YOLO12: Attention-Centric Object Detection

Overview#

YOLO12, released in early 2025, introduces an attention-centric architecture that departs from the traditional CNN-based approaches used in previous YOLO models, yet retains the real-time inference speed essential for many applications. It achieves high object detection accuracy through changes to its attention mechanisms and overall network architecture. Despite those advantages, YOLO12 remains a community-driven release that may exhibit training instability, elevated memory consumption, and slower CPU throughput due to its heavy attention blocks, so Ultralytics recommends YOLO11 or YOLO26 for most production workloads.

Watch: How to Use YOLO12 for Object Detection with the Ultralytics Package | Is YOLO12 Fast or Slow?

Key Features#

• Area Attention Mechanism: A new self-attention approach that processes large receptive fields efficiently. It divides feature maps into l equal-sized regions (defaulting to 4), either horizontally or vertically, avoiding complex operations and maintaining a large effective receptive field. This significantly reduces computational cost compared to standard self-attention.
• Residual Efficient Layer Aggregation Networks (R-ELAN): An improved feature aggregation module based on ELAN, designed to address optimization challenges, especially in larger-scale attention-centric models. R-ELAN introduces:
  • Block-level residual connections with scaling (similar to layer scaling).
  • A redesigned feature aggregation method creating a bottleneck-like structure.
• Optimized Attention Architecture: YOLO12 streamlines the standard attention mechanism for greater efficiency and compatibility with the YOLO framework. This includes:
  • Using FlashAttention to minimize memory access overhead.
  • Removing positional encoding for a cleaner and faster model.
  • Adjusting the MLP ratio (from the typical 4 to 1.2 or 2) to better balance computation between attention and feed-forward layers.
  • Reducing the depth of stacked blocks for improved optimization.
  • Using convolution operations (where appropriate) for their computational efficiency.
  • Adding a 7x7 separable convolution (the "position perceiver") to the attention mechanism to implicitly encode positional information.
• Comprehensive Task Support: YOLO12 supports a range of core computer vision tasks: object detection, instance segmentation, image classification, pose estimation, and oriented object detection (OBB).
• Enhanced Efficiency: Achieves higher accuracy with fewer parameters compared to many prior models, demonstrating an improved balance between speed and accuracy.
• Flexible Deployment: Designed for deployment across diverse platforms, from edge devices to cloud infrastructure.

Supported Tasks and Modes#

YOLO12 supports a variety of computer vision tasks. The table below shows task support and the operational modes (Training, Validation, Inference, and Export) enabled for each:

All YOLO12 architectures support every mode once a trained checkpoint is available. The Pretrained Weights column indicates only whether Ultralytics publishes an official pretrained .pt on ultralytics/assets: for segmentation, pose, classification, and OBB, you must train your own checkpoint from the corresponding .yaml before running inference, validation, or export.

Performance Metrics#

YOLO12 improves accuracy at all model scales, with some trade-offs in speed compared to the fastest prior YOLO models. Below are quantitative results for object detection on the COCO validation dataset:

Detection Performance (COCO val2017)#

• Inference speed measured on an NVIDIA T4 GPU with TensorRT FP16 precision.
• Comparisons show the relative improvement in mAP and the percentage change in speed (positive indicates faster; negative indicates slower). Comparisons are made against published results for YOLOv10, YOLO11, and RT-DETR where available.

Usage Examples#

This section provides examples for training and inference with YOLO12.

The examples below focus on YOLO12 Detect models (for object detection). For other supported tasks (segmentation, classification, pose estimation, and oriented object detection), refer to the respective task-specific documentation: Segment, Classify, Pose, and OBB.

from ultralytics import YOLO

# Load a COCO-pretrained YOLO12n model
model = YOLO("yolo12n.pt")

# Train the model on the COCO8 example dataset for 100 epochs
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

# Run inference with the YOLO12n model on the 'bus.jpg' image
results = model("path/to/bus.jpg")

Key Improvements#

1. Enhanced Feature Extraction:

   • Area Attention: Efficiently handles large receptive fields, reducing computational cost.
   • Optimized Balance: Improved balance between attention and feed-forward network computations.
   • R-ELAN: Enhances feature aggregation using the R-ELAN architecture.

2. Optimization Innovations:

   • Residual Connections: Introduces residual connections with scaling to stabilize training, especially in larger models.
   • Refined Feature Integration: Implements an improved method for feature integration within R-ELAN.
   • FlashAttention: Incorporates FlashAttention to reduce memory access overhead.

3. Architectural Efficiency:

   • Reduced Parameters: Achieves a lower parameter count while maintaining or improving accuracy compared to many previous models.
   • Streamlined Attention: Uses a simplified attention implementation, avoiding positional encoding.
   • Optimized MLP Ratios: Adjusts MLP ratios to more effectively allocate computational resources.

Requirements#

The Ultralytics YOLO12 implementation, by default, does not require FlashAttention. However, FlashAttention can be optionally compiled and used with YOLO12. To compile FlashAttention, one of the following NVIDIA GPUs is needed:

• Turing GPUs (e.g., T4, Quadro RTX series)
• Ampere GPUs (e.g., RTX30 series, A30/40/100)
• Ada Lovelace GPUs (e.g., RTX40 series)
• Hopper GPUs (e.g., H100/H200)

Citations and Acknowledgments#

If you use YOLO12 in your research, please cite the original work by University at Buffalo and the University of Chinese Academy of Sciences:

@inproceedings{tian2025yolov12,
  title={YOLOv12: Attention-Centric Real-Time Object Detectors},
  author={Tian, Yunjie and Ye, Qixiang and Doermann, David},
  booktitle={Advances in Neural Information Processing Systems},
  volume={38},
  pages={78433--78457},
  year={2025},
  url={https://proceedings.neurips.cc/paper_files/paper/2025/file/7103444259031cc58051f8c9a4868533-Paper-Conference.pdf}
}

@software{yolo12,
  author = {Tian, Yunjie and Ye, Qixiang and Doermann, David},
  title = {YOLO12: Attention-Centric Real-Time Object Detectors},
  year = {2025},
  url = {https://github.com/sunsmarterjie/yolov12},
  license = {AGPL-3.0}
}

The YOLO12 paper was published in the NeurIPS 2025 proceedings, with a preprint on arXiv.

FAQ#