Three Neural Networks Drew Web References in Five Styles: From Fintech to Web 1.0

Neural networks are becoming a useful tool for web designers. The author of the StudyAI blog decided to test this in practice: they created web assets for five completely different industries using three modern image generation models. The results turned out to be mixed.

Three models in the test:

Grok Image 2, Flux 2 Pro, Nano Banana Pro

The author chose three popular models: Grok Image 2, Flux 2 Pro, and Nano Banana Pro. Each was asked to perform identical tasks: create web assets with precise requirements for style, color palette, typography, and functionality. This isn't just image generation — it's a test of how well neural networks understand the specific requirements of web designers.

14 examples across five industries

Esports and gaming (4 examples): • HUD panel for an esports shooter (futuristic minimalism, neon cyan accents, transparent PNG) • RPG inventory in dark fantasy style (gold frames, parchment texture, Diablo-inspired) • Mobile buttons for a casual puzzle game (candy-like shapes, glossy gradients) • Web 1.0 animated buttons (beveled 3D borders, GeoCities aesthetic, sprite sheet)

Fintech and banking (3 examples): • Crypto exchange dashboard (dark mode, candlestick charts, glassmorphism, monochrome palette) • Bank transfer card (neumorphic soft UI, 24px rounded, iOS guidelines) • Investment app onboarding (3 screens, isometric style, Stripe-inspired)

Web 1.0 nostalgia (1 example): • GIF banners 468x60 as in the 2000s (Comic Sans, Impact, flashing "Under Construction," Y2K)

Ecology (3 examples): • Carbon footprint dashboard (organic soft UI, sage green, Patagonia aesthetic) • Eco-product cards for a marketplace (recycled kraft paper, botanical illustrations) • Mobile recycling tracker (playful flat design, Duolingo-style gamification)

AI services (3 examples): • AI assistant chat interface (streaming animation, typing indicator, gradient edge glow) • ML model control panel (loss/accuracy charts, confusion matrix, neural network visualization) • Image generator onboarding (prompt input, style presets, progress ring)

Results: two models excelled, one fell short

Grok 2.0 and Nano Banana Pro demonstrated a good sense of visual style and delivered acceptable results almost immediately. Both models understood stylistic requirements well — from minimalism to nostalgia, from glassmorphism to Y2K aesthetics. The results looked professional and ready to be included in a brief or even serve as a basis for final design.

Flux 2 Pro, despite solid training and good generation logic, made frequent errors. Working with text proved especially problematic: the model repeatedly wrote in an unidentified language. This is odd, since Flux uses a flow-matching approach that should handle sequential logic and text labels well. Nevertheless, the errors it produced were systematic.

Neural networks showed a good ability to generate web interface elements in the required style. With their help, a designer can sketch out a full set of references for any project in an hour or two — from a fintech startup to a nostalgic Web 1.0 website. If you invest time in crafting more detailed prompts and allocate a budget for a series of generations, you can get ready-made components for a production site.