-source tool that converts raster images such as JPG and PNG into SVG. Its README lists apps for Windows, macOS and Linux and a command-line program, and labels the current release 1.0.0-alpha.4. The three commands below are from the README. --preset bw suits black-and-white line art, and --palette snaps the output to the hex colors you list, which fits a logo with two or three brand colors. Two more options, --max-colors and --simplify, cap the number of colors and fit smoother curves, and the README lists curve simplification as often half the file size.

import base64
from pathlib import Path

from openai import OpenAI

client = OpenAI()

logos = client.images.generate(
    model=model,
    prompt=prompt,
    size="1024x1536",
    quality="medium",
    background="transparent",
    output_format="png",
    n=4,
)
for index, image in enumerate(logos.data, start=1):
    Path(f"logo-generation-{index}.png").write_bytes(base64.b64decode(image.b64_json))

• Start from the largest PNG the app will give you. A trace follows the pixels it is handed, so a small or blurry source gives ragged paths.
• Open the SVG in a browser and zoom in far. Ragged edges or specks mean the trace needs another try with different options.
• Keep the transparent PNG as well. Some websites and apps still ask for a raster file.

Check the name and the mark before you use them

Wikipedia's Trademark article defines a trademark as a word, phrase, symbol, design or a combination of them that identifies a product or service from a particular source. A trademark check means looking for existing marks that could be confused with yours. Infringement is the use of "the same or a confusingly similar trademark for the same or similar products", so a similar logo in a different field matters less than one next door to yours.

No image model runs that search. The word "non-infringing" in a prompt does not change that, and a logo that looks original to you can still sit close to a registered one. Wikipedia names the registers to search: the US Patent and Trademark Office keeps a public database, WIPO's Global Brand Database covers international marks and the EUIPO has an eSearch plus tool for EU marks. Search the name, any word inside the mark and the shape, in your own field.

Rights can come from registering a mark and, in some countries, from using it in business, so a name that someone used before you can hold the senior rights.

Copyright is a separate question. Wikipedia's summary of AI and copyright says US courts have declined to grant copyright to works generated without human intervention. It adds that the Copyright Office's January 2025 clarification treats works "where the expressive elements are determined by a machine" as uncopyrightable.

Other countries, the UK and China among them, have their own rules. Choosing, combining and editing the artwork yourself leaves a person's work in it. For a brand you plan to grow, a trademark attorney in your country can say what applies. This lesson is not legal advice.

Common mistakes

• Asking for a scene, a photo effect or a gradient. A logo with a background picture loses it the moment it goes on a stamp or an embroidered cap.
• Judging the candidates at chat size only. The browser-tab size shows which ones hold up.
• Accepting a wordmark with one wrong letter because the symbol is good.
• Reading "non-infringing" in the prompt as a clearance. It clears nothing.
• Tracing a screenshot of the logo. Export the largest file the app offers instead.
• Using the name on products before searching the register for it.

Try it yourself

Pick a business you know. Your own works, and so does a friend's. The aim is a logo file and a note on what you searched.

• Write the brand line and generate four versions with the Field & Flour prompt.
• Check them on a white page, a dark page and at browser-tab size, and keep one.
• Make two follow-ups that each change one thing, and keep the best result.
• Convert the winner with VTracer and zoom in on the edges of the SVG.
• Search the name in your country's trademark register and write down what you found, even if it is nothing.

You are done when the SVG looks right at every size, the name is spelled correctly and the search note is saved with the file.

Next: put words in an image and have them spell right

A logo is one picture with a few words. Posters, covers and ads carry a lot more, and that is where models misspell. The next lesson compares GPT Image, Gemini, Qwen-Image and Ideogram on text.

The prompt habits behind this lesson are in AI image prompts, and the AI image generator comparison lists more apps to try.

Frequently asked questions

Sources

1. OpenAI API docs: Image prompting (read November 6, 2026)https://developers.openai.com/api/docs/guides/image-prompting
2. OpenAI API docs: Image generation (read November 6, 2026)https://developers.openai.com/api/docs/guides/image-generation
3. GitHub: visioncortex/vtracer (README) (read November 6, 2026)https://github.com/visioncortex/vtracer
4. Wikipedia: Trademark (read November 6, 2026)https://en.wikipedia.org/wiki/Trademark
5. Wikipedia: Artificial intelligence and copyright (read November 6, 2026)https://en.wikipedia.org/wiki/Artificial_intelligence_and_copyright