Training Taste — Thais Castello Branco, Taste Labs
AI did not start web-design sameness, but it accelerates it. Builders should give design agents explicit brand context, varied references, and checks for repetition, fit, and coherence.
Taste Labs analyzed **more than 2 million websites** spanning roughly **10 years** and found that palettes and layouts were converging before generative AI. Its classifiers use features such as color, typography, layout, and audience to detect repeated design patterns.
Treat visual quality as an inference-time system problem, not just a model problem. Give coding agents a coherent brand system, retrieve relevant references, encourage deliberate variation, and verify that output still fits its audience and context.
Taste Labs analyzed **more than 2 million websites** spanning roughly **10 years** and found that palettes and layouts were converging before generative AI. Its classifiers use features such as color, typography, layout, and audience to detect repeated design patterns. Treat visual quality as an inference-time system problem, not just a model problem. Give coding agents a coherent brand system, retrieve relevant references, encourage deliberate variation, and verify that output still fits its audience and context. The talk reports strong detection and better brand fidelity, but provides no exact scores or public evaluation details. Its proposed creativity API and brand repository were described as work in progress, and none of the methods eliminates the need for human judgment.
This adds an empirical diagnosis to the design-skill candidates: web sameness predates generative AI, so avoiding generic output requires an inference-time system of brand constraints, retrieved references, deliberate variation, audience checks, and human judgment. It reinforces persistent design context and iterative review, while undisclosed evaluation details and unfinished infrastructure prevent broad claims about detection or fidelity gains.