Repository

AISmith-3D

AISmith 3D — Local AI 3D Studio

Workflow
Result
Generate
Generate
Refine
Refine
Rig + Animate
Rig + Animate

Specs

  • WorkflowResult
  • GenerateGenerate
  • RefineRefine
  • Rig + AnimateRig + Animate
  • Texture PaintTexture Paint
  • VariablePurpose
  • BLENDER_EXEExplicit Blender executable path
  • TRELLIS_CPP_DIRLocal trellis.cpp runtime directory
  • TRELLIS_CPP_MODEL_DIRGeometry-model storage location
  • AUTOREMESHER_EXEOptional AutoRemesher executable path
  • PYTORCH_INDEX_URLOverride the CUDA PyTorch package index

README

AISmith 3D — Local AI 3D Studio

AISmith 3D is a Windows-local studio for taking a reference image through 3D generation, AI-assisted refinement, texture painting, and browser-based rigging/animation. It keeps assets and inference on the local machine; large model files are deliberately downloaded from the app only when a user selects a workflow.

Hackathon note: this repository contains the application and its source dependencies, but does not include any model weights. Use the in-app download controls after startup.

What it does

Workflow Result Prerequisites
Generate Image → geometry GLB through the native trellis.cpp CUDA worker NVIDIA CUDA-capable GPU
Refine Image + mesh → reconstructed, simplified, PBR-textured GLB with TRELLIS.2 FP8 NVIDIA CUDA-capable GPU
Rig + Animate Fit a skeleton, bind weights, preview/select animation, export GLB Modern browser; no GPU or Blender required
Texture Paint Paint a mesh in the browser, bake a base-color texture, export GLB Blender 4.2+

The app runs its API locally at 127.0.0.1:8000 and its Vite UI locally at 127.0.0.1:5173. It does not expose a ComfyUI server: the TRELLIS.2 refiner is invoked in an isolated local worker.

Requirements

  • Windows 10 or 11
  • Python 3.11 or 3.12
  • Node.js 20+ with npm
  • For Generate and Refine: NVIDIA GPU, current NVIDIA driver, and CUDA-compatible PyTorch. CPU-only machines can still use the UI and Rig + Animate workflow, but inference buttons return a clear CUDA requirement error.
  • For Texture Paint and Blender-assisted conversion: Blender 4.2 or later.

Disk space is required only when a model workflow is chosen in the UI: Generate downloads approximately 9.6 GB of geometry GGUFs; Refine downloads approximately 8.8 GB for its FP8 runtime and models. Keep additional working space for generated assets.

Install and run

Clone the repository, then run the one-command launcher from PowerShell:

git clone https://github.com/intisarGIT/AISmith-3D.git
cd AISmith-3D
.\start.ps1 -Install

This creates an isolated .venv, installs the Python and frontend dependencies, detects NVIDIA hardware, and starts the API and UI. Open http://127.0.0.1:5173.

To let the setup script install Blender through winget when it is missing:

.\start.ps1 -Install -InstallBlender

The initial setup must be run in a PowerShell session permitted to create the virtual environment and install packages. Later launches need only:

.\start.ps1

If a previous AISmith API is still listening on port 8000, replace it with:

.\start.ps1 -RestartApi

Model downloads

After the UI opens, use its Download models action in the Generate or Refine workflow. The project intentionally does not download model weights during installation. Downloads are stored below studio-data/, which is ignored by Git.

Optional configuration

Copy .env.example to .env only when you need non-default paths:

Copy-Item .env.example .env
Variable Purpose
BLENDER_EXE Explicit Blender executable path
TRELLIS_CPP_DIR Local trellis.cpp runtime directory
TRELLIS_CPP_MODEL_DIR Geometry-model storage location
AUTOREMESHER_EXE Optional AutoRemesher executable path
PYTORCH_INDEX_URL Override the CUDA PyTorch package index

Verification

From the repository root:

.\.venv\Scripts\python.exe -m compileall -q backend
.\.venv\Scripts\python.exe -m unittest discover -s backend/tests -v
npm.cmd --prefix frontend run build

Architecture

reference image → native trellis.cpp geometry → TRELLIS.2 FP8 refine / reconstruct / PBR → Mesh2Motion rig + animation

Heavyweight inference is serialized through a coordinator lease. TRELLIS.2 runs in a short-lived process, Blender stages are headless and short-lived, and tabs pass artifact IDs/files rather than large meshes through browser memory.

License and acknowledgements

AISmith 3D’s original application code is available under the MIT License. It incorporates or integrates third-party projects under their own licenses; their license texts remain with the relevant vendored source. See THIRD_PARTY_NOTICES.md for attribution, source links, and licensing details.

In particular, thank you to the maintainers and contributors of TRELLIS.2, trellis.cpp, trellis2cpp, ComfyUI, ComfyUI-Trellis2, AutoRemesher, and Mesh2Motion.

Model weights are downloaded separately and may have terms different from this repository. Review the applicable Hugging Face model cards and upstream project licenses before redistributing models or outputs.

Built with Codex

Codex/GPT-5.6 assisted with application architecture, native worker boundaries, GPU/VRAM policy, artifact API, UI, texture painter, retargeting layer, tests, and documentation. Retain the primary Codex task/session evidence with the hackathon submission materials if the event requires it.