Official PyTorch implementation of YOLOv12.
Comparisons with others in terms of latency-accuracy (left) and parameter-accuracy (right) trade-offs.
YOLOv12: Attention-Centric Real-Time Object Detector.\
Yunjie Tian, Qixiang Ye, and David Doermann\
UPDATES 🔥
Abstract
Enhancing the network architecture of the YOLO framework has been long crucial yet focused on CNN-based improvements, despite the proven superiority of attention mechanisms in modeling capabilities. This is because attention-based models cannot match the speed of CNN-based models. This paper proposes an attention-centric YOLO framework, namely YOLOv12, that matches the speed of CNN-based ones while harnessing the performance benefits of attention mechanisms.
YOLOv12 surpasses all popular real-time object detectors in both speed and accuracy. For example, YOLOv12-N achieves $40.4$ mAP with an inference latency of $1.4$ ms on a T4 GPU, outperforming the advanced YOLOv10-N/YOLOv11-N by $1.9/1.0$ mAP and being $x.x\%/x.x\%$ faster. This advantage extends to other model scales. Furthermore, YOLOv12-S achieves comparable accuracy to RT-DETR-R18/xxxx while running $86\%/xx\%$ faster, using only $35\%/xx\%$ of the computation and $45\%/xx\%$ of the parameters
Main Results
COCO
| Model |
size
(pixels)
| mAPval
50-95
| Speed
T4 TensorRT10
| params
(M)
| FLOPs
(B)
|
| YOLO12n |
640 |
xx.x |
x.xx |
2.5 |
6.6 |
| YOLO12s |
640 |
xx.x |
x.xx |
8.9 |
22.0 |
| YOLO12m |
640 |
xx.x |
x.xx |
19.9 |
69.7 |
| YOLO12l |
640 |
xx.x |
x.xx |
28.3 |
97.2 |
| YOLO12x |
640 |
xx.x |
xx.x |
63.2 |
216.5 |
Installation
wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu11torch2.2cxx11abiFALSE-cp311-cp311-linux_x86_64.whl
conda create -n yolov12 python=3.11
source activate yolov12
pip install -r requirements.txt
pip install -e .
Validation
yolov12n
yolov12s
yolov12m
yolov12l
yolov12x
from ultralytics import YOLO
model = YOLO.from_pretrained('sunsmarterjie/yolov12{n/s/m/b/l/x}')
# or
# wget https://github.com/sunsmarterjie/yolov12/releases/download/v1.0/yolov12{n/s/m/l/x}.pt
model = YOLO('yolov12{n/s/m/b/l/x}.pt')
model.val(data='coco.yaml', batch=128)
Training
from ultralytics import YOLO
model = YOLO('yolov12n.yaml')
# Train the model
results = model.train(
data='coco.yaml',
epochs=600,
batch=128,
imgsz=640,
device="0,1,2,3",
)
# Evaluate model performance on the validation set
metrics = model.val()
# Perform object detection on an image
results = model("path/to/image.jpg")
results[0].show()
# Export the model to ONNX format
path = model.export(format="onnx") # return path to exported model
Finetuning
# If you want to finetune the model with pretrained weights, you could load the
# pretrained weights like below
# model = YOLO.from_pretrained('sunsmarterjie/yolov12{n/s/m/l/x}')
# or
# wget https://github.com/sunsmarterjie/yolov12/releases/download/v1.0/yolov12{n/s/m/l/x}.pt
# model = YOLO('yolov12{n/s/m/b/l/x}.pt')
Prediction
from ultralytics import YOLO
model = YOLO.from_pretrained('sunsmarterjie/yolov12{n/s/m/l/x}')
# or
# wget https://github.com/sunsmarterjie/yolov12/releases/download/v1.0/yolov12{n/s/m/l/x}.pt
model = YOLO('yolov12{n/s/m/b/l/x}.pt')
model.predict()
Export
from ultralytics import YOLO
model = YOLO.from_pretrained('sunsmarterjie/yolov12{n/s/m/l/x}')
# or
# wget https://github.com/sunsmarterjie/yolov12/releases/download/v1.0/yolov12{n/s/m/l/x}.pt
model = YOLOv10('yolov12{n/s/m/b/l/x}.pt')
model.export(...)
Demo
python app.py
# Please visit http://127.0.0.1:7860
Acknowledgement
The code base is based on ultralytics. Thanks for their excellent work!
Citation
If our code or models help your work, please cite our paper:
@article{tian2025yolov12,
title={YOLOv12: Attention-Centric Real-Time Object Detectors},
author={Tian, Yunjie and Ye, Qixiang and Doermann, David},
journal={arXiv preprint arXiv:2502.xxxxx},
year={2025}
}
