AI Generates Web Design References in Five Styles, from Fintech to Web 1.0

Habr has published an experiment in generating web design assets with AI. The author picks five themes: gaming, fintech, Web 1.0 pastiche, ecology and AI services. The announcement does not say which models or prompts were used, so it is too early to judge the quality of the results. But the choice of themes shows that the test targets the stylistic flexibility of generative models, not their ability to produce a single good-looking picture. - Five themes, five different visual languages. - The task: references and assets, not finished layouts. - The main question: how much time this saves a designer.

AI-processed from Habr AI; edited by Hamidun News
AI Generates Web Design References in Five Styles, from Fintech to Web 1.0
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1Five themes, five different visual languages.
The taskreferences and assets, not finished layouts.
The main questionhow much time this saves a designer.

An experiment in generating web design references with AI has appeared on Habr. The author creates assets for five themes: gaming, fintech, Web 1.0 pastiche, ecology and AI services. The post is presented as a first attempt, and the word "conditional" appears right in the announcement. That is an honest framing: this is material for inspiration and moodboards, not a replacement for a designer.

The public part of the post contains no model names, versions or prompt texts. So below we do not review the author's results. Instead, we look at what this particular set of themes tests and how people usually work with it.

Why These Five Styles Were Chosen

The set is built to pull the most varied visual languages out of a model. Each theme brings its own vocabulary of shapes, color and typography:

  • Gaming. Dark backgrounds, neon or fantasy textures, aggressive accents, large hero images.
  • Fintech. A restrained palette, plenty of white space, cards, charts, and a sense of trust and order.
  • Web 1.0. Borders, gradient buttons, cramped tables, default fonts. The model has to imitate an aesthetic that nobody seriously designs in today.
  • Ecology. Natural textures, soft green and earth tones, organic shapes, photographs instead of illustrations.
  • AI services. Gradients, glows, abstract "neural" motifs. Here the model risks producing a cliché we have already seen hundreds of times.

The most interesting pair is Web 1.0 and AI services. The first tests how accurately a model can imitate a style. The second shows whether it can move away from the most overused image of AI.

Where Generative Models Usually Stumble

Web design is harder than illustration for three reasons. First, text. Headlines, button labels and menus in generated images often turn into pseudo-letters. Second, the grid. The model draws something "that looks like a website," but alignment, spacing and hierarchy often fall apart on closer inspection. Third, consistency. It is hard to turn five images from one project into a unified set with shared colors, corner radii and icons.

That is why it makes more sense to treat such results as raw material. You can take a composition, a color scheme or a mood, and then build the text, grid and components by hand in Figma or in code.

How to Apply This in Your Work

If you want to repeat the experiment, start with a prompt formula: theme, page type (landing page, dashboard, catalog), mood, palette, and a reference era or genre. Ask for individual blocks, such as a hero section or a pricing card, rather than a whole site at once. This makes it easier to control the details.

Next, lock in the parameters that worked and reuse them across all variants: the same format, the same description of lighting and palette. This gives you a better chance of getting a consistent style.

One last point. Judge the process by time, not by how good the picture looks. If you get ten directions to discuss with a client in 20 minutes, the tool has paid off, even if none of the layouts is fit for production. For the full results, the tools used and the author's prompts, it is best to check the original on Habr: that is where it will become clear how convincing the result is in each of the five themes.

Key Questions Answered

Why These Five Styles Were Chosen?

The set is built to pull the most varied visual languages out of a model. Each theme brings its own vocabulary of shapes, color and typography: Gaming. Dark backgrounds, neon or fantasy textures, aggressive accents, large hero images. Fintech. A restrained palette, plenty of white space, cards, charts, and a sense of trust and order. Web 1.0. Borders, gradient buttons, cramped tables, default fonts. The model has to imitate an aesthetic that nobody seriously designs in today. Ecology. Natural textures, soft green and earth tones, organic shapes, photographs instead of illus

Where Generative Models Usually Stumble?

Web design is harder than illustration for three reasons. First, text. Headlines, button labels and menus in generated images often turn into pseudo-letters. Second, the grid. The model draws something "that looks like a website," but alignment, spacing and hierarchy often fall apart on closer inspection. Third, consistency. It is hard to turn five images from one project into a unified set with shared colors, corner radii and icons. That is why it makes more sense to treat such results as raw material. You can take a composition, a color scheme or a mood, and then build th

How to Apply This in Your Work?

If you want to repeat the experiment, start with a prompt formula: theme, page type (landing page, dashboard, catalog), mood, palette, and a reference era or genre. Ask for individual blocks, such as a hero section or a pricing card, rather than a whole site at once. This makes it easier to control the details. Next, lock in the parameters that worked and reuse them across all variants: the same format, the same description of lighting and palette. This gives you a better chance of getting a consistent style. One last point. Judge the process by time, not by how good the pi

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering

Build & Ship Production AI Apps with Vibe Coding

Join Zhemal Khamidun's hands-on Vibe Coding Intensive at hamidun.com/vibecoding — master Claude Code, Cursor, and autonomous agentic workflows.

Explore Vibe Coding →Vibe Academy Platform
What do you think?
Loading comments…