How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces

A coding agent built a 3D Paris gallery by chaining two Hugging Face Spaces, Ideogram 4 and TripoSplat, with zero manual image or 3D tooling.

Tuesday June 9, 2026 Source: huggingface.co
TL;DR — Quick Answer

Hugging Face engineer Mishig asked a coding agent to build a 3D gallery of Paris monuments as Gaussian splats. The agent chained two Spaces — Ideogram 4 for image generation and VAST-AI/TripoSplat for single-image 3D reconstruction — using each Space's agents.md file for call instructions. The result shipped as mishig/monuments-de-paris, and the same two-Space pipeline spun up analogous galleries for Egypt and Japan.

Key Takeaways

How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces — AI news article illustration

Hugging Face engineer Mishig asked a coding agent to build a website showcasing the monuments of Paris as 3D Gaussian splats — and never opened an image generator or a 3D tool. The agent produced every asset by chaining two Hugging Face Spaces, then deployed the whole gallery as a static Space at mishig/monuments-de-paris.

The agent called two Spaces in sequence: Ideogram 4 generated clean, dark-background specimen shots of each monument, and VAST-AI/TripoSplat reconstructed a 3D Gaussian splat from each single image. Prompt to image to 3D — a full multimedia pipeline with zero hand-written integration code.

Every Space Is a Building Block

The unlock is agents.md. Every Gradio Space now exposes a plain-text file that tells an agent exactly how to call it — the API schema, call and poll endpoints, file upload format, and auth hint. No client library, no hardcoded SDK: the agent reads the file and drives the Space end to end. Set an HF_TOKEN and the pipeline is running.

The Glue Work Was the Story

The agent handled the hard parts of production entirely on its own:

Two Prompts, Many Galleries

The real test of a building block is cheap reuse, and it passed: the same two Spaces, with only the prompts changed, produced the Monuments of Egypt and Monuments of Japan. The marginal cost of a new multimedia app falls toward the cost of describing it.

Why This Matters

What This Means

The building-block economy has reached multimedia. Hugging Face’s open-weight catalog is becoming a library of callable primitives — and the agents already know how to glue.

Frequently Asked Questions

What is agents.md on Hugging Face?

A plain-text file every Gradio Space exposes that tells an agent exactly how to call it — the API schema, call and poll endpoints, file upload format, and auth hint.

How was the 3D Paris gallery built?

A coding agent chained Ideogram 4 to generate images of each monument, then fed each image to VAST-AI/TripoSplat to reconstruct a 3D Gaussian splat, and deployed the result as a static Space.

What models did the agent use?

Ideogram 4 for image generation and TripoSplat for single-image to 3D Gaussian splat reconstruction, both hosted as Hugging Face Spaces.

Can I try this myself?

Yes. Fetch any Space's agents.md with curl, set an HF_TOKEN, and ask a coding agent to build something — the full pipeline lives in the monuments-de-paris Space repo.

This article is based on the official announcement from huggingface.co . Read the original for full technical details.

Related Articles

Back to all news