Four Muse Spark demos show an AI building, browsing and inspecting images

From prompt to a changing 3D scene

In one published example, a coding agent iterates on a 3D scene using browser screenshots as feedback. The sequence makes the work visible: generate a change, inspect the rendered result, then adjust it. It is a practical glimpse of a model contributing to a build loop rather than returning a single answer.

A browser agent works from screenshots

A separate browser example uses screenshots as the agent’s view of the page. The authors say it acts on pixels rather than reading the page’s DOM or HTML, and show an eight-action run. That is a useful distinction: the demo tests visual interaction, while a DOM-enabled agent has access to a different and often richer interface.

Voice shopping and visual inspection

Another clip shows a spoken request guiding a browser toward products that fit stated criteria. Liquid AI’s d1-ultra example takes a different route: given a circuit-board image, it returns probability scores for missing, extra or scratched components. The published example assigns 0.80 probability to a missing component.

These are demonstrations from their publishers, not independent benchmark results. Together they show four concrete tasks—iterative scene building, screenshot-based browser interaction, voice-guided shopping and visual defect classification—and the kinds of checks needed before relying on each one.