README.md

metadata

title: Scene Machine 9
emoji: 🖥️
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.10.0
python_version: 3.11
app_file: app.py
short_description: Turn text prompts into procedural Three.js scenes.
startup_duration_timeout: 10m
models:
  - deepseek-ai/DeepSeek-V4-Flash-0731
tags:
  - threejs
  - text-to-3d
  - procedural-generation
  - inference-providers
license: mit

Scene Machine 9

Scene Machine 9 turns a text description into a procedural Three.js scene. It uses DeepSeek V4 Flash through Hugging Face Inference Providers, validates the returned JSON against a small scene contract, and renders only whitelisted primitives inside a sandboxed iframe.

Design

The pipeline adapts the central idea from img2threejs: build recognizable 3D objects from code, named parts, procedural geometry, and explicit materials instead of producing an opaque mesh. The original project is an agent skill for image-guided, multi-pass reconstruction. This Space is intentionally text-driven and completes a smaller pipeline in one request:

1. silently inventory identity-defining details;
2. plan blockout, structure, form, materials, and surface details;
3. emit a constrained JSON scene;
4. normalize ranges and reject unsafe fields server-side;
5. render the result with Three.js in the browser.

Generated JavaScript is never evaluated. The renderer accepts seven primitive families, transforms, colors, and bounded PBR material values only.

Configuration

The Space requires an HF_TOKEN secret with Inference Providers access. No token is sent to the browser or included in API responses.

API

The public Gradio endpoint is /generate and accepts:

• prompt: object or small-scene description;
• detail: one of the three geometry budgets shown in the UI;
• finish: one of the art-direction presets shown in the UI.

It returns an interactive preview document, the normalized JSON scene, and a short status.

Model choice

DeepSeek V4 Flash 0731 was selected after checking Hugging Face's live inference catalog. It is available through several routed providers and is particularly strong on current code-agent and full-stack evaluations, which closely match this structured code-generation task. See RESEARCH.md for the implementation notes and tradeoffs.