Presidium Shape Studio

Website: https://presidiumdev.github.io/presidium-shapestudio/ · PyPI: https://pypi.org/project/presidium-shapestudio/ · Releases: https://github.com/presidiumdev/presidium-shapestudio/releases

A free, local image-to-3D studio and MCP server. One picture in, a textured mesh out (GLB and OBJ with vertex colours, plus a turntable PNG), with permissively licensed models on your own machine. A desktop app for Windows, macOS and Linux, a web UI, a command line, and an MCP server for Presidium, Claude, Cursor or any other agent. It runs on the CPU if that is all there is, and uses an NVIDIA GPU when you have one.

Part of the Presidium family, sibling of Presidium Image Studio, built on the same architecture and the same rules. Design: Hivemind ADR-051; brand: ADR-065 and 066. Licence: Apache-2.0.

The policy

Quality is a setting, never a consequence. The default preset is best: the finest surface grid on the strongest permissively licensed model that fits your machine. The machine decides how long that takes (cores, GPU, memory), never how good the result is. Nothing downgrades silently to be faster: the studio estimates the time and tells you.

Faster only with consent. Presets balanced and fast exist and must be chosen explicitly. Each states its trade-off in words, in the tool description and in capabilities():

The network pass is identical at every preset; only the surface extraction changes. Nothing else is lowered on its own.

Two tiers, chosen by hardware, never silently. The standard tier is TripoSR (MIT): it runs on a CPU. The best tier is TRELLIS image-large (MIT): clearly higher fidelity, and it needs a discrete NVIDIA GPU with 16 GB of memory or more. On a machine without that GPU the plan says so in fit_note ("Running TripoSR because TRELLIS needs an NVIDIA GPU and this machine has none. This is a hardware fit, not a speed choice."), and asking for best by name is refused with the reason instead of being swapped for something else.

Permissive licences only. TripoSR weights and code: MIT. TRELLIS weights: MIT. PyTorch: BSD. IS-Net and U²-Net background models: Apache-2.0. Excluded by policy, whatever their quality: Hunyuan3D 2.x (Tencent Hunyuan Non-Commercial License, with territorial restrictions), Stable Fast 3D, Stable Zero123 and Stable Point Aware 3D (Stability AI Community License, free only below a revenue threshold).

Offline after setup. The models and, if you want it, the CUDA build of PyTorch download once, each after explicit consent (--yes or consent=true), into the studio's own data folder, pinned by SHA-256. After that nothing touches the network; there are no keys or accounts.

Measured, not promised. Every time estimate comes from runs on your machine (per model, per device) and is labelled not measured yet until the first one. This README contains no timing claim this project has not measured: as of 0.1.1 the real model has not been run by the project yet (see Status).

Status, stated plainly

Sizes, stated plainly

Memory: TripoSR's weights are 1.68 GB in fp32; the planner budgets about 4.2 GiB of RAM for a 256³ grid on the CPU (weights plus working memory) and refuses over-budget runs unless you unlock them. TRELLIS is budgeted at 16 GiB of GPU memory, the authors' figure.

Install

Windows app: download ShapeStudio-<version>-Setup.exe from the latest release and run it: Start menu entry, uninstall from Settings → Apps, no administrator rights. Prefer portable? …-windows-x64.zip is the same app: unzip, run ShapeStudio.exe. The build is not code-signed yet, so SmartScreen asks once per download ("More info" → "Run anyway"); the .sha256 file next to each download lets you verify it. ShapeStudio.exe cli doctor reaches the command line. Logs: %LOCALAPPDATA%\shapestudio\logs\app.log.

macOS (Apple silicon): ShapeStudio-<version>-macos-arm64.zip; unzip, right-click "Shape Studio.app" → Open the first time (not notarized yet). Linux (x86_64): ShapeStudio-<version>-linux-x86_64.tar.gz; unpack and run ShapeStudio, which opens the studio in your browser. Details in packaging/README-LINUX-MACOS.txt.

Everywhere else (and for the MCP server and CLI), Python 3.11 or newer:

pipx install presidium-shapestudio

or the same wheel with pip install / uv tool install. The PyPI package is not published yet: it will be registered once as presidium-shapestudio when its publisher exists; until then the wheel on the GitHub release is the Python distribution. PyTorch is a dependency and comes along (see Sizes). Then:

shapestudio doctor

shapestudio download-model triposr --yes

shapestudio download-model isnet-general-use --yes

shapestudio reconstruct chair.png --out meshes/

doctor shows the hardware found, the PyTorch runtime (CPU or CUDA, and whether the CUDA build can be installed here), both tiers with whether they are downloaded and whether they fit, and the expected time per preset, each labelled measured or not. reconstruct queues one job, follows it, and copies the GLB, OBJ and turntable into --out. shapestudio ui opens the studio in your browser; shapestudio benchmark runs one real reconstruction of a built-in test image and writes profile.json so every later estimate is measured.

On a machine with an NVIDIA GPU:

shapestudio install-runtime --cuda --yes

installs the pinned CUDA build of PyTorch (pip into the same interpreter; in the packaged Windows app a hash-verified wheel unpacked under the data folder). It is offered, never automatic, and doctor says which build actually loaded.

The GitHub Releases of this repository and the presidium-shapestudio package on PyPI are the official builds (see NOTICE and TRADEMARK.md).

The app

• Home: a getting-started checklist (runtime, model, background model, first mesh, benchmark), the two tiers with their state on this machine, and the policy.
• Create: drop a picture, pick a tier and preset, see the expected time and a load meter before you commit, reconstruct. The queue shows every job with its phase and an honest ETA (or "measuring" once the estimate is spent).
• Gallery: every mesh with a 3D viewer (model-viewer, Apache-2.0, loaded from jsdelivr pinned to 4.1.0; offline the PNG turntable is shown instead), downloads (GLB, OBJ, turntable), favourites, notes, projects (named series with one-zip export), select several for a zip or the trash, trash with restore.
• Settings: hardware and runtime, benchmark and the measured profile, every model and tool with licence, size, status and Install/Show/Remove (install dialogs show the files, sizes, sources and SHA-256), preferences (device, memory limit, threads, over-budget unlock, pause), projects, trash, about (with the MCP command).

Dark and light themes in the Presidium palette (Shape Studio coral #D5582F), phone-friendly layout, the Presidium spinner while a job runs.

MCP tools

shapestudio serve speaks MCP over stdio. Every tool declares annotations (read-only, destructive, idempotent, open-world) and its description is written for the agent.

Client configs for Claude Desktop, Claude Code, Cursor and VS Code are in examples/mcp.

Licences

The code is free (Apache-2.0, LICENSE, NOTICE, third_party/). The names "Hivemind", "Presidium Shape Studio" and "Presidium Shape Studio" and the logo identify the official project and are not covered by the licence; the policy is in TRADEMARK.md. Official builds carry the line "Presidium Shape Studio · made with ♥ by Ratul Roy (royal-programmer)".

Development

uv venv && uv pip install -e ".[dev,verify]"

uv run pytest -q

The suite (about 50 tests) runs with a fake engine and needs no weights; the vendored TripoSR is exercised on random weights, and the ViT cross-check runs when transformers is installed. ruff check src tests and ruff format --check src tests are the lint. The data folder is %LOCALAPPDATA%\shapestudio (Windows), ~/Library/Application Support/shapestudio (macOS) or ~/.local/share/shapestudio (Linux); override with SHAPESTUDIO_DATA_DIR.

Roadmap

1. First real TripoSR run on a GPU machine; measured numbers into this README and the site.
2. TRELLIS engine integration behind the existing tier (CUDA stack packaging).
3. Texture baking (UV atlas) in addition to vertex colours.
4. Multi-view input and mesh clean-up (hole filling, decimation presets).