How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces
A coding agent built a Paris monument gallery by chaining an image Space and a 3D splat Space, then assembled the viewer itself.
Intelligence analysis by GPT-5.4 Mini
The post argues that Hugging Face Spaces have become callable building blocks for agents: one Space can generate images, another can turn them into 3D Gaussian splats, and an agent can stitch the pieces together. The author uses a Paris monuments gallery to show how much of multimedia software can now be assembled through documented interfaces instead of custom integration work.
An agent made a Paris gallery by using one tool to draw pictures and another tool to turn those pictures into 3D objects, like stacking LEGO bricks instead of carving a statue from scratch.
Analysis
What happened
The author asked a coding agent to build a website showcasing Paris monuments as 3D Gaussian splats. The agent did not rely on a human-operated image tool or a manual reconstruction workflow. Instead, it called two Hugging Face Spaces directly: one to generate clean source images and another, VAST-AI/TripoSplat, to reconstruct each image into a 3D splat.
Why that worked
The post's main claim is that Hugging Face Spaces have become building blocks for agents because every Gradio Space now exposes an agents.md file. That file tells an agent how to fetch the schema, call the endpoint, poll for results, upload files, and authenticate with HF_TOKEN. In other words, the integration instructions are machine-readable, so the agent can operate the Space end to end without a bespoke client library.
What the agent had to do
The agent chained prompt to image to 3D. It also handled the glue work: it flipped TripoSplat's Y-down output upright, auto-framed the monuments, compressed .ply files into smaller .ksplat files for faster loading, built a Three.js viewer with scroll-to-switch and drag-to-rotate behavior, and deployed the whole thing as a static Space. The only human input described in the post was taste-level feedback such as making the view zoomed out or choosing a different object for splatting.
The broader point
The author frames this as a preview of multimedia software development. Models become composable primitives, and agents tend to prefer tools that are documented and reachable. The article's core message is that the hard part was not the models themselves, but the integration, and that barrier is now much lower when tools publish clear agent-facing interfaces.
Key points
- A coding agent built a Paris monument gallery by chaining an image-generation Space and a 3D reconstruction Space.
- The post highlights `agents.md` as the mechanism that lets agents call Hugging Face Spaces directly.
- The agent handled integration tasks such as flipping output orientation, auto-framing, compressing files, and building the viewer.
- The author argues that multimedia AI is moving toward a building-block model where agents glue together documented tools.
- The example is presented as evidence that the main bottleneck is integration, not just model quality.
If this pattern spreads, more AI tools could be used together like plug-and-play parts instead of separate, hard-to-wire systems. That would make it easier to build rich multimedia apps from open models without a lot of custom glue code.
The approach still depends on tools exposing reliable, well-documented interfaces and on agents handling format quirks correctly. If a Space changes its API or produces weak outputs for a given input, the whole chain can break or require manual correction.
