SIAgentNews.com
← All news
DevelopingGemini

A developer built five interactive 3D palaces from prompts in Google AI Studio

A dev.to author used Google's reported new Gemini model to make a playable 3D world, and found the tool mattered as much as the model.

· First reported by dev.to

12,000roughly how many lines of TypeScript the finished project has

The short version

  • The project has five palaces, five camera views, a day/night switch, adjustable petals and a PNG snapshot button.
  • AI Studio's app-building mode produced a real multi-file React, TypeScript and Three.js project, not one blob of HTML.
  • The finished code is roughly 12,000 lines of TypeScript, with no image or audio files; textures and sound are generated in code.
  • The Gemini chat app gave flatter scenes, partly because free users get automatic model selection.

Why it matters

You can test what an AI model does with a visual project in a browser tab, with no local setup. The author's main finding is that the same model can look mediocre or impressive depending on the tool wrapped around it. Judge a model by how you use it, not only by the demos you see online.

One prompt grew into a 3D world

As dev.to reported, a developer saw a wave of 3D scenes and generative art on X after Google released a model called Gemini 4 Hydrogen. They decided to build something themselves. The result, called Imperial Palaces, is an interactive 3D world with five palaces in Japanese, Russian, Victorian, Mughal and Persian styles.

The first version was a single Edo-era Japanese castle in a cherry-blossom garden. It had a tiered castle keep, an arched red bridge, stone lanterns, a koi pond and a sunset mountain backdrop. Over a few rounds of follow-up requests, the author added the other four palaces and a set of controls.

The viewer can pick from five camera views, such as the pond, the bridge or the horizon, and the camera glides between them. A toggle switches from golden hour to a moonlit night, with stars, glowing lanterns and a moon streak on the water. Petal density runs from 240 to 640 to 1,200 particles, and a button saves the current view as a PNG image.

Why the chat app disappointed

The author's first try was the Gemini app, where they described the castle scene in plain words. The scenes came out flatter and less detailed than what they had seen online, and refining them felt clumsy.

They give two reasons. Free users in the app get the model through automatic selection, so you can't always tell which model is answering. And a chat window is built for conversation, not for steadily improving a multi-file 3D project.

Moving to Google AI Studio changed the outcome. There the author picked the model by name, used the app-building mode and knew which thinking level was running. The output was a proper project they could preview live and keep editing in place.

What is inside the code

The project uses React and TypeScript, two common web-development tools, plus Three.js, a code library for drawing 3D graphics in a browser. It has no image or audio files to speak of. Everything you see and hear is produced by code at runtime.

The stone, roof tiles, plaster and wood are painted by code onto surfaces in the scene, a technique usually called procedural texturing. Thousands of petals, leaves and grass blades are drawn with a method that reuses one shape many times, which keeps animation smooth.

The sound is built the same way. Using the browser's Web Audio API, the code makes stream and wind noise and plays koto-style plucks, a Japanese string instrument sound, on a five-note scale. The author also notes that camera moves and the day-to-night blend run inside the animation loop, so the page doesn't redraw its interface on every frame.

What the author says about the model

According to the post, Hydrogen is the fast, cheaper tier of the Gemini 4 family, replacing the old Flash branding. The author lists a launch date of November 12, 2026 and a 1M-token context window, meaning how much text the model can consider at once. Google is said to claim it finishes tasks about 2.5× faster than Gemini 3.8 Flash.

The listed API prices are $1.50 per 1M input tokens and $7.50 per 1M output tokens. A token is the small chunk of text a model bills by. The post says the model is available in the Gemini app, Search, AI Studio, the Gemini API and Antigravity.

These are the author's account, and we haven't seen Google's own announcement ourselves. Check which model names your own AI Studio account actually lists before relying on any of this.

How to read a showcase like this

This is one person's best result, not a benchmark. The author also leaned on other people's demos for inspiration, including a scene of a desert ranch that Google reportedly boosted. Social media demos show what is possible, not what is typical.

Still, the practical lessons are useful. Get a decent first version, then iterate. The author says the model was fast enough that a request like adding four more palaces became a quick loop rather than a big commitment.

The author's last point is the one to remember: the tool around a model matters as much as the model. If a result in one app looks weak, it may be worth trying the same model somewhere that lets you choose it and see the project structure.

What to watch

  • Does Google's own announcement confirm the model name, launch date and speed claim that the author reports?
  • Do other builders get similar results in AI Studio's app-building mode, or is this one author's best case?

Try it

Open Google AI Studio, check which Gemini models your account lists, and ask the app-building mode for one small 3D scene. Note where it breaks before you try anything bigger.

Sources and links

First reported by dev.to. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news