PPF Contact Solver

Why it's ranked here

This is a serious simulation stack with an unusually complete path from research to production tooling. The contact method has a peer-reviewed foundation, successful runs receive an intersection check, and automated workflows repeatedly exercise examples. Blender, Python, JupyterLab, Docker, Windows, remote hosts, and MCP offer several entry points. The verdict remains qualified because it needs specific NVIDIA hardware, targets offline work, and documents meaningful SSH and MCP security weaknesses.

What's good

The workflow supports much more than a bare solver. Artists can group objects, tune measured fabric presets, add moving pins and invisible colliders, resume from checkpoints, fetch animation, and bake results that play without the add-on. Strain limits place strict bounds on triangle extension. Both contact and elasticity calculations run on the GPU in single precision. Documentation includes galleries, scripts, tutorials, troubleshooting guidance, parameter references, and explicit disclosure of AI-assisted code and documentation.

Tradeoffs

The hardware envelope is narrow: x86 only, no arm64, and an NVIDIA GPU supporting CUDA 12.8 or newer. This is an offline simulator, not a real-time engine. Docker deployment uses an image around 1 GB, while the native Windows bundle is roughly 320 MB. First-time SSH connections silently trust unknown host keys. The local MCP service has no authentication or sandbox and exposes arbitrary Python and shell execution. The authors also warn that add-on algorithms received less scrutiny than the underlying solver research and may retain rough edges.

How to use it well

It best suits simulation engineers, technical artists, and researchers producing cloth, soft-body, rod, rigid-body, or granular animation on NVIDIA-equipped workstations or remote GPU hosts. Build scenes and constraints in Blender, iterate on materials and global parameters, checkpoint long runs, then bake approved motion for downstream rendering. Use JupyterLab for headless sweeps and batch variants. Keep MCP local and review agent actions. Choose another tool when the requirement is real-time interaction, arm64 deployment, non-NVIDIA compute, or a general renderer and content-authoring suite.

Observed

License

Apache License 2.0, permitting commercial and proprietary use.

Primary implementation

Rust 2021 workspace for formats, core logic, Python bindings, server, and solver, with a Python Blender front end.

Packaging

Docker image and self-contained Windows executable are provided; the PyO3 extension is loaded directly rather than installed as a wheel.

Interfaces

Blender add-on, documented Python API, JupyterLab workflow, TCP solver protocol, and local MCP server.

Platform support

Linux through Docker or local and remote hosts, plus Windows 10/11 native or Docker; x86 and NVIDIA CUDA 12.8 or newer are required, while arm64 is unsupported.

Testing structure

Pytest targets cover the frontend, PyO3 bindings, and add-on host tests; separate Blender and example workflows are described.