Roboflow

Best Instance Segmentation Models

Compare the best 15 instance segmentation models and try 6 of them on your own image, free in the Roboflow Playground. 15 are open-weight, so you can self-host them for free under their licenses.

15 models · 15 open-weight · 6 free to try · prices synced Aug 3, 2026

We haven't benchmarked instance segmentation yet; scores and rankings appear on this site only where we've measured them. Until then, compare the models below, and for fixed categories in production expect a model fine-tuned on your own data to win.

Newest

Meta

SAM 3

Most recently added (Nov 2025)

Open-Source Instance Segmentation Models

15 models with downloadable weights you can self-host under their licenses (Custom, Apache 2.0, and AGPL 3.0). 6 run live in the Playground through hosted APIs, so self-hosting is optional.

Meta
SAM 3Meta
—CustomNov 2025
Try
Meta
SAM 3D ObjectsMeta
—CustomNov 2025
RF-DETR SegmentationRoboflow
33.6M-38.6MApache 2.0Oct 2025
Try
YOLO26Ultralytics
2.4M-55.7MAGPL 3.0Oct 2025
Try
Azure
Florence-2Microsoft
230MMITJun 2025
Try
YOLOETHU-MIG
10M-50MAGPL 3.0Mar 2025
YOLOv12THU-MIG
2.6M-59.1MAGPL 3.0Feb 2025
YOLO11Ultralytics
2.6M-56.9MAGPL 3.0Sep 2024
Try

How to Choose an Instance Segmentation Model

Instance segmentation models differ mainly in how masks get requested: trained on fixed classes, prompted interactively, or driven by text. Which one you want depends on whether segmentation is a production pipeline or a labeling and editing workflow.

RF-DETR Seg, Mask R-CNN lineage models, and YOLO segmentation variants learn to segment your specific classes and run fast enough for production video and high-volume pipelines. Like detectors, they need labeled masks to train, but they then produce consistent instance masks with no prompting, which is what automated measurement, counting, and robotics pipelines need.

SAM-family models segment almost anything from a click, box, or, in newer variants, a text phrase, with no training. They are the fastest way to get pixel-accurate masks for labeling datasets, cutting objects out in editing tools, and one-off analysis. The tradeoff: they do not know your classes, so something else (a person, a detector, or a text prompt) has to decide what to segment.

The standard workflow uses a promptable segmenter to build a mask dataset quickly, then trains a compact segmentation model on those masks for production. You get SAM-quality boundaries in the dataset and millisecond inference in deployment.

The bottom line: Interactive masks and labeling: SAM family. Fixed classes at production speed: train a segmentation model, ideally on masks SAM helped you label.

Frequently Asked Questions About Instance Segmentation Models

Instance segmentation is the task of producing a pixel-accurate mask, class label, and score for each individual object in an image. It is strictly finer than object detection (masks instead of boxes) and finer than semantic segmentation because touching or overlapping objects of the same class stay separate: ten cells become ten masks, not one blob. Classic approaches like Mask R-CNN add a mask head to a detector, while transformer models like Mask2Former treat boxes and masks as one set-prediction problem. Quality is measured by mask AP, mean average precision computed on mask overlap rather than box overlap. The pixel-exact boundaries matter for robotic grasping, medical and biological imaging, defect measurement, and image editing. This page lists 15 instance segmentation models, including 15 open-weight options you can self-host; 6 of them run live in the Playground so you can test them on your own images.

It depends on your task and constraints. For fixed categories in production, a model fine-tuned on your own data typically beats any general-purpose model. Compare the instance segmentation models on this page and try them on your own image to see which fits.

Yes. 15 of the 15 instance segmentation models here are open-weight (for example SAM 3, SAM 3D Objects, and RF-DETR Segmentation), free to self-host under their licenses (Custom, Apache 2.0, and AGPL 3.0).

Yes. You can run 6 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.

This page lists all 15 instance segmentation models in the Roboflow Playground catalog, all of them open-weight and free to self-host; 6 of them run live in the Roboflow Playground on your own images. Compare licenses, parameters, and release dates side by side, or open any model page for full details.