Gridfinity Pic-to-Bin

Generate 3D-printable gridfinity bins with custom tool cutouts from phone camera photos.

A printed ArUco marker template handles perspective correction and automatic scale calibration. The rest of the pipeline — SAM2 segmentation, vectorization, layout packing, and Fusion 360 bin generation — runs automatically.

Contents

• Quick start
• Installation
  • Installing pipx
  • 1. The core code
  • 2. The web app
  • 3. The Fusion 360 add-in
  • Where to run these commands
  • pipx, venv, or plain pip
  • Upgrading and uninstalling
  • From source
• Step 1: Print the template
• Step 2: Take the photo
• Step 3: Run the pipeline
  • All pic-to-bin options
  • Bin sizing logic
  • Output files
• Web app (browser frontend)
• Running individual steps
• Fusion 360 integration
• Troubleshooting
• Running tests
• License

Quick start

1. Install the pipeline — pipx install gridfinity-pic-to-bin. One command, and it covers all three pieces: the CLI, the web app, and the bundled Fusion 360 add-in.
2. Print the ArUco template — generate-phone-template --paper-size letter. One-time setup; print at exactly 100% scale, no fit-to-page.
3. Photograph the tool on the template — lay the tool in the dotted zone, shoot from above with all 8 markers in frame, and measure the tool's depth with calipers.
4. Run the pipeline — pic-to-bin photo.jpg --tool-height 17. Produces bin_config.json, a combined DXF, and 1:1 fit-test printouts you can lay the real tool on before printing.
5. Build the bin in Fusion 360 — click Solid → Create → Gridfinity Pic-to-Bin, then export STL or STEP and slice it.

Prefer a browser to a terminal? pic-to-bin-web replaces step 4 with drag-and-drop upload, live progress, and a layout preview — see Web app.

Installation

Recommended — one command:

pipx install gridfinity-pic-to-bin

No extras, no second step: the web app's dependencies are part of the base package. (Earlier releases put them behind a [web] extra. That still resolves, so old instructions keep working — it just no longer adds anything the plain install lacks.)

pipx is the standard installer for Python applications — as opposed to libraries you import. It builds a private virtual environment for this package and puts its commands on your PATH, so there is no venv to create, activate, or remember, and nothing this package drags in can disturb the rest of your system. Don't have pipx yet? See Installing pipx. If you would rather use a venv and plain pip, that still works — see pipx, venv, or plain pip.

That single command covers all three pieces of the project:

There is no second install to run. pic-to-bin and pic-to-bin-web both come from that one command, and the Fusion add-in is bundled in the package itself — pic-to-bin-fusion install is a file copy into Fusion 360's own add-ins folder, not a download.

The server dependencies (fastapi, uvicorn, …) are a handful of small pure-Python packages next to the multi-GB PyTorch download that dominates the install either way, so there is nothing to save by leaving them out.

One feature is still optional: the "check with LLM" fit review, which calls the Anthropic API and bills your account per call. It is off unless you pass --enable-llm, and its SDK installs separately:

pipx install --force "gridfinity-pic-to-bin[llm]"

Keep those quotes — zsh, the default shell on macOS, reads [llm] as a glob pattern and answers zsh: no matches found before pipx ever runs. The quotes are harmless in bash, so quote it everywhere.

Install from wherever you like — but where you later run the commands does matter; see Where to run these commands.

Installing pipx

Windows — PowerShell, not "Run as Administrator":

py -m pip install --user pipx
py -m pipx ensurepath

macOS — Terminal:

brew install pipx        # or: python3 -m pip install --user pipx
pipx ensurepath

ensurepath edits your shell profile, so close the terminal and open a new one before continuing — that is what puts pipx, and the commands it installs, on your PATH.

pipx builds each app's environment with your default Python. This package needs Python 3.10 or newer; if your default is older, point pipx at a newer one: pipx install --python 3.12 gridfinity-pic-to-bin.

1. The core code

pipx install gridfinity-pic-to-bin

Everything that actually does the work: ArUco template generation, photo preprocessing (marker detection, perspective correction, scale calibration), SAM2 segmentation and vectorization, layout packing, the Fusion 360 config writer, and the add-in installer.

Install it if you are happy driving the pipeline from a terminal. It is also the base for both other pieces, so there is no way to skip it.

It gives you these commands:

Dependencies (installed automatically): ultralytics (SAM2), opencv-python, numpy, ezdxf, potracer, pyclipper, matplotlib, Pillow, pillow-heif, plus the web app's fastapi, uvicorn[standard], python-multipart, sse-starlette, and python-dotenv.

2. The web app

pipx install gridfinity-pic-to-bin   # the same single install as above
pic-to-bin-web --port 8000           # http://localhost:8000

This is not a second installation — it is the same package. The server and its dependencies come with the core install, so if pic-to-bin works, pic-to-bin-web works.

A FastAPI server plus a Lit frontend layered on top of the core code. It does not reimplement the pipeline — it calls the same run_pipeline() the CLI calls. What it adds is the interaction layer: drag-and-drop photo upload, live step-by-step progress streamed over SSE, an in-browser layout preview with printable fit-test downloads, a cheap Re-do loop that re-packs the layout without re-tracing, and one-click downloads of the DXF/PDF/JSON.

Install it if:

• You want a GUI instead of memorizing CLI flags.
• You want to review the layout preview and tweak parameters interactively before committing to a 3D print.
• You want to serve other people. The web app is multi-user by design: per-job UUID directories, a GPU semaphore so concurrent jobs queue instead of fighting over SAM2, and a TTL sweep for old jobs.

Serving dependencies: fastapi, uvicorn[standard], python-multipart, sse-starlette, python-dotenv — all core. The only piece held back is the anthropic SDK for the opt-in "check with LLM" review; add it with pipx install --force "gridfinity-pic-to-bin[llm]" if you plan to pass --enable-llm.

See Web app below for usage, and pic_to_bin/web/README.md for hosting it behind a public reverse proxy.

3. The Fusion 360 add-in

pic-to-bin-fusion install

No download and no second install — the add-in files ship inside the Python package. This command copies them into Fusion 360's own add-ins directory (%APPDATA%\Autodesk\Autodesk Fusion 360\API\AddIns\pic_to_bin\ on Windows, ~/Library/Application Support/Autodesk/Autodesk Fusion 360/API/AddIns/pic_to_bin/ on macOS). Nothing is installed into Fusion 360 itself — you need Fusion already installed, and it only runs on Windows and macOS.

Once enabled inside Fusion, a Gridfinity Pic-to-Bin button appears under Solid → Create. Clicking it reads a bin_config.json and builds the whole parametric bin — body, stacking lip, deck, tool pockets, finger slots, and gridfinity base pads — in a fresh document, with optional STL/STEP/PNG export.

Install it if you want the finished 3D model. Steps #1 and #2 stop at bin_config.json plus DXF/PDF files; this is the piece that turns those into printable geometry. Skip it if you only want the 1:1 fit-test printouts, or if you plan to import the DXF into some other CAD package yourself.

The command is idempotent — re-run it after upgrading the package to refresh the installed copy. pic-to-bin-fusion uninstall removes it.

See Fusion 360 integration below for enabling the add-in inside Fusion and what gets built.

Where to run these commands

pipx install does not care which directory you are in, and because pipx puts the commands on your PATH there is no environment to activate later either. But three things that come after the install do care about your current directory, so it is worth setting up one working folder now and running everything from there:

• pic-to-bin writes its results to generated/ relative to the current directory, and with no image arguments it processes every PNG/JPG in the current directory.
• pic-to-bin-web creates its web_jobs/ directory in the current directory.
• Ultralytics downloads the SAM2 weights (sam2.1_l.pt, several hundred MB) on the first trace, into the directory you ran from.

So: make a dedicated folder in your home directory and run every command from there.

Windows — PowerShell, not "Run as Administrator":

pipx install gridfinity-pic-to-bin   # once, from anywhere
mkdir $HOME\pic-to-bin
cd $HOME\pic-to-bin                         # then work from here

$HOME\pic-to-bin resolves to C:\Users\<you>\pic-to-bin.

macOS — Terminal:

pipx install gridfinity-pic-to-bin   # once, from anywhere
mkdir -p ~/pic-to-bin
cd ~/pic-to-bin                             # then work from here

Every later session, cd back to that folder before running pic-to-bin. With pipx there is nothing to activate — the commands are already on your PATH. (If you installed into a venv instead, re-activate it first: .\.venv\Scripts\Activate.ps1 on Windows, source .venv/bin/activate on macOS.)

Locations to avoid on both platforms:

pic-to-bin-fusion install is the one exception to the working-folder rule: it copies files into Fusion 360's own add-ins directory, so it can be run from anywhere.

pipx, venv, or plain pip

pipx is the recommended installer here because this package is an application — you run pic-to-bin from a terminal, you don't import pic_to_bin from your own code. pipx is built for exactly that case: it creates a dedicated virtual environment behind the scenes and exposes only the commands.

Why isolation matters for this package in particular:

• It is heavy. ultralytics pulls in PyTorch and torchvision — a couple of GB — and ultralytics, opencv-python and matplotlib each constrain the numpy version. Installed machine-wide with pip, that can break an unrelated project that wanted a different numpy or torch build.
• Some Pythons refuse a machine-wide install outright. Homebrew Python on macOS and most Linux distro Pythons mark themselves "externally managed" (PEP 668) and reject pip install outside a virtual environment with error: externally-managed-environment. pipx sidesteps this entirely; plain pip does not.
• Clean uninstall. pipx uninstall gridfinity-pic-to-bin takes all ~100 transitive dependencies with it. Undoing a machine-wide pip install means chasing them individually.
• Nothing to activate. A venv only works once you remember to activate it, in every new terminal, which is easy to get wrong when you also have to cd to a working folder. pipx commands are simply on your PATH.

When plain pip in a venv is still the right call:

• You are working on the source. Editable installs (pip install -e .) and the [dev] test dependencies belong in a venv — see From source.
• You want to import pic_to_bin from your own scripts. pipx's environment is deliberately not on your import path; a venv is.
• You need a specific PyTorch build, e.g. a CUDA wheel from download.pytorch.org. That is easier to control in a venv. It is still possible under pipx via pipx runpip gridfinity-pic-to-bin install torch --index-url …, just fiddlier.

For a venv install, replace the pipx install line with the usual three:

python3 -m venv .venv          # Windows: py -m venv .venv
source .venv/bin/activate      # Windows: .\.venv\Scripts\Activate.ps1
pip install gridfinity-pic-to-bin

On Windows, if Activate.ps1 is blocked by execution policy, run Set-ExecutionPolicy -Scope CurrentUser RemoteSigned once, then retry.

Further reading:

• pipx documentation
• Python docs — venv
• Python Packaging User Guide — installing packages with pip and virtual environments

Upgrading and uninstalling

pipx upgrade gridfinity-pic-to-bin      # keeps the spec you installed with
pipx uninstall gridfinity-pic-to-bin    # removes the package and every dependency

pipx records the exact spec you installed — including any extras — so pipx upgrade preserves it. To change the spec later (adding [llm], say), use pipx install --force with the one you want.

Two things live outside the package and are not touched by either command:

• The Fusion 360 add-in copy. Re-run pic-to-bin-fusion install after an upgrade to refresh it, and pic-to-bin-fusion uninstall before removing the package.
• The SAM2 weights (sam2.1_l.pt) and the generated/ and web_jobs/ folders in your working directory. Delete those by hand if you want the disk space back.

From source

pipx installs the tool for using; to work on the code itself, use a venv and an editable install. Clone wherever you normally keep code — C:\Users\<you>\source\ on Windows, ~/src on macOS — again avoiding cloud-synced folders:

git clone https://github.com/justinhardin/gridfinity-pic-to-bin.git
cd gridfinity-pic-to-bin
python3 -m venv .venv          # optional; Windows: py -m venv .venv
source .venv/bin/activate      # optional; Windows: .\.venv\Scripts\Activate.ps1
pip install -e .               # add ".[dev]" for the test deps

An editable install still runs from anywhere, so you can keep the checkout here and run pic-to-bin from your working folder — or run it from the checkout and let generated/ land next to the source (it is gitignored).

Step 1: Print the template

Generate and print a template for your paper size. Print at exactly 100% scale — no fit-to-page.

generate-phone-template --paper-size letter --output template.pdf

Options:
  --paper-size {a4,letter,legal}   Paper size (default: a4)
  --output PATH                    Output PDF path (default: phone_template_<size>.pdf)

The template places 8 ArUco markers (IDs 0–7) — 4 corners and 4 edge midpoints — around a dotted placement zone. The markers are 20 mm squares with ~20 mm margins from the paper edge.

Placement zone sizes by paper:

Step 2: Take the photo

1. Lay the printed template on a flat surface.
2. Place the tool inside the dotted placement zone.
3. Photograph from above. Moderate angles are fine — perspective is corrected automatically.
4. All 8 markers should be visible (3 minimum, 8 ideal).
5. Measure the tool's depth with calipers — you'll need it for --tool-height.

Lighting tips: Use diffuse overhead light. Avoid harsh shadows across the markers. Do not cover markers with the tool.

Format: JPEG, PNG, or HEIC/HEIF (iPhone) are all supported.

Step 3: Run the pipeline

Single tool

pic-to-bin photo.jpg --tool-height 17

Multiple tools (one photo each)

pic-to-bin a.jpg b.jpg --tool-height 0=17 --tool-height 1=14

Tool indices correspond to image order. Each tool's DXF is traced separately and packed into one bin.

Specifying paper size

pic-to-bin photo.jpg --tool-height 17 --paper-size a4

The paper size must match what you printed. Default is letter.

Shallow drawer (no stacking lip)

pic-to-bin photo.jpg --tool-height 17 --stacking false

Drops the 4.4 mm stacking lip for shorter bins in shallow drawers. Pocket depth is unchanged.

All pic-to-bin options

positional:
  images                        Photo files to process (default: all PNG/JPG in cwd)

required:
  --tool-height VALUE           Tool depth in mm. Use INDEX=VALUE per tool
                                (e.g. --tool-height 0=17 --tool-height 1=14)

optional:
  --paper-size {a4,letter,legal}  Template paper size (default: legal)
  --tolerance MM                  Extra clearance on top of a 2 mm baseline
                                  (default: 0 = 2 mm physical clearance).
                                  Positive = looser fit, negative = tighter,
                                  -2 = exact-trace match.
  --axial-tolerance MM            Extra clearance only along the tool's
                                  principal axis (default: 'auto'; 2 mm
                                  floor + taper-proportional bonus).
                                  Compensates for SAM2 length under-
                                  detection.
  --phone-height MM               Camera height above template, mm (default: 480).
                                  Drives the parallax-compensation scale-down.
  --gap MM                        Minimum gap between tools in layout, mm (default: 3.0)
  --bin-margin MM                 Extra clearance from tool extent to bin wall (default: 0)
  --min-units-x N                 Minimum X grid size in units (default: 1)
  --min-units-y N                 Minimum Y grid size in units (default: 1)
  --min-units-z N                 Minimum Z grid size in height units (default: 1).
                                  Floor on the auto height; ignored when
                                  --height-units is set.
  --max-units N                   Max gridfinity grid size per axis (default: 7)
  --height-units N                Force bin height in gridfinity units (default: auto)
  --stacking BOOL                 Generate stacking lip (default: true). Set
                                  false for a shorter bin without the lip.
  --slots BOOL                    Generate finger-access slots (default: true)
  --output-dir DIR                Output directory (default: generated/)
  --straighten-threshold DEG      Max degrees to auto-straighten trace (default: 45, 0=off)
  --max-refine-iterations N       SAM2 cleanup iterations (default: 5)
  --max-concavity-depth MM        Max acceptable concavity loss, mm (default: 3.0)
  --mask-erode MM                 Post-SAM mask erosion (default: 0). Use 0.3-0.5
                                  only if your photos have a clear shadow halo.
  --sam-model WEIGHTS             SAM2 model file (default: sam2.1_l.pt)
  --skip-trace                    Skip tracing, reuse existing DXFs in generated/

Bin sizing logic

• The bin auto-sizes to the smallest gridfinity unit count that fits the tool: ceil((tool_height + 1mm) / 7mm) height units.
• The pocket floor sits 1 mm above the bin floor.
• The deck rises to half the tool's height — the upper half of the tool stands proud for finger access; the lower half is buried in the pocket.
• The combined cutout (pocket + finger slot) is centered in the bin floor; slack from rounding up to whole gridfinity units is distributed evenly on all four sides.

Output files

generated/
  <stem>/
    <stem>_rectified.png       Perspective-corrected image (scanner equivalent)
    <stem>_trace.dxf           Tool outline DXF (inner + tolerance + finger slot)
    <stem>_trace.svg           SVG preview of the trace
  combined_layout.dxf          All tools packed into a bin footprint
  layout_preview.png           Screen-viewable layout preview (matplotlib, 150 DPI)
  layout_actual_size.pdf       1:1 scale fit-test drawing (PDF page = bin footprint)
  layout_actual_size.svg       1:1 scale fit-test drawing (SVG width/height in mm)
  bin_config.json              Fusion 360 input config

The layout_actual_size.pdf and .svg files are sized to the bin's exact mm dimensions — print at "Actual size" / 100% scale (NOT "Fit to page") and lay your real tool on top to verify the fit before committing to a 3D print.

After Fusion runs:

generated/
  gridfinity_bin.stl
  gridfinity_bin.step
  gridfinity_bin_preview.png  Viewport screenshot of the finished bin

Web app (browser frontend)

A FastAPI + Lit web wrapper exposes the same pipeline through a browser. Multi-user ready: per-job UUID directories, GPU semaphore around SAM2 so concurrent submissions queue rather than fight over the GPU, SSE-streamed progress.

pipx install gridfinity-pic-to-bin     # from a checkout: pip install -e .
pic-to-bin-web --port 8000

Open http://localhost:8000, drag in a photo, fill in the tool height, watch the step tracker, review the layout preview (with print-at-actual-size PDF / SVG downloads to test fit), then click Proceed to generate the bin config.

The browser back button navigates between screens (form → progress → preview → downloads). The form fields all have an (i) info button next to their label that opens a modal with a multi-paragraph explanation.

Lit is vendored inside the package (pic_to_bin/web/static/lit-all.min.js), so the page loads with no CDN request and works offline. To re-download it after deleting it, or to bump the pinned version:

python -m pic_to_bin.web.vendor_lit

Running individual steps

You can run each stage of the pipeline independently.

Preprocess a photo

Detect markers, correct perspective, and save the rectified image:

preprocess-phone photo.jpg --paper-size letter --output-dir generated/photo

positional:
  image                         Phone photo file

optional:
  --paper-size {a4,letter,legal}  Template paper size (default: a4)
  --output-dir DIR                Output directory (default: generated/<stem>)

Outputs <stem>_rectified.png at the computed effective DPI (typically 100–250).

Trace a rectified image

Run SAM2 segmentation + vectorization on a rectified image:

trace-tool rectified.png --dpi 150

Pack tool DXFs into a layout

layout-tools tool1.dxf tool2.dxf --gap 3 --max-units 5

Generate the Fusion 360 config

prepare-bin generated/combined_layout.dxf --tool-height 17

Fusion 360 integration

pic-to-bin ships in two flavors for Fusion 360 — a toolbar add-in (recommended) and a classic script. One install command sets up both:

pic-to-bin-fusion install

This copies the add-in to …/API/AddIns/pic_to_bin/ and the script to …/API/Scripts/pic_to_bin/, sharing the build code (_bin_builder.py) between them.

Add-in (recommended) — toolbar button

1. Open Fusion 360.
2. Press Shift+S → Add-Ins tab.
3. Select pic_to_bin → Run (toggle Run on Startup to keep the button available every session).
4. In a Design workspace, the Solid > Create panel now contains a Gridfinity Pic-to-Bin button.
5. Click the button. The script auto-loads <project>/generated/bin_config.json if it exists; otherwise a file dialog opens defaulting to your Desktop.
6. Bin gets built, exported as STL + STEP, and a viewport screenshot is saved alongside bin_config.json.

What gets built

The bin is generated in a fresh document with these timeline groups for easy navigation:

• Bin Body — outer rectangular block. Appearance: ABS (White).
• Stacking Lip (if enabled) — solid block + corner fillets + base-profile mating cutout + 0.6 mm top recess.
• Deck — recessed surface around the pocket.
• Tool Pockets — one cut per tool.
• Finger Slots — one cut per tool, same floor as the pocket.
• Base Pads — gridfinity baseplate-mating geometry, one Join extrude for all wide pads, one for the narrow posts, plus the two chamfers.

Reinstalling and reload

Re-running pic-to-bin-fusion install overwrites the add-in folder. The add-in entry point reloads _bin_builder.py from disk on every invocation, so most code changes land on the next button click without restarting Fusion. Only changes to the entry-point file itself (pic_to_bin_addin/pic_to_bin.py) require a Stop/Run on the add-in or a Fusion restart.

Uninstall

pic-to-bin-fusion uninstall

Removes the add-in, plus the Scripts/pic_to_bin folder left behind by versions that installed a separate script entry point.

Troubleshooting

Common photo issues

• Markers partially cut off: Keep all 8 markers visible. At 3 minimum the pipeline will run but accuracy drops.
• Blurry markers: Tap to focus on the template before shooting; avoid camera shake.
• Shadows on markers: Use diffuse or overhead lighting, not a side lamp.
• HEIC images (iPhone): Supported natively — the pipeline converts automatically via pillow-heif.

Running tests

python -m pytest tests/ -v

License

MIT — see LICENSE.md.