LibreYOLO documentation

All tasks

Object detection

Boxes and classes. The default task, and the one most families serve.

Instance segmentation

Per-object masks, one mask per detection.

Semantic segmentation

A class for every pixel, without separating objects.

Panoptic segmentation

Semantic and instance segmentation in one output.

Pose estimation

Keypoints per object, top-down or bottom-up.

Image classification

One label per image, including zero-shot with CLIP and SigLIP2.

Oriented detection

Rotated boxes, for aerial and document imagery.

Point detection

Centroids instead of boxes. Counting, and very small models.

Depth estimation

Per-pixel depth from a single image.

Edge detection

Surface normals

Image restoration

Denoising, deblurring and upscaling.

Background removal

Alpha mattes for cutouts.

OCR

Text detection and recognition in one pass.

Face recognition

Face embeddings and identity galleries.

Gaze estimation

Where a person is looking.

Body mesh

Object tracking

Identities across video frames, over any detector.

Open-vocabulary detection

Detect classes you name at runtime, with no training.

Promptable segmentation

Segment anything you point at, with SAM and friends.

Start with a flagship unless you have a reason not to. Every feature is designed and GPU-validated against those first, and each page carries the checkpoints, the export matrix and the licensing for that family.

RF-DETR

4 tasks, 19 checkpoints.

YOLOv9

Detection, 9 checkpoints.

DEIM

Detection, 13 checkpoints.

D-FINE

2 tasks, 10 checkpoints.

EdgeCrafter

3 tasks, 12 checkpoints.

RT-DETR

Detection, 16 checkpoints.

YOLO-NAS

2 tasks, weights distributed by their authors, not by us.

ConvNeXt

Detection, 3 checkpoints.

DINOv2

3 tasks, weights distributed by their authors, not by us.

EfficientNetV2

Detection, 4 checkpoints.

FOMO

Detection, 3 checkpoints.

LingBot-Vision

Detection, 3 checkpoints.

MobileNetV4

Detection, 3 checkpoints.

NAFNet

Detection, 1 checkpoint.

PicoDet

Detection, 3 checkpoints.

ResNet

Detection, 4 checkpoints.

RTMDet

2 tasks, 10 checkpoints.

SegFormer

Detection, 6 checkpoints.

YOLOv7

Detection, 1 checkpoint.

YOLOX

Detection, 6 checkpoints.

AlexNet

Detection, 1 checkpoint.

BiRefNet

Detection, 1 checkpoint.

CenterNet

Detection, 2 checkpoints.

DeepLabv3

Detection, 3 checkpoints.

Deformable DETR

Detection, 5 checkpoints.

Depth Anything 3

Detection, 1 checkpoint.

Depth Anything V2

Detection, 3 checkpoints.

DETR

Detection, 4 checkpoints.

DexiNed

Detection, weights distributed by their authors, not by us.

DINO-DETR

Detection, 3 checkpoints.

EfficientDet

Detection, 5 checkpoints.

EoMT

3 tasks, 6 checkpoints.

Faster R-CNN

Detection, 4 checkpoints.

FCN

Detection, 2 checkpoints.

FCOS

Detection, 1 checkpoint.

FeyNobg

Detection, 3 checkpoints.

HRNet

Detection, 2 checkpoints.

L2CS-Net

Detection, weights distributed by their authors, not by us.

LibreFaceRec

Detection, weights distributed by their authors, not by us.

LW-DETR

Detection, 5 checkpoints.

Mask R-CNN

2 tasks, 1 checkpoint.

MiDaS

Detection, weights distributed by their authors, not by us.

MoGe-2

Detection, weights distributed by their authors, not by us.

PIDNet

Detection, 3 checkpoints.

PP-OCRv5

Detection, 2 checkpoints.

Real-ESRGAN

Detection, 3 checkpoints.

RetinaNet

Detection, 2 checkpoints.

SAM 3D Body

Detection, 2 checkpoints.

SSD

Detection, 1 checkpoint.

Swin Transformer

Detection, 4 checkpoints.

SwinIR

Detection, 3 checkpoints.

TEED

Detection, weights distributed by their authors, not by us.

VGG

Detection, 4 checkpoints.

ViT

Detection, 4 checkpoints.

ZipDepth

Detection, 2 checkpoints.

DeiT

Detection, 3 checkpoints.

YOLOv1

Detection, 1 checkpoint.

YOLOv2

Detection, 2 checkpoints.

YOLOv3

Detection, 3 checkpoints.

YOLOv4

Detection, 2 checkpoints.

CLIP

2 tasks, 2 checkpoints.

EdgeTAM

Detection, 1 checkpoint.

Florence-2

Detection, weights distributed by their authors, not by us.

Grounding DINO

Detection, 2 checkpoints.

InternVL3

Detection, weights distributed by their authors, not by us.

Kosmos-2

Detection, weights distributed by their authors, not by us.

LFM2-VL

Detection, weights distributed by their authors, not by us.

LibreMODUS

4 tasks, weights distributed by their authors, not by us.

LocateAnything

2 tasks, weights distributed by their authors, not by us.

MobileSAM

Detection, 1 checkpoint.

OMDet-Turbo

Detection, 1 checkpoint.

OV-DEIM

Detection, 3 checkpoints.

OWLv2

Detection, 2 checkpoints.

PicoSAM3

Detection, 1 checkpoint.

Qwen3-VL

Detection, weights distributed by their authors, not by us.

SAM

Detection, weights distributed by their authors, not by us.

SAM 2

Detection, 4 checkpoints.

SAM 3

Detection, weights distributed by their authors, not by us.

SenseNova-Vision

7 tasks, 1 checkpoint.

SigLIP2

2 tasks, 2 checkpoints.

SmolVLM2

Detection, weights distributed by their authors, not by us.

Every accuracy and latency number in these docs comes from a run we publish, with the hardware, runtime and precision recorded alongside it. The chart below is live.

Train

Datasets, arguments, augmentation, multi-GPU and experiment loggers.

Predict

Images, folders, video, webcams and RTSP streams.

Export and deploy

Twelve targets, with a support matrix per family.

Command line

Everything the Python API does, without writing Python.

LibreYOLO's code is MIT. It does not require you to open source your application, it does not reach your model weights, and it does not change if you sell what you build. Pretrained weights are separate: each one carries the license of whoever trained it, and the docs state that per checkpoint rather than averaging it into a claim. A few are non-commercial, and those say so.

How licensing works here