3D-Agent

A Django chat application that wraps an LLM agent capable of turning text prompts and reference images into textured 3D models. The agent reasons over a set of tools — image generation, web/image search, GPU-hosted 3D pipelines, and two quality "judges" — and streams its progress back to a chat UI.

How it works

Browser (chat UI, SSE stream)
        │
        ▼
Django (core/ + chat_interface/)  ──  SQLite (sessions, messages, gallery)
        │
        ▼
LangChain agent (agent.py)  ──  Gemini (gemini-flash-latest)
        │
        ├─ Gemini image model      → concept images, Objaverse restyle
        ├─ Tavily                  → web search, image search
        ├─ Modal GPU endpoints     → TRELLIS.2, PartCrafter, Hunyuan3D-2
        ├─ pyrender / trimesh      → render mesh views
        └─ CLIP                    → alignment scoring (judge)

The agent (agent.py) is built with langchain.agents.create_agent over the tools in tools.py. It can:

• Generate a concept image from text (Gemini) or download one from image search.
• Restyle an image into the clean "Objaverse" single-object look.
• Submit an image to a 3D pipeline (TRELLIS.2, PartCrafter, or Hunyuan3D-2) hosted on Modal, then poll and download the resulting .glb.
• Render multi-view PNGs of a generated mesh.
• Judge quality two ways: a VLM judge (tool_inspect_image, Gemini vision) and a CLIP alignment score (tool_score_alignment) that flags the worst- matching camera view.

Project layout

Prerequisites

• Python (see .python-version)
• uv for dependency management
• A Modal account with the CLI installed and authenticated (modal token new) — the 3D pipelines run there, and result download falls back to modal volume get.
• API keys for Google Gemini and Tavily.

Setup

1. Install dependencies (uses pyproject.toml / uv.lock):

   uv sync
   

   (Or pip install -r requirements.txt into a virtualenv.)

2. Create a .env in the project root:

   GEMINI_API_KEY=your-gemini-key
   TAVILY_API_KEY=your-tavily-key
   TRELLIS_WORKSPACE=your-modal-workspace
   

   TRELLIS_WORKSPACE is your Modal workspace slug. tools.py builds the pipeline URLs from it, e.g. https://<workspace>--trellis2-generator-web.modal.run. .env is gitignored — keep your real keys out of version control.

3. Deploy the Modal apps (one per pipeline). See each folder's MODAL.md for details, then:

   modal deploy TRELLIS.2/modal_app.py
   modal deploy PartCrafter/modal_app.py
   modal deploy hunyuan3d-2/modal_app.py
   

   These expose the *-web.modal.run endpoints the agent calls. The expected deployed function names are:

   • trellis2-generator-web, trellis2-texturer-web
   • partcrafter-objectgenerator-web, partcrafter-scenegenerator-web
   • hunyuan3d-2-generator-web

4. Run Django migrations:

   python manage.py migrate
   

Running

python manage.py runserver

Then open http://127.0.0.1:8000/:

• / — landing page with the central prompt box
• /chat/<id>/ — a chat session
• /gallery/ — all generated assets

Type a request (e.g. "make a 3D model of a cyberpunk drone") or upload a reference image. The agent streams friendly status updates over Server-Sent Events while it runs the pipeline, and the resulting .glb shows up in the chat and the gallery.

Notes

• First CLIP call is slow. tool_score_alignment loads openai/clip-vit-large-patch14 into the Django process on first use; on a machine without CUDA it runs on CPU.
• Modal cold starts. The job runner in tools.py has generous retry, backoff, and polling budgets to tolerate GPU container cold starts; a first generation can take a while.
• Stopping a job. The UI's stop button cancels the in-flight Modal call via its fc-... call ID.

License

See LICENSE.