Examples Monorepo

A pixi workspace containing multiple computer vision example projects, each with its own isolated environment and dependencies.

Packages

Package	Prod env	Dev env	Python	GPU	Description
monoprior	monoprior	monoprior-dev	3.12	CUDA 12.9	Monocular geometric priors (depth, normals)
prompt-da	prompt-da	prompt-da-dev	3.12	CUDA 12.9	Prompt Depth Anything — depth completion on Polycam data
wilor-nano	wilor	wilor-dev	3.12	CUDA 12.9	Hand pose estimation
sam3d-body-rerun	sam3d	sam3d-dev	3.12	CUDA 12.9	3D body segmentation with SAM + Rerun
sam3-rerun	sam3-rerun	sam3-rerun-dev	3.12	CUDA 12.9	SAM3 video segmentation with Rerun
robocap-slam	robocap	robocap-dev	3.10	None	Multi-camera visual odometry & SLAM
pysfm	pysfm	pysfm-dev	3.12	CUDA 12.9	COLMAP SfM reconstruction with Rerun + Gradio
vistadream	vistadream	vistadream-dev	3.12	CUDA 12.9	Single-image 3D reconstruction via 3D Gaussians
gsplat-rust-renderer	gsplat-rust-renderer	gsplat-rust-renderer-dev	3.12	CUDA 12.9	Rust-based Gaussian splatting renderer
pyvrs-viewer	pyvrs-viewer	pyvrs-viewer-dev	3.12	None	VRS file viewer with Rerun

Quick start

Direnv setup (recommended, one-time)

direnv auto-activates the correct pixi environment when you cd into a package directory. No more -e and --frozen flags.

pixi global install direnv                      # install direnv
echo 'eval "$(direnv hook bash)"' >> ~/.bashrc  # or: echo 'eval "$(direnv hook zsh)"' >> ~/.zshrc
source ~/.bashrc                                # restart shell or source your rc file
direnv allow                                    # approve root .envrc
for d in packages/*/; do (cd "$d" && direnv allow); done  # approve all packages

To reduce output noise, create ~/.config/direnv/direnv.toml:

[global]
hide_env_diff = true

Development workflow

With direnv, just cd into a package and use tools directly:

cd packages/robocap-slam/   # direnv activates robocap-dev
python --version             # → 3.10.x
ruff check .                 # just works
pytest -q                    # just works

cd ../../packages/pysfm/     # direnv switches to pysfm-dev
python --version             # → 3.12.x

For tasks with depends-on chains (like auto-downloading data), use pixi run from the repo root:

pixi run -e monoprior --frozen relative-depth   # downloads data, then runs demo
pixi run -e robocap --frozen robocap-track-slam
pixi run -e pysfm --frozen sfm-reconstruction-demo

Override environment per directory

Each .envrc defaults to the *-dev environment. To use a different env (e.g., prod), create a .envrc.local (gitignored):

# packages/robocap-slam/.envrc.local
PIXI_ENV=robocap

CUDA driver troubleshooting

GPU environments require CUDA 12.9. If pixi install fails with Virtual package '__cuda >=12.9' does not match, your NVIDIA driver is too old.

Workaround (no reboot):

export CONDA_OVERRIDE_CUDA=12.9
pixi install -a

Proper fix — upgrade your NVIDIA driver to one that reports CUDA 12.9+ (check with nvidia-smi). On Ubuntu:

sudo apt install nvidia-driver-590-open && sudo reboot

Listing tasks

pixi task list -e robocap       # Demo/app tasks only
pixi task list -e robocap-dev   # Demo/app tasks + lint, typecheck, tests

Architecture

How pixi.toml is organized

All pixi configuration lives in the root pixi.toml. Per-package pyproject.toml files only contain standard Python packaging metadata ([project], [build-system], [tool.ruff], etc.) — no [tool.pixi.*] sections.

The root pixi.toml is structured around features that compose into environments:

[workspace]                     # Channels, platforms, pixi version
[pypi-options]                  # Workspace-level: no-build-isolation, dependency-overrides

[feature.common]                # Shared base deps (all envs get these)
[feature.cuda]                  # CUDA 12.9 toolkit + PyTorch GPU
[feature.dev]                   # ruff, pytest, beartype, pyrefly, hypothesis + PIXI_DEV_MODE=1
[feature.ide]                   # Standalone IDE/editor tooling (Python 3.12)

[feature.monoprior]             # Package-specific: conda deps, pypi deps, tasks, PACKAGE_DIR
[feature.prompt-da]             #   "
[feature.wilor]                 #   "
[feature.sam3d]                 #   "
[feature.sam3-rerun]            #   "
[feature.robocap]               #   "
[feature.pysfm]                 #   "
[feature.vistadream]            #   "
[feature.gsplat-rust-renderer]  #   "
[feature.pyvrs-viewer]          #   "

[environments]                  # Compose features into named environments (prod + dev per package)

Each package has a prod and dev environment sharing the same solve-group (no extra resolution cost). no-default-feature = true means each env is fully defined by its features.

Per-feature deps can tighten common's loose versions for their solve-group (e.g., vistadream tightens pytorch-gpu >= 2.10.0 over common's >= 2.8.0). Features compose by intersection, so a per-package feature can only narrow a shared constraint, never relax it.

pypi-options (workspace-level)

[pypi-options]
no-build-isolation = ["moge", "gsplat"]   # These need access to torch at build time

[pypi-options.dependency-overrides]
iopath = ">=0.1.10"      # Override transitive iopath pins
fsspec = ">=2025.3"       # Override transitive pins

Project structure

pixi.toml                         # Workspace manifest (features, envs, tasks, deps)
pyrefly.toml                      # Root pyrefly config (single source of truth for IDE + tasks)
.envrc                            # Root direnv config (activates dev environment)
packages/
  <name>/
    .envrc                        # Direnv config (activates <name>-dev environment)
    pyproject.toml                # [project] + build backend + [tool.ruff]
    <module>/                     # Python package
    tools/                        # CLI scripts (demos/, apps/)
    tests/

Note: sam3d-body-rerun uses tool/ (singular) not tools/.

pyrefly is configured only at the repo root — do not add [tool.pyrefly] to per-package pyproject.toml files.

Adding a new package

1. Create packages/<name>/ with pyproject.toml, source module, tools/, tests/
2. Add a feature in root pixi.toml:
   • [feature.<name>.dependencies] — Python pin + package-specific conda deps
   • [feature.<name>.pypi-dependencies] — editable install + git deps
   • [feature.<name>.activation.env] — set PACKAGE_DIR = "packages/<name>"
   • [feature.<name>.tasks.*] — demo/app tasks with cwd = "packages/<name>"
3. Add two environments in [environments]:
   • <name> = { features = ["common", "cuda", "<name>"], solve-group = "<name>", no-default-feature = true }
   • <name>-dev = { features = ["common", "cuda", "<name>", "dev"], solve-group = "<name>", no-default-feature = true }
4. Create packages/<name>/.envrc:

   watch_file ../../pixi.lock
   PIXI_ENV="${PIXI_ENV:-<name>-dev}"
   source_env_if_exists .envrc.local
   eval "$(pixi shell-hook -e "$PIXI_ENV" --manifest-path ../.. --frozen)"

5. Add packages/<name>/data/ to .gitignore
6. Run pixi install -e <name>-dev to verify

Gotchas

• Never use pip — all dependency management goes through Pixi
• hf download not huggingface-cli — conda's huggingface_hub provides hf, not huggingface-cli
• gradio from PyPI, not conda — conda's gradio package has missing transitive deps
• moge/gsplat need no-build-isolation — they require torch at build time (configured in [pypi-options])