Why you need it

• Netfabb shut down. Autodesk retired its mesh-repair tool and there is no modern free replacement for printability pre-check — until now.
• "Upload and print" fails. Non-manifold meshes, thin walls, overhangs and holes waste hours of print time and filament.
• Print shops quote by hand. Weight, time, material, electricity and labor estimated manually — slow and inconsistent.
• Failed prints stay a mystery. Stringing? Warping? Z-band? New users stare at a failed part and don't know which knob to turn.

Print Doctor answers all three: is this model printable? what does it cost? what went wrong?

One CLI, three moments of a print

Showcase

Check & quote (before printing)	Diagnose (after a failure)
Watch (during printing)	3D defect report (interactive)

Features

Printability check (check)

Detects non-watertight meshes, non-manifold edges, inverted normals, degenerate faces, thin walls (ray-cast), overhangs, self-intersections, isolated components and sliver triangles — each with a severity level and an actionable fix. Outputs a 0-100 score.

Interactive 3D report (check --3d): a self-contained HTML file with the model rendered in 3D and defect faces highlighted by type (overhang red, thin-wall yellow, self-intersection purple). Drag to rotate, scroll to zoom — no network needed to view.

Mesh repair (repair)

Fixes common issues automatically: face normals, winding, stitching and degenerate faces. Reports honestly what it cannot fix (large holes, self-intersections — use a dedicated tool for those).

Print orientation (orient)

Finds a good orientation before slicing: searches rotations, scores overhang vs bed contact, and recommends "rotate X 30°" — or saves the re-oriented model. A fast sanity check before you commit to a direction.

Hollowing (hollow)

Saves material: hollows a model with a chosen wall thickness and reports the percentage saved (e.g. a 1.5mm shell of a benchy saves ~66% material). Rejects models too thin for the requested wall.

Support estimation (support)

Estimates how much support material a model needs (overhang area → projected height → volume → weight → cost). Directly useful for quoting prints with supports. Estimate only — calibrate the density factor against your slicer.

G-code analysis (gcode-info)

Parses sliced G-code (OrcaSlicer/PrusaSlicer ;LAYER_CHANGE and Bambu ;LAYER:n) to report layer count, max Z, total extrusion, and locate the current print state at any E position (layer + progress %).

Cost quoting (check --quote, quote-sheet)

Full shop pricing: material, electricity, machine depreciation, labor, waste allowance and a suggested retail price. Emits Markdown, HTML, or machine-readable JSON — print-shop ready.

Real-time monitoring (watch)

Samples a webcam, a watched directory, or a Moonraker (Klipper) snapshot URL, classifies each frame with ML, and alerts on defects — saving an evidence screenshot and optionally POSTing a webhook.

Progress-aware: give it your sliced G-code plus progress (manual --progress or auto from Moonraker --moonraker http://printer:7125) and the monitor shows [47% · layer 123] alongside defect alerts — so you know where in the print a failure happened. Ships with an OctoPrint plugin (octoprint-print-doctor/) that can pause the print on defect.

Photo diagnosis (diagnose)

Trained on real photos scraped from 3D Printing StackExchange. Classifies stringing, warping, layer shift, under/over-extrusion and first-layer failure, then explains why and what to change.

Honest status: healthy-vs-defect detection is reliable (~100% accuracy, 0% false positives on healthy prints). Classifying which specific defect is still prototype-grade: strict top-1 accuracy is ~0.3 (unaugmented) / ~0.5 (data-augmented), because the training labels are weak — derived from forum post text, not verified against each image. Treat the defect type as a ranked candidate, not ground truth.

Diagnosis also reports anomaly region localization (best-effort bounding boxes of where something looks unusual) — see the "Anomaly Regions" table in the report.

Shop & batch workflows

• check-batch — analyze directories, sortable comparison table
• --json — stable machine-readable schema (integration / APIs)
• quote-sheet — customer-facing printable HTML quote
• Exit codes for CI (critical issues → exit 1; defects → exit 2)

Quick start

pip install print-doctor

# Check a model (printability score + cost estimate)
print-doctor check model.stl

# Diagnose a failed print from a photo
print-doctor diagnose failed_print.jpg

# Monitor your printer's webcam in real time
print-doctor watch 0

Requires Python 3.11+. Diagnosis needs OpenCV: pip install print-doctor[vision].

How it compares

Full usage

Check a model

print-doctor check model.stl                    # analysis + cost (PLA)
print-doctor check model.stl -m PETG -p 30      # pick material/price
print-doctor check model.stl -o report.md       # save report
print-doctor check model.stl --html -o r.html   # HTML report
print-doctor check model.stl --json             # machine-readable JSON
print-doctor check model.stl --quote            # full shop pricing

Exit code 1 when the model has critical issues → CI friendly.

Diagnose a failed print

print-doctor diagnose print_1.jpg
print-doctor diagnose top.jpg side.jpg -m PLA --temperature 210
print-doctor diagnose failed_print.jpg -o diagnosis.md

Exit code 2 when defects are detected.

Monitor in real time

print-doctor watch 0                                      # webcam
print-doctor watch ./snapshots -i 5                       # photo dir
print-doctor watch "http://printer:7125/server/webcams/snapshot?name=webcam" \
  -i 5 -e ./evidence -w https://hooks.example.com/alert   # Moonraker

Quote for a customer

print-doctor quote-sheet model.stl --shop "My Print Shop" \
  --customer "Alice" --quote-number Q-001 -o quote.html

Batch analysis

print-doctor check-batch ./models/ --quote --json -o summary.json

Python API

from print_doctor import check, estimate_cost, diagnose

analysis = check("model.stl")                    # MeshAnalysis
est = estimate_cost(14.6, PrintConfig(material_type="PLA"), 25.0, 0.12, 200.0)
diag = diagnose(["photo.jpg"], hints={"material": "PLA"})

Reproducing the ML classifier

Trained on 600 real photos (URLs + weak labels in data/stackexchange_manifest.json):

python scripts/fetch_dataset.py --out /tmp/dataset/stackexchange
python scripts/train_classifier.py --data /tmp/dataset/stackexchange --augment

Measured honestly (5-fold CV, strict top-1): healthy-vs-defect ~100% (0% false positives); which-defect ~0.3 unaugmented / ~0.5 augmented. Labels are weak (forum-text-derived) and partly noisy — scripts/audit_dataset.py lists suspect samples for human review. Accuracy improves with cleaner labels, not just more images.

Print-farm integration

• OctoPrint — see octoprint-print-doctor/ for a plugin (monitor on print start, tab alerts, optional pause-on-defect).
• Klipper / Moonraker — point watch at the webcam snapshot URL (see octoprint-print-doctor/README.md).

Development

uv sync --group vision
uv run pytest
uv run python tests/generate_fixtures.py
uv run python tests/generate_diagnose_fixtures.py

Development & AI disclosure

This project is developed with heavy AI assistance. Concretely:

• AI assistants were used for code generation, test writing, documentation, refactoring, and debugging across the majority of commits.
• The ML defect classifier was trained on real forum photos with weak labels; its measured limitations (healthy-vs-defect ~100% reliable, which-defect top-1 ~0.3-0.5) are documented honestly in this README and the docs — see How it works.
• Core algorithms (mesh checks, cost model) and all quantitative claims are human-reviewed; no claim in this repo is machine-asserted without inspection.

We disclose this because the project targets the 3D printing community, where trust in tool outputs matters. If you find a claim that overstates what the tool does, please open an issue — accuracy over hype.

License

MIT License