A developer at Hugging Face has demonstrated an AI coding agent chaining two separate machine learning models to build interactive 3D web galleries directly from text prompts. By querying standardized agents.md documentation endpoints on Gradio Spaces, the agent orchestrated image generation and 3D reconstruction without manual API integration. This modular workflow shows how AI agents can autonomously compose independent open-weights tools into complex multimedia applications.
Standardized Endpoints Power Autonomous Model Chaining
The demonstration relies on agents.md, a plain-text documentation format exposed by Gradio Spaces hosted on Hugging Face. The file provides coding agents with complete instructions to interact with hosted models, including API schema URLs, POST call structures, polling endpoints, and file upload parameters.
By passing agents.md endpoints to a coding assistant alongside an authentication token (HF_TOKEN), agents can discover and execute model capabilities programmatically. Key technical features of this integration protocol include:
- Zero-SDK Execution: Agents execute HTTP requests using basic cURL templates rather than requiring hardcoded Python or JavaScript SDKs.
- Universal Input Handlers: Standardized file upload endpoints streamline transferring generated assets between disparate spaces.
- Direct Agent Orchestration: Documentation enables agents to dynamically parse parameters and construct multi-step execution graphs.
Building 3D Galleries Through Model Chaining
To build the interactive galleries, the agent executed a two-stage generative workflow. It first called the ideogram-ai/ideogram4 Space to produce clean specimen images on dark backgrounds from text descriptions of national monuments. Next, the agent passed those output images directly into VAST-AI/TripoSplat to reconstruct 3D Gaussian splat models in .ply format.
Beyond running model inference, the coding agent handled post-processing and frontend deployment automatically. The agent corrected Y-axis spatial alignment, auto-framed each monument, and compressed .ply files to .ksplat format to reduce file sizes by 3x. It then wrote a custom Three.js web viewer with scroll-to-switch and drag-to-rotate features, deploying complete six-monument galleries for Paris, Japan, and Egypt via single-prompt instructions.
Single-Image Reconstruction Limitations
Despite the efficiency of this modular approach, single-view 3D Gaussian splatting introduces technical constraints. Because models like TripoSplat must infer missing geometry for the back of objects from a single image, certain complex structures fail to reconstruct cleanly.