Hugging Face Spaces agents.md enables automatic 3D multimedia pipelines

Hugging Face Spaces agents.md enables automatic 3D multimedia pipelines

TL;DR

A coding agent automatically built a static Hugging Face Space that showcases Paris monuments as 3‑D Gaussian splats by chaining two existing Spaces—one for image generation and one for single‑image 3‑D reconstruction—demonstrating how agents.md turns any Space into a composable building block for multimedia pipelines.

The building‑block economy reaches multimedia

The "building block economy" described by Mitchell Hashimoto argues that modern software development favors small, well‑documented components that can be assembled by agents rather than monolithic codebases. While this idea has been applied to code libraries, the same principle now applies to multimedia AI. The most difficult part of using state‑of‑the‑art image, video, TTS, or 3‑D models has traditionally been integration: SDKs, weight files, GPU requirements, and input‑output formats. By exposing each model as a callable component, agents can glue them together just as they do with npm packages.

agents.md makes every Hugging Face Space a callable block

Every Gradio Space on the Hugging Face Hub provides a plain‑text agents.md file that describes how to invoke the Space programmatically. A typical agents.md contains:

API schema:   GET  .../gradio_api/info
Call endpoint: POST .../gradio_api/call/v2/{endpoint} {"param_name": value, ...}
Poll result:  GET  .../gradio_api/call/{endpoint}/{event_id}
File inputs:  POST .../gradio_api/upload -F "files=@file.ext"
Auth:         Bearer $HF_TOKEN

No client library or hard‑coded SDK is required—an agent can read these instructions, set an HF_TOKEN, and drive the Space end‑to‑end. The real power emerges when the output of one Space becomes the input of another, enabling pipelines such as prompt → image → 3‑D splat.

Worked example: Paris monuments → Gaussian splats

The author instructed a coding agent to create a gallery of Paris monuments using two Spaces:

  1. Image generation – The ideogram-ai/ideogram4 Space produced clean, dark‑background images of each monument (e.g., the Pantheon, Opéra, Arc de Triomphe, Sacré‑Cœur, and a diorama of the Eiffel Tower).
  2. 3‑D reconstruction – The VAST-AI/TripoSplat Space converted each single image into a 3‑D Gaussian splat (.ply).

The agent performed additional glue work:

  • Detected that TripoSplat outputs are Y‑down and automatically flipped them upright.
  • Auto‑framed each monument for optimal viewing.
  • Compressed the .ply files to .ksplat (≈3× smaller) for fast loading.
  • Generated a Three.js viewer with scroll‑to‑switch and drag‑to‑rotate interactions.
  • Deployed the entire viewer as a static Space (mishig/monuments-de-paris).

Human input was limited to high‑level preferences such as "zoom out more" or "replace the obelisk with a better shape for splatting." The agent also reacted to visual feedback (e.g., a wide glass pyramid splatted poorly) and iterated accordingly, illustrating an "outsourced R&D" loop.

Reusing the pipeline with just a prompt

Because the pipeline is fully described by two agents.md files, creating a new gallery requires only a new textual prompt. The author asked the same agent to generate galleries for Egypt and Japan, resulting in:

  • Monuments of Egypt – Great Pyramid, Sphinx, Abu Simbel, Tutankhamun’s mask, Karnak, Colossi of Memnon.
  • Monuments of Japan – Tokyo Tower, Himeji Castle, Kinkaku‑ji, Osaka Castle, Great Buddha of Kamakura, Itsukushima torii.

Each gallery was built with the same two Spaces, the same compression and viewer logic, and only the textual description changed. This demonstrates that the marginal cost of a new multimedia app approaches the cost of describing it.

Why this matters

  • Composable models – State‑of‑the‑art image and 3‑D reconstruction models from different organizations can be chained without writing integration code, turning the Hub’s open‑weights catalog into a library of callable multimedia primitives.
  • Agent‑friendly documentation – agents.md provides a machine‑readable contract, making a Space instantly reachable for agents and encouraging its reuse over manually set‑up models.
  • Integration barrier removed – Tasks that previously required a dedicated engineering effort (e.g., "turn a prompt into a rotating 3‑D monument") are now a single step in an agent‑driven pipeline.

Try it yourself

To replicate the workflow, point any coding agent (e.g., Claude Code) at the agents.md files of the two Spaces and provide an HF_TOKEN:

# Image generation Space
curl https://huggingface.co/spaces/ideogram-ai/ideogram4/agents.md

# Single‑image to 3‑D Gaussian splat Space
curl https://huggingface.co/spaces/VAST-AI/TripoSplat/agents.md

Insert the URLs into the agent, describe the desired gallery, and let the agent orchestrate image creation, 3‑D reconstruction, compression, viewer generation, and deployment. The full reproducible pipeline and supporting scripts are available in the monuments‑de‑paris Space repository.

Conclusion

agents.md turns every Hugging Face Space into a documented, callable component, enabling agents to compose complex multimedia applications with minimal human direction. The Paris 3‑D gallery showcases how a simple prompt can drive a full pipeline—from image synthesis to 3‑D rendering and web deployment—signaling a shift toward a building‑block economy for AI‑powered media creation.

Sources