01
Prompt the way you think
Text, include-boxes, exclude-boxes — and combinations. SAM understands "person," "label," "dog." Shift-click to draw an and-not region.
v1.0 · macOS 14+ · Apple Silicon
SAM 3 · SAM 3.1 · MLX · Metal · Accelerate
Segment anything. On your Mac. In bulk.
Contour turns Meta's SAM 3 and SAM 3.1 into a real macOS workspace — drop in a folder, write a prompt, get clean masks, boxes, and dataset-ready exports.
- Free for personal use
- Pro · $49 one-time
- 1.7 GB model
412ms median infer ~200 instances/image 0 bytes uploaded
IMG_0412.jpg · SAM 3.1
5 instances · 412ms
Before ← drag → After
Features
An instrument, not a magic wand.
Every feature is there because someone who segments images for a living needed it. No confetti, no auto-enhance, no one-click magic.
02
Instant threshold tuning
Raw output is cached per image. Drag confidence, NMS, alpha, and minimum area — results re-render in hundreds of milliseconds, no re-inference.
03
Batch a whole collection
Point at a folder, set a prompt and model once, run. Status is tracked in SwiftData. Completed images skip on re-runs; crashes recover cleanly.
04
Three overlay modes
Pixel mask, axis-aligned bbox, or oriented bbox — switch per export.
05
On device, via MLX
MLX for compute, Metal for GPU, Accelerate for SIMD mask composition. No cloud, no API key, no telemetry.
06
Real export formats
COCO JSON (with optional RLE), YOLO TXT bbox or polygon, mask PNG, transparent cutout PNG. Optional source-image copy.
Models
Pick the precision that fits the box.
Six SAM variants ship ready to cache. Downloaded once from Hugging Face, color-coded in the toolbar so you always know what's loaded.
SAM 3 full
- Size
- ~1.7 GB
- Mem
- 8 GB+
- Infer
- 412 ms
Reliable baseline. Fast on most Apple Silicon.
SAM 3.1 full
- Size
- ~3.5 GB
- Mem
- 16 GB+
- Infer
- 680 ms
Latest. Adds box-geometry prompts.
SAM 3 · int8 quant
- Size
- ~900 MB
- Mem
- 8 GB+
- Infer
- 298 ms
Quantized · near-lossless for most prompts.
SAM 3 · int4 quant
- Size
- ~450 MB
- Mem
- 8 GB+
- Infer
- 215 ms
Smallest. Trade accuracy for throughput.
SAM 3.1 · mxfp8 μfp
- Size
- ~900 MB
- Mem
- 8 GB+
- Infer
- 340 ms
Microscaling FP — modern low-precision.
SAM 3.1 · mxfp4 μfp
- Size
- ~500 MB
- Mem
- 8 GB+
- Infer
- 240 ms
Smallest 3.1 variant. Box prompts supported.
Workflow
Five steps, from folder to dataset.
No modal popups. No "AI magic." Each step is a tool in the workstation — toolbar-driven, menu-driven, keyboard-driven.
- 01
Collect
Drag a folder of images into a new collection. SwiftData tracks per-image status.
- 02
Pick a model
SAM 3 or 3.1, full-precision or quantized. Color-coded in the toolbar.
- 03
Prompt
Type a label, draw include / exclude boxes, or combine.
- 04
Tune
Sliders re-render from cache — no re-inference while you dial it in.
- 05
Export
Mask PNG, cutout PNG, COCO JSON, or YOLO TXT. Pick a folder, run it.
Exports
Real formats. Nothing to reshape.
What comes out of Contour drops straight into your training pipeline — no round-tripping, no script-to-reshape, no CSV middleware.
COCO · annotations.json
COCO-format instance annotations with optional RLE-encoded masks.
{
"images": [...],
"annotations": [
{
"id": 1,
"image_id": 412,
"category_id": 1,
"bbox": [142,288,72,312],
"area": 18472,
"segmentation": {
"counts": "PYm03L4M...",
"size": [1920,1080]
},
"score": 0.97
}
]
} YOLO · IMG_0412.txt
YOLO txt — bbox or polygon format. Douglas-Peucker simplified.
0 0.186 0.554 0.038 0.520
0 0.452 0.554 0.038 0.520
0 0.742 0.576 0.037 0.510
0 0.867 0.598 0.036 0.498
0 0.092 0.620 0.024 0.391
# class_id cx cy w h
# normalized 0..1
# 5 detections MASK · IMG_0412_mask.png
Per-instance binary alpha masks. One PNG per detection.
IMG_0412/
├─ mask_001.png · 62 KB
├─ mask_002.png · 58 KB
├─ mask_003.png · 61 KB
├─ mask_004.png · 47 KB
└─ mask_005.png · 34 KB
5 masks · 262 KB total
1920×1080 · 8-bit alpha CUTOUT · IMG_0412_cut.png
Transparent-background cutouts — ready for compositing and matting.
IMG_0412/
├─ cutout_001.png · 240 KB
├─ cutout_002.png · 221 KB
├─ cutout_003.png · 204 KB
├─ cutout_004.png · 178 KB
└─ cutout_005.png · 132 KB
5 cutouts · RGBA
alpha channel preserved On device
Your client's unreleased imagery never leaves the machine.
Every model runs locally via MLX — Apple's machine learning framework for Apple Silicon — with Metal for GPU compute and Accelerate for SIMD-accelerated mask composition.
- No cloud inference. MLX runs on your GPU, not someone else's.
- No API key. No quota. No sign-in to segment.
- No telemetry. Contour makes zero outbound requests once models are cached.
- Model weights downloaded once from Hugging Face. Silent thereafter.
- Collections, masks, and caches stored in the app's SwiftData container — yours.
- Sandboxed. App-Store-ready. Notarized and hardened runtime.
Pricing
Free to try. One-time for Pro.
No subscription. Buy Pro once, own it. Free tier covers personal projects and exploration — Pro unlocks batch, exports, and the 3.1 quantizations.
Free
$0 forever
Personal use. One image at a time.
- SAM 3 · quantized (int8)
- Text + box prompts
- Slider tuning with live re-render
- Mask PNG + cutout PNG export
- Up to 50 detections per image
- One collection, 100 images
Pro
ONE-TIME $49 pay once · yours
Everything. For people who segment for a living.
- All models · SAM 3, 3.1, int8, mxfp8, mxfp4
- Batch processing across collections
- COCO JSON (with RLE) + YOLO TXT export
- OBB · oriented bounding box output
- Up to 200 detections per image
- Unlimited collections and images
- All future 1.x updates
FAQ
Questions we actually get.
Missing one? The app's Help menu opens a short, searchable manual — or email support@mgcrea.io .
Q01 Does Contour need an internet connection?
Only once — to download model weights from Hugging Face on first use. After models are cached locally, Contour runs entirely offline. The app makes zero outbound requests during segmentation.
Q02 Which Macs can run it?
macOS 14 (Sonoma) or later on Apple Silicon. 8 GB of unified memory for quantized SAM 3; 16 GB+ recommended for full-precision SAM 3.1. Intel Macs are not supported — MLX and Metal require Apple Silicon.
Q03 How is this different from a one-click background remover?
A background remover is about the outline — one subject, one clean cutout. Contour is about the partition — many instances, each one identified, each one bounded, exportable as a dataset. Different problem, different tool.
Q04 Can I script it or automate batches?
Pro includes CLI export and AppleScript / Shortcuts bindings for running collections headlessly. SwiftData is queryable via the app's URL scheme.
Q05 Do I need to accept Meta's SAM license?
Yes. Contour gates SAM downloads behind a one-time license-acceptance prompt. After you accept, subsequent model pulls are silent.
Q06 What about my training data — is anything collected?
Nothing. Contour ships with no analytics, no crash reporter, no "help improve the product" telemetry. Images, masks, and caches live in the app's sandboxed SwiftData container on your machine.
Q07 Will there be a Linux or Windows version?
No. MLX is Apple-Silicon-only, and the app is built around macOS idioms — SwiftUI, SwiftData, Quick Look, Finder integration. A web or Linux port would be a different product.