YOLOv13 – Hypergraph-Enhanced Real-Time Object Detection

YOLOv13 is a real-time object detection model introduced in 2025 by Mengqi Lei, Siqi Li, Yihong Wu, Han Hu, You Zhou, Xinhu Zheng, Guiguang Ding, Shaoyi Du, Zongze Wu, and Yue Gao. It was presented in the research paper YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception.

YOLOv13 focuses on improving how a neural network models relationships between visual features. Earlier YOLO architectures mainly rely on local feature aggregation or pairwise relationships. YOLOv13 instead introduces hypergraph-based modeling to capture more complex multi-to-multi, high-order relationships between visual features.

Its main innovations include:

• HyperACE
• FullPAD
• Hypergraph computation
• High-order feature correlation modeling
• Cross-location feature fusion
• Cross-scale feature fusion
• Depthwise separable convolution blocks
• Reduced parameters and computational cost
• Real-time object detection

What Is YOLOv13?

YOLOv13 is a lightweight real-time object detector designed to improve both feature representation and information flow across the network.

The model addresses a limitation found in many previous object detectors: visual relationships are often modeled only locally or between two features at a time.

YOLOv13 introduces a hypergraph-based approach that allows the network to model more complex relationships involving multiple visual regions simultaneously.

This is especially useful in complex scenes where understanding relationships between multiple objects, locations, and feature scales can improve detection quality.

YOLOv13 Main Innovations

The official YOLOv13 project highlights three major architectural ideas:

HyperACE

HyperACE stands for:

Hypergraph-based Adaptive Correlation Enhancement

It is designed to capture high-order relationships between features.

FullPAD

FullPAD stands for:

Full-Pipeline Aggregation-and-Distribution

It distributes correlation-enhanced features across different parts of the network.

DS-Based Lightweight Blocks

YOLOv13 replaces certain large-kernel convolutions with depthwise separable convolution based modules to reduce parameter count and computation.

What Is HyperACE?

HyperACE is one of the most important components of YOLOv13.

The mechanism treats pixels or feature positions from multi-scale feature maps as vertices in a hypergraph.

A standard graph typically connects two nodes with an edge.

A hypergraph is more flexible because a single hyperedge can connect multiple nodes simultaneously.

This allows YOLOv13 to represent multi-to-multi relationships rather than only pairwise relationships.

Why Hypergraph Computation Matters

Images often contain relationships involving more than two regions.

For example, detecting an object may depend on:

• Nearby objects
• Background context
• Multiple spatial locations
• Features at different scales
• Structural relationships between several image regions

Pairwise modeling can capture relationships between two features, but hypergraph modeling can represent a larger group of related features at the same time.

YOLOv13 uses this idea to improve visual understanding in complex scenes.

Learnable Hyperedge Construction

HyperACE includes a learnable hyperedge construction module.

Instead of manually defining relationships between feature locations, the network learns which features should be grouped together.

This allows YOLOv13 to adaptively discover meaningful high-order relationships during training.

The learned hyperedges can connect information across:

• Different spatial locations
• Different object regions
• Different feature-map scales

Hypergraph Message Passing

After constructing hyperedges, YOLOv13 uses message passing to exchange information between connected features.

The official project describes a message passing mechanism with linear complexity for aggregating multi-scale features using high-order correlations.

This is important because hypergraph computation could otherwise become expensive for real-time object detection.

The design attempts to preserve strong modeling capability while maintaining practical inference efficiency.

What Is FullPAD?

FullPAD stands for:

Full-Pipeline Aggregation-and-Distribution Paradigm

It builds on HyperACE and distributes correlation-enhanced features across the entire detection pipeline.

Instead of using enhanced features at only one point in the network, FullPAD forwards them to multiple locations.

How FullPAD Works

According to the official repository, FullPAD uses three separate information paths.

Correlation-enhanced features are forwarded to:

1. The connection between the backbone and neck
2. Internal layers of the neck
3. The connection between the neck and detection head

This creates more consistent information sharing across the network.

Full-Pipeline Information Flow

A typical detector contains:

Input → Backbone → Neck → Head → Detection

YOLOv13 attempts to improve communication between these components.

With FullPAD, high-order correlation information is distributed through multiple stages instead of being isolated inside one feature aggregation block.

This improves:

• Feature reuse
• Information flow
• Gradient propagation
• Cross-scale representation
• Detection performance

The official repository specifically notes that FullPAD improves gradient propagation and detection quality.

YOLOv13 Lightweight Design

YOLOv13 is not focused only on accuracy.

The model also attempts to reduce unnecessary computational cost.

The authors replace some traditional large-kernel convolutions with depthwise separable convolution based blocks.

The repository lists modules including:

• DSConv
• DS-Bottleneck
• DS-C3k
• DS-C3k2

These blocks are designed to preserve a useful receptive field while reducing parameters and FLOPs.

Depthwise Separable Convolution

Depthwise separable convolution breaks a standard convolution into more computationally efficient operations.

Instead of performing every spatial and channel interaction simultaneously, the operation separates:

• Spatial filtering
• Channel combination

This can significantly reduce computational cost.

YOLOv13 uses this approach to make parts of the architecture lighter without sacrificing the broader receptive field required for object detection.

YOLOv13 Model Variants

The official YOLOv13 repository provides four main model sizes:

• YOLOv13-N
• YOLOv13-S
• YOLOv13-L
• YOLOv13-X

Unlike several other YOLO generations, the official repository does not currently list a Medium model in its main benchmark family.

YOLOv13-N

YOLOv13-N is the Nano model.

The official COCO benchmark reports:

• 2.5 million parameters
• 6.4 GFLOPs
• 41.6 AP50-95
• 57.8 AP50
• 45.1 AP75
• 1.97 ms latency

This model targets lightweight and low-latency applications.

YOLOv13-S

YOLOv13-S is the Small model.

The official benchmark reports:

• 9.0 million parameters
• 20.8 GFLOPs
• 48.0 AP50-95
• 65.2 AP50
• 52.0 AP75
• 2.98 ms latency

It provides a balance between speed and detection accuracy.

YOLOv13-L

YOLOv13-L is the Large model.

The official benchmark reports:

• 27.6 million parameters
• 88.4 GFLOPs
• 53.4 AP50-95
• 70.9 AP50
• 58.1 AP75
• 8.63 ms latency

This version offers higher capacity for accuracy-focused applications.

YOLOv13-X

YOLOv13-X is the Extra-Large model.

The official benchmark reports:

• 64.0 million parameters
• 199.2 GFLOPs
• 54.8 AP50-95
• 72.0 AP50
• 59.8 AP75
• 14.67 ms latency

Among the official N/S/L/X variants, YOLOv13-X provides the highest reported detection accuracy.

YOLOv13 Performance Table

These figures are reported in the official repository’s MS COCO benchmark.

YOLOv13 vs YOLOv12

YOLOv13 introduces a major conceptual change compared with YOLOv12.

YOLOv12 emphasizes attention-centric feature modeling, while YOLOv13 focuses on hypergraph-based high-order correlations.

The YOLOv13 paper argues that attention mechanisms in YOLOv12 are still mainly based on pairwise correlations, while HyperACE can model multi-to-multi relationships.

The official benchmark reports:

• YOLOv12-N: 40.1 AP50-95
• YOLOv13-N: 41.6 AP50-95

This gives YOLOv13-N a reported improvement of 1.5 AP points over YOLOv12-N.

YOLOv13 vs YOLO11

The paper also compares YOLOv13-N with YOLO11-N.

The authors report that YOLOv13-N improves mAP by approximately 3.0 points over YOLO11-N while using a similar lightweight model scale.

This improvement is mainly attributed to:

• HyperACE
• FullPAD
• Better high-order feature modeling
• Efficient lightweight blocks

YOLOv13 vs Earlier YOLO Models

YOLOv13 differs from earlier YOLO generations mainly in how it models relationships between visual features.

The major innovation is the use of hypergraphs for global cross-location and cross-scale feature relationships.

YOLOv13 Object Detection

YOLOv13 performs real-time object detection.

A typical detection includes:

• Bounding box location
• Object class
• Confidence score

The model processes an image through its feature extraction and aggregation pipeline before producing final detections.

The general workflow is:

1. Load an image.
2. Resize it to the model input size.
3. Extract backbone features.
4. Build high-order correlations with HyperACE.
5. Distribute enhanced features using FullPAD.
6. Process multi-scale features through the neck.
7. Generate predictions through the detection head.
8. Return final bounding boxes and classes.

YOLOv13 Input Size

The official training configuration uses an image size of:

640 × 640 pixels

The repository’s example training configuration specifies:

imgsz=640

This is also consistent with the standard COCO evaluation workflow used by the project.

YOLOv13 Training

The official repository provides training through the Ultralytics-style Python API.

The example configuration uses:

• 600 epochs
• Batch size 256
• Image size 640
• Mosaic augmentation
• MixUp configuration
• Copy-paste augmentation
• Multi-GPU training

The exact settings vary by model size.

Custom Dataset Training

YOLOv13 can also be trained on custom object detection datasets.

A custom dataset typically requires:

• Training images
• Validation images
• Bounding box labels
• Object class definitions
• Dataset configuration

Developers can replace the default COCO configuration with their own dataset configuration in the training workflow.

YOLOv13 Validation

The repository provides validation through the Ultralytics YOLO API.

The official example loads a pretrained YOLOv13 model and evaluates it on the COCO dataset.

Validation can measure:

• Precision
• Recall
• AP
• AP50
• AP75
• AP50-95

These metrics help evaluate both classification and localization quality.

YOLOv13 Prediction

YOLOv13 supports direct object detection through Python.

The repository demonstrates loading a YOLOv13 model and running predictions through the Ultralytics-style interface.

This makes it possible to integrate YOLOv13 into:

• Image analysis systems
• Video pipelines
• Robotics applications
• Industrial monitoring
• Research projects

YOLOv13 Export

The official repository supports model export to:

• ONNX
• TensorRT

Exporting allows YOLOv13 to run outside the original PyTorch environment.

YOLOv13 and TensorRT

YOLOv13 can be exported as a TensorRT engine.

The repository demonstrates:

model.export(format="engine", half=True)

for FP16 TensorRT export.

TensorRT can provide optimized inference on supported NVIDIA GPUs.

YOLOv13 and ONNX

YOLOv13 can also be exported to ONNX.

ONNX is useful for:

• Cross-platform deployment
• C++ applications
• Edge inference
• Hardware-specific runtimes
• Production environments

The official export workflow explicitly supports ONNX.

Flash Attention Support

The YOLOv13 repository states that the model supports Flash Attention acceleration.

The installation instructions provide a Flash Attention package as part of the recommended environment setup.

This can help improve the efficiency of certain attention-related operations.

YOLOv13 Installation

The official quick start uses:

• Python 3.11 environment
• Repository requirements
• Editable package installation
• Optional Flash Attention acceleration

The repository provides commands for creating a dedicated Conda environment and installing the package.

YOLOv13 Repository Structure

The repository includes major folders such as:

• assets
• docker
• examples
• tests
• ultralytics

It also includes:

• LICENSE
• README.md
• paper-yolov13.pdf
• pyproject.toml
• requirements.txt

The project states that its implementation is based on Ultralytics code.

YOLOv13 Android Deployment

The official repository records support for Android deployment through a community contribution.

This makes YOLOv13 potentially useful for mobile and edge-oriented object detection projects.

YOLOv13 FastAPI Support

The repository also lists a community-supported FastAPI REST API integration.

This can be useful for deploying YOLOv13 as a web-accessible inference service.

YOLOv13 Huawei Ascend and RKNN Support

The repository records community support for conversion to:

• Huawei Ascend OM
• Rockchip RKNN formats

These formats extend deployment possibilities to additional edge and accelerator hardware.

YOLOv13 and COCO

YOLOv13 was extensively evaluated on the MS COCO object detection dataset.

The official paper and repository use COCO to compare YOLOv13 with:

• YOLOv6
• YOLOv8
• YOLOv9
• YOLOv10
• YOLO11
• YOLOv12
• RT-DETR
• Gold-YOLO

The benchmark demonstrates the model’s accuracy, parameter efficiency, FLOPs, and latency across multiple model sizes.

Advantages of YOLOv13

High-Order Feature Modeling

HyperACE allows the network to capture multi-to-multi relationships that are difficult to represent with ordinary pairwise modeling.

Cross-Scale Feature Fusion

YOLOv13 can combine information from multiple feature-map scales through hypergraph computation.

Global Feature Relationships

The architecture can model relationships between distant spatial positions rather than relying only on local context.

Improved Information Flow

FullPAD distributes enhanced features throughout the detection pipeline.

Efficient Lightweight Blocks

Depthwise separable convolution based modules reduce parameters and computational cost.

Multiple Model Sizes

Nano, Small, Large, and Extra-Large models allow users to balance accuracy and performance.

Export Support

The official repository provides ONNX and TensorRT export.

YOLOv13 Limitations

Hypergraph Modeling Adds Complexity

YOLOv13 introduces more advanced architectural concepts than many earlier YOLO models.

This can make the model more difficult for beginners to understand or modify.

Larger Variants Require More Hardware

YOLOv13-L and YOLOv13-X require substantially more computation than Nano or Small variants.

Latency Increases with Model Size

The official benchmarks show latency increasing from 1.97 ms for YOLOv13-N to 14.67 ms for YOLOv13-X.

Deployment Performance Depends on Hardware

Actual performance varies depending on:

• GPU
• Runtime
• Precision
• Input resolution
• Export format
• Batch size

YOLOv13 Use Cases

YOLOv13 can be used for:

• Real-time object detection
• Vehicle detection
• Person detection
• Traffic monitoring
• Industrial inspection
• Robotics
• Drone vision
• Wildlife monitoring
• Agriculture
• Manufacturing
• Security systems
• Retail analytics
• Complex scene analysis
• Edge AI research
• Computer vision experiments

The model may be especially useful where relationships between multiple objects or regions are important.

YOLOv13 Release

The YOLOv13 paper was first submitted to arXiv on June 21, 2025 and later revised in September 2025.

The official repository states:

• Code open-sourced: June 21, 2025
• Model weights released: June 22, 2025
• Paper linked in repository: June 24, 2025

YOLOv13 Naming

The official model name is:

YOLOv13

The standard checkpoints use names such as:

• YOLOv13-N
• YOLOv13-S
• YOLOv13-L
• YOLOv13-X

This naming should be used consistently when referring to the official project.

YOLOv13 Paper

The original research paper is titled:

YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception

Authors:

• Mengqi Lei
• Siqi Li
• Yihong Wu
• Han Hu
• You Zhou
• Xinhu Zheng
• Guiguang Ding
• Shaoyi Du
• Zongze Wu
• Yue Gao

The paper focuses on overcoming the limitations of local feature aggregation and pairwise correlation modeling.

Why YOLOv13 Is Important

YOLOv13 introduces a new direction in real-time YOLO architecture design by bringing hypergraph computation into the feature modeling process.

Its main contributions include:

• HyperACE
• FullPAD
• Multi-to-multi correlation modeling
• Cross-location feature fusion
• Cross-scale feature fusion
• Better gradient propagation
• Depthwise separable lightweight blocks
• Real-time detection performance

The model demonstrates that object detectors can move beyond local and pairwise relationships and instead model more complex high-order visual interactions.

Frequently Asked Questions

What is YOLOv13?

YOLOv13 is a real-time object detector introduced in 2025 that uses hypergraph computation to model high-order visual relationships.

Who created YOLOv13?

YOLOv13 was introduced by Mengqi Lei, Siqi Li, Yihong Wu, Han Hu, You Zhou, Xinhu Zheng, Guiguang Ding, Shaoyi Du, Zongze Wu, and Yue Gao.

When was YOLOv13 introduced?

The original paper was submitted on June 21, 2025.

What is the YOLOv13 paper called?

The paper is titled:

YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception.

What is HyperACE?

HyperACE stands for Hypergraph-based Adaptive Correlation Enhancement.

It models high-order relationships between multiple visual features using hypergraph computation.

What is FullPAD?

FullPAD stands for Full-Pipeline Aggregation-and-Distribution.

It distributes correlation-enhanced features throughout the backbone, neck, and head connections of the network.

What YOLOv13 models are available?

The official family includes:

• YOLOv13-N
• YOLOv13-S
• YOLOv13-L
• YOLOv13-X

Is there a YOLOv13-M model?

The main official repository benchmark currently lists Nano, Small, Large, and Extra-Large variants and does not list a Medium model.

What input size does YOLOv13 use?

The official example training configuration uses 640 × 640 images.

Can YOLOv13 be trained on custom datasets?

Yes. The repository supports custom dataset configuration through the Ultralytics-style training workflow.

Does YOLOv13 support ONNX?

Yes. The official repository provides ONNX export.

Does YOLOv13 support TensorRT?

Yes. TensorRT engine export is supported.

Does YOLOv13 support Flash Attention?

Yes. The repository states that YOLOv13 supports Flash Attention acceleration.

Is YOLOv13 more accurate than YOLOv12?

In the official Nano-scale comparison, YOLOv13-N reports 41.6 AP50-95 compared with 40.1 for YOLOv12-N, an improvement of 1.5 points.

What is the smallest YOLOv13 model?

YOLOv13-N is the Nano variant with approximately 2.5 million parameters and 6.4 GFLOPs.

Which YOLOv13 model has the highest reported AP?

Among the official N/S/L/X variants, YOLOv13-X has the highest reported AP50-95 at 54.8.

Is YOLOv13 suitable for real-time detection?

Yes. YOLOv13 is explicitly designed as a real-time object detector, with the official project emphasizing both accuracy and efficiency.

Conclusion

YOLOv13 advances real-time object detection by introducing hypergraph-based high-order visual modeling.

Its HyperACE mechanism allows the network to discover multi-to-multi relationships across image locations and feature scales, while FullPAD distributes these enhanced features throughout the detection pipeline.

Combined with depthwise separable lightweight blocks, YOLOv13 aims to provide strong detection accuracy without unnecessary computational overhead.

With four official model sizes, custom training, Flash Attention acceleration, ONNX and TensorRT export, and support for several deployment environments, YOLOv13 represents an important evolution of the YOLO family toward richer global visual understanding and efficient real-time detection.