Show HN: UIPalooza – Do AI models have a recognizable design taste?

UI Palooza, a Show HN project, publishes a gallery that gives different AI models the same versioned design brief and places their resulting interfaces side by side to test whether models show a recognizable design taste. Each run starts in an isolated workspace, with Pi using Lunaroute and other rows using Claude Code or Codex, under a default generation budget of 30 minutes, 120 model responses and high reasoning. The project separates browser interaction and accessibility checks, which measure functionality, from human design judgments, and states that its two baselines are exploratory evidence rather than a statistically reliable ranking.

Rows are models and columns group selected samples by skill variant. Click a preview to open the working artifact. No models match these filters. Reset filters Evidence, and your eye. Latest attempts for the selected brief, samples, and skills. Functional checks and human design judgments remain separate. Different schools of thought. A curated selection based on adoption, relevance, and distinct approaches to interface design. “Top five” is a selection, not a proven quality ranking. A fairer starting point. Hold the task steady. Change the model or the skill. Keep the evidence. 01 Same task, fresh workspace Every sample starts from its versioned brief, local content and assets. The live briefs use browser state; the original backend Forma experiment is preserved in an offline archive. Baselines are independent; Alternative direction receives recorded descriptions of the two baselines. Whacky starts fresh with an experimental-design prompt. 02 Isolated coding agents Each run starts in an isolated workspace. Pi uses Lunaroute; other rows use Claude Code or Codex. The model, coding agent, skill, and sample condition are shown with each result. 03 Versioned ingredients Skill commits, supporting references, model catalog, task and asset hashes, coding-agent versions, prompts, and transcripts are saved. Reruns are new experiments; model generation is not deterministic. 04 Checks, not invented scores Each brief has browser interaction and accessibility checks. These measure functionality, not design taste. Use desktop and mobile previews to make your own comparison. Forma compares each skill independently with vanilla. The frontend suite includes Baseline A and Baseline B as independent attempts, plus Alternative direction as a deliberately different design prompted by descriptions of both baselines. Whacky asks for an experimental interface without showing earlier outputs, initially in Lunaroute vanilla. These conditions test range, not unprompted defaults. Two baselines are exploratory evidence, not a statistically reliable ranking. Provider aliases and reasoning settings may change upstream; manifests record requested IDs and settings. GLM 5.2 maps explicitly to glm-5.2-vision. Native-agent rows compare both the model and its coding agent. System prompts, tools, and reasoning budgets differ. Codex has no equivalent per-response output-token cap; manifests record this difference and each agent's response-counting method. The public gallery contains selected static artifacts and sanitized measurements. Interactive designs run on a separate origin. Original records and private experiments remain in the research workspace. Default generation budget: 30 minutes, 120 model responses, high reasoning; agent-specific output limits are recorded per run. Monetary costs are not compared across providers and subscription plans. A test of taste. Give different AI models the same brief. What do they choose for the type, the layout, the way an interface works? UI Palooza puts those choices side by side. Compare models, add a design skill, or look across attempts to see what changes and what keeps showing up.