Generative design and neural networks in logo creation
Case: how a startup got a logo in the evening and what had to be redone
A small coffee chain ordered a logo not from an agency but decided to try Midjourney. The first prompt was simple: «logo of a coffee shop, minimalism, brown». The result was a beautiful image with a cup, but with the text «COFFEE», which was distorted with crooked letters, and with an extra background of grains and leaves that nobody had asked for.
After several iterations, the prompt turned into: «minimalist logo mark, coffee bean and cup silhouette combined, warm brown and cream colors, on white background, no text, no letters, flat vector style». They removed the text from the generation entirely — and the quality of the mark immediately improved: the distorted letters disappeared, the form became cleaner, and the symmetry became predictable.
Next, we ran the file (a standard PNG) through Vectorizer.io, cleaned up unnecessary nodes in Illustrator, manually added the text in Cyrillic in the brand's font, and combined everything into a single file with the logo and mark separately. It took about one and a half hours of a freelance designer's work for the refinements. The final cost of the project is several times lower than ordering a logo from an agency from scratch, but without this hour of work, the result could not have been used for signage or in print.
Where neural networks are really good and where they are not
The main strength of generative models is the speed of idea exploration. Where a designer draws three to five concepts per day, a neural network produces dozens of variations in a minute. This is an excellent tool for brainstorming at the start of a project, especially when the client themselves cannot formulate what they want to see.
| What it does well | What it does poorly |
|---|---|
| Generating a large number of concepts and stylistic directions | Text and letters — especially Cyrillic |
| Combining unexpected images (animal + geometry) | Exact symmetry and mathematical precision of form |
| Fast search for color and style references | Vector purity — raster output is always guaranteed |
| Imitation of fashionable visual trends (gradients, 3D, minimalism) | Control over uniqueness — risk of accidental similarity to someone else's logo |
A separate issue — hallucinations. AI may 'invent' non-existent details, produce asymmetrical forms where strict geometry is required, or add an extra element that was not in the prompt. Specialized services like Logo Diffusion offer more control by allowing users to combine their sketch with a text prompt, but they still do not eliminate the need for final cleanup.
Why Cyrillic 'floats' and how to work around it
Models like Midjourney and DALL-E are mainly trained on English-language data, so Latin scripts come out reasonably well, but Cyrillic letters are not understood by the model as symbols — it draws shapes resembling letters rather than text. Hence the distorted 'Ya' and merged 'Sh' on any generated logo with a Russian name.
The working solution is simple and proven: do not ask the neural network for text at all. A direct prohibition is added to the prompt — 'no text, no letters, no typography' — and only the symbol is generated: a pictogram, monogram, or abstract form. The text is added manually on top of the finished symbol in Figma or Illustrator, with font selection tailored to the character of the logo. This is a separate, but short stage of work that completely resolves the issue.
Copyright on a logo created by a neural network
The situation here is less optimistic than one would wish to think. According to Article 1228 of the Civil Code of the Russian Federation, the author of a work is recognized as the citizen whose creative work it was created — a neural network is not a subject of rights and cannot be an author. But even a person who simply entered a prompt risks not receiving full protection: in the ruling of the Presnensky District Court of Moscow 2024 the court indicated that textual commands to the user of the model are technical actions, not a creative contribution. This ruling is under appeal and does not constitute a final precedent.
The same logic was also applied abroad: in 2023 year, the American copyright office (USCO) reclassified the registration of the comic Zarya of the Dawn created with Midjourney: it refused protection for the generated images, but retained protection for the text and composition of the work.
The conclusion is simple: relying on automatic copyright for 'clean' generation is a poor strategy. The reliable path is to register the logo as a trademark at Rospatent. This costs money and time, but provides protection for 10 years and is cheaper than court disputes over proving authorship.
How to check that the logo does not copy someone else's mark
Neural networks are trained on millions of existing images and may accidentally produce a composition similar to an already registered trademark. The responsibility for this lies with the user, not the platform — this is legally stipulated in the user agreements of most services.
- Image search in Yandex and Google Images — a basic but effective check for visual similarity
- Open trademark registry of Rospatent — search by class of goods and services (MKTO)
- Manual analysis of logos of direct competitors in the niche — coincidences often occur precisely there
- With a budget — patent search through a specialized lawyer before filing an application
Full cycle: from prompt to finished file
Working with generative design is not 'pressed a button, got a logo', but a sequence of steps, each of which cannot be skipped if the logo is needed for real business.
- Prompt formulation according to the formula: sign type (minimalist, emblem, monogram) + meaning that should be read + style + specific colors (not 'corporate', but exact shades) + background ("on a white background") + prohibitions ("without text, without letters"). The order is important — if you start with color, rather than sign type, the model will add extra details.
- Generation of 20–40 variants, selection of 3–5 concepts closest to the task.
- Vectorization of a raster file — automatically via Vectorizer.io or Image Trace in Illustrator, or manual tracing for complex forms.
- Cleaning vector nodes, aligning symmetry, eliminating form 'hallucinations'.
- Manual addition of typography — font selection, kerning, adaptation for Cyrillic.
- Checking for similarity with third-party marks and preparing final files: SVG, EPS, PNG in various sizes, black-and-white version.
For small businesses and startups, this approach provides real savings — 70–90% compared to placing a full order with a branding agency. But this is a time-saving measure for the designer, not an excuse to do without one: without final manual refinement, a logo from an AI remains a pretty picture, not a working brand mark.
Case: how a startup got a logo in the evening and what had to be redone
A small coffee chain ordered a logo not from an agency but decided to try Midjourney. The first prompt was simple: «logo of a coffee shop, minimalism, brown». The result was a beautiful image with a cup, but with the text «COFFEE», which was distorted with crooked letters, and with an extra background of grains and leaves that nobody had asked for.
After several iterations, the prompt turned into: «minimalist logo mark, coffee bean and cup silhouette combined, warm brown and cream colors, on white background, no text, no letters, flat vector style». They removed the text from the generation entirely — and the quality of the mark immediately improved: the distorted letters disappeared, the form became cleaner, and the symmetry became predictable.
Next, we ran the file (a standard PNG) through Vectorizer.io, cleaned up unnecessary nodes in Illustrator, manually added the text in Cyrillic in the brand's font, and combined everything into a single file with the logo and mark separately. It took about one and a half hours of a freelance designer's work for the refinements. The final cost of the project is several times lower than ordering a logo from an agency from scratch, but without this hour of work, the result could not have been used for signage or in print.
Where neural networks are really good and where they are not
The main strength of generative models is the speed of idea exploration. Where a designer draws three to five concepts per day, a neural network produces dozens of variations in a minute. This is an excellent tool for brainstorming at the start of a project, especially when the client themselves cannot formulate what they want to see.
| What it does well | What it does poorly |
|---|---|
| Generating a large number of concepts and stylistic directions | Text and letters — especially Cyrillic |
| Combining unexpected images (animal + geometry) | Exact symmetry and mathematical precision of form |
| Fast search for color and style references | Vector purity — raster output is always guaranteed |
| Imitation of fashionable visual trends (gradients, 3D, minimalism) | Control over uniqueness — risk of accidental similarity to someone else's logo |
A separate issue — hallucinations. AI may 'invent' non-existent details, produce asymmetrical forms where strict geometry is required, or add an extra element that was not in the prompt. Specialized services like Logo Diffusion offer more control by allowing users to combine their sketch with a text prompt, but they still do not eliminate the need for final cleanup.
Why Cyrillic 'floats' and how to work around it
Models like Midjourney and DALL-E are mainly trained on English-language data, so Latin scripts come out reasonably well, but Cyrillic letters are not understood by the model as symbols — it draws shapes resembling letters rather than text. Hence the distorted 'Ya' and merged 'Sh' on any generated logo with a Russian name.
The working solution is simple and proven: do not ask the neural network for text at all. A direct prohibition is added to the prompt — 'no text, no letters, no typography' — and only the symbol is generated: a pictogram, monogram, or abstract form. The text is added manually on top of the finished symbol in Figma or Illustrator, with font selection tailored to the character of the logo. This is a separate, but short stage of work that completely resolves the issue.
Copyright on a logo created by a neural network
The situation here is less optimistic than one would wish to think. According to Article 1228 of the Civil Code of the Russian Federation, the author of a work is recognized as the citizen whose creative work it was created — a neural network is not a subject of rights and cannot be an author. But even a person who simply entered a prompt risks not receiving full protection: in the ruling of the Presnensky District Court of Moscow 2024 the court indicated that textual commands to the user of the model are technical actions, not a creative contribution. This ruling is under appeal and does not constitute a final precedent.
The same logic was also applied abroad: in 2023 year, the American copyright office (USCO) reclassified the registration of the comic Zarya of the Dawn created with Midjourney: it refused protection for the generated images, but retained protection for the text and composition of the work.
The conclusion is simple: relying on automatic copyright for 'clean' generation is a poor strategy. The reliable path is to register the logo as a trademark at Rospatent. This costs money and time, but provides protection for 10 years and is cheaper than court disputes over proving authorship.
How to check that the logo does not copy someone else's mark
Neural networks are trained on millions of existing images and may accidentally produce a composition similar to an already registered trademark. The responsibility for this lies with the user, not the platform — this is legally stipulated in the user agreements of most services.
- Image search in Yandex and Google Images — a basic but effective check for visual similarity
- Open trademark registry of Rospatent — search by class of goods and services (MKTO)
- Manual analysis of logos of direct competitors in the niche — coincidences often occur precisely there
- With a budget — patent search through a specialized lawyer before filing an application
Full cycle: from prompt to finished file
Working with generative design is not 'pressed a button, got a logo', but a sequence of steps, each of which cannot be skipped if the logo is needed for real business.
- Prompt formulation according to the formula: sign type (minimalist, emblem, monogram) + meaning that should be read + style + specific colors (not 'corporate', but exact shades) + background ("on a white background") + prohibitions ("without text, without letters"). The order is important — if you start with color, rather than sign type, the model will add extra details.
- Generation of 20–40 variants, selection of 3–5 concepts closest to the task.
- Vectorization of a raster file — automatically via Vectorizer.io or Image Trace in Illustrator, or manual tracing for complex forms.
- Cleaning vector nodes, aligning symmetry, eliminating form 'hallucinations'.
- Manual addition of typography — font selection, kerning, adaptation for Cyrillic.
- Checking for similarity with third-party marks and preparing final files: SVG, EPS, PNG in various sizes, black-and-white version.
For small businesses and startups, this approach provides real savings — 70–90% compared to placing a full order with a branding agency. But this is a time-saving measure for the designer, not an excuse to do without one: without final manual refinement, a logo from an AI remains a pretty picture, not a working brand mark.
Recent articles
MegaFon, BMW, VTB and more 240+ companies. See how it looks in the showcase reel:
Our branding agency is approached to order:
Branding agency services
